Analysis of the Functionality of the Feed Chain in Olive Pitting, Slicing and Stuffing Machines by IoT, Computer Vision and Neural Network Diagnosis

senso s A icle Analysis o he Func ionali y o he Feed Chain in Oli e Pi ing, Slicing and S u ing Machines by IoT, Compu e Vision and Neu al Ne wo k Diagnosis Albe o Lucas Pascual 1, An onio Madueño Luna 2, Manuel de Jóda Láza o 1,* , JoséMiguel Molina Ma ínez 1, An onio Ruiz Canales 3, JoséMiguel Madueño Luna 4 and Me i xell Jus icia Sego ia 3 1Food Enginee ing Depa men , Technical Uni e si y o Ca agena, 30203 Ca agena, Spain; [email p o ec ed] (A.L.P.); [email p o ec ed] (J.M.M.M.) 2Ae ospace Enginee ing and Fluid Mechanical Depa men , Uni e si y o Se ille, 41013 Se ille, Spain; [email p o ec ed] 3Enginee ing Depa men , Miguel He nández Uni e si y o Elche, 03312 O ihuela, Spain; [email p o ec ed] (A.R.C.); [email p o ec ed] (M.J.S.) 4G aphics Enginee ing Depa men , Uni e si y o Se ille, 41013 Se ille, Spain; [email p o ec ed] *Co espondence: [email p o ec ed] Recei ed: 15 Janua y 2020; Accep ed: 6 Ma ch 2020; Published: 10 Ma ch 2020   Abs ac : Oli e pi ing, slicing and s u ing machines (DRR in Spanish) a e cha ac e ized by he ac ha hei op imal unc ioning is based on app op ia e adjus men s. T adi ional sys ems a e no comple ely eliable because hei minimum e o a e is 1–2%, which can esul in ui loss, since he pi ing p ocess is no in allible, and ood sa e y issues can a ise. Such minimum e o s a e impossible o emo e h ough mechanical adjus men s. In o de o achie e his objec i e, an inno a i e solu ion mus be p o ided in o de o emo e e o s a ope a ing speed a es o e 2500 oli es/min. This wo k analyzes he app op ia e placemen o oli es in he pocke s o he eed chain by using he ollowing i ems: (1) An IoT Sys em o con ol he DRR machine and he da a analysis. (2) A compu e ision sys em wi h an ex e nal sho came a and a LED ligh ing sys em, which akes a pic u e o e e y pocke passing in on o he came a. (3) A chip wi h a neu al ne wo k o classi ica ion ha , once ained, classi ies be ween ou possible pocke cases: emp y, no mal, inco ec ly de-s oned oli es a any angles (also known as a “boa ”), and an anomalous case ( o eign elemen s such as lea s, small b anches o s ones, wo oli es o small pa s o oli es in he same pocke ). The main objec i e o his pape is o illus a e how wi h he use o a sys em based on IoT and a physical chip (Neu oMem CM1K, Gene al Vision Inc.) wi h neu al ne wo ks o so ing pu poses, i is possible o op imize he unc ionali y o his ype o machine by emo ely analyzing he da a ob ained. The use o classi ying ha dwa e allows i o wo k a he nominal ope a ing speed o hese machines. This would be limi ed i o he classi ying echniques based on so wa e we e used. Keywo ds: In e ne o hings (IoT); able oli e pi ing; slicing and s u ing machines; a i icial neu al ne wo ks (ANNs); CM1K chip; In el Cu ie chip; Teensy 1. In oduc ion Oli es we e pi ed and s u ed by hand un il he 1970s. As labo cos s g adually inc eased, he need o mechanize bo h p ocesses a ose. The indus ializa ion o he oli e indus y p og essed a a slow pace, and he i s con inuous oli e pi ing machines appea ed in 1975 [ 1 ]. Cu en models ha e imp o ed and a e able o pi oli es a a a e up o 2500 oli es/min. Howe e , he e has been a lack o elec onic imp o emen s because he ongoing mechanical imp o emen s in oli e pi ing machines Senso s 2020,20, 1541; doi:10.3390/s20051541 www.mdpi.com/jou nal/senso s Senso s 2020,20, 1541 2 o 22 ha e limi ed he inco po a ion o new echnologies and ine adjus men s. I is wo h men ioning howe e some o he ew ad ances in his a ea, including pe o mance op imiza ion and emo e e o de ec ion [ 2 ]. The oli e pi ing minimum e o a e o adi ional sys ems is 1–2% (i he machine is co ec ly adjus ed) and canno be imp o ed h ough mechanical adjus men s al hough his e o can inc ease due o he poo calib a ion o an oli e pi ing machine. The consequences o hese e o inc eases include a g ea e ui loss and ood sa e y issues, because he pi ing p ocess is no comple ely ensu ed, i may esul in a low-quali y p oduc . The g owing compe i i eness in p oduce coun ies, due o a globalized ma ke , and he inc ease o ood sa e y measu es ha e led o new echnological solu ions. These solu ions a e aimed a imp o ing he eliabili y o oli e pi ing machines du ing he pi ing p ocess by emo ing he s one comple ely and inc easing hei p oduc i i y o be e compe i ion wi h o he p oduce s. As a biological p oduc cha ac e ized by a he e ogeneous mo phology, oli es challenge de ec ion echnologies. The e o e, neu al ne wo ks a e he ools chosen o iden i y e e y possible case. Cases using deep lea ning and compu e ision ha e been con i med o deli e g ea esul s and a e p omising solu ions o sol ing he a o emen ioned p oblem. Yunchao e al. [ 3 ] p esen ed an a icle o in e es on he use o eal- ime de ec ion in ecycled agg ega e conc e e echnology. Howe e , o his pa icula case, i does no p io i ize he p ocessing speed. The e o e, his echnology does no ha e he same applicabili y in he p esen a icle. Ano he s udy ha could be conside ed o de e mining an applicable solu ion is he analysis de eloped by Mingyou e al. [ 4 ], who used a mul i-came a sys em, which is in e es ing o he analysis o se e al pi ing machines o inc ease he pe cep ion ange o ision. Howe e , he use o se e al came as is no he subjec o his s udy. The e o e, i is necessa y o explo e o he solu ions ha p o ide su icien p ocessing speed o analyze images a he speed p esc ibed by he pi ing machines. Al hough hese s udies use in e es ing echnology o he de ec ion and classi ica ion o images, hei esul s a e no applicable he e because p ocessing speed is no ele an . The e a e applica ions which use compu e isions wi h apid de ec ion echnologies o indus ial pu poses like p oduc classi ica ion o image analysis. These sys ems use mo phological ope a ions o ex ac he sough cha ac e is ics in he image h ough il e s and speci ic analysis algo i hms like Canny Edge, O su, local bina y pa e n algo i hm, K-means, e c. In o de o de elop hese algo i hms, speci ic so wa e is used, such as he in eg a ed de elopmen se ing o Mic oso Visual S udio and speci ic compu e ision lib a ies, as in he case wi h Openc . Maoyon, e al. [ 5 ] p oposes a s udy using hese ypes o echniques in he classi ica ion o apples on an indus ial le el. Al hough he desc ibed p ocedu es a e o g ea in e es , hese ypes o echniques equi e high-pe o mance compu ing ha dwa e and ele a ed cos s o he ype o applica ion e alua ed in he cu en analysis, whe e mo e economic solu ions and educed ha dwa e size a e sough . Lucas e al. [ 6 ] p opose a speci ic example done wi h oli es which uses hese as de ec ion echniques based on C++ p og amming language and OpenCV a i icial ision lib a ies. Speci ic s udies in he ag i- ood sec o and he ela ed beha io o neu al ne wo ks ha e been conside ed. I is wo h men ioning he s udies conduc ed by Guichao e al. [ 7 – 9 ] on de ec ion sys ems o he a i icial isualiza ion o ui s h ough RGB space colo analysis and obo ic sys ems. In he case o he classi ica ion o apples, Yang e al. [ 10 ] p o ide posi i e esul s. Yang’s in es iga ion used h ee laye s o 9-6-3 neu ons, wi h 96.6% accu acy. Naga a e al. [ 11 ] de eloped a g ading sys em o ui and ege ables using neu al ne wo k echnologies, ob aining a high pe cen age o accu acy o s awbe ies and g een peppe s (94% o 98% and 89%, espec i ely). Likewise, Beh ooz e al. [ 12 ] applied machine ision and a i icial neu al ne wo k (ANN) o modelling and con olling he g ape d ying p ocess, o e ing a new me hod o p edic i e modelling o he g ape d ying p ocess o he on-line moni o ing and con olling o he p ocess. The e ha e been some echniques used o classi y able oli es using compu e ision and neu al ne wo ks. Ga ika e al. [ 13 ] p oposed an in e es ing pape abou oli e ui ecogni ion using neu al ne wo ks. Oli e ui ecogni ion is pe o med by analyzing RGB images aken om oli e ees. Ano he Senso s 2020,20, 1541 3 o 22 s udy used a neu al ne wo k based on he backp opaga ion me hod. Mancuso e al. [ 14 ] iden i ied oli e (Olea eu opaea) cul i a s using a i icial neu al ne wo ks. Backp opaga ion neu al ne wo ks (BPNNs) we e used o dis inguish 10 oli e (Olea eu opaea L.) cul i a s ha o igina e h oughou he Medi e anean basin. The book Compu e Vision Technology o Food Quali y E alua ion, 2nd Edi ion [ 15 ], includes a speci ic chap e based on he di e en echnologies use o analyze he quali y o ood. Chap e s 11 o 13 desc ibe he quali y e alua ions o apples, ci us ui s, and s awbe ies, espec i ely. O pa icula in e es o he p oblem a hand, Chap e 14 [ 16 ] explains he classi ica ion and e alua ion o able oli es and desc ibes how o classi y hem by colo , shape, o ex e nal de ec s made by insec s. The usual way o classi y oli es is by using compu e ision. Tha pape analyzes he images cap u ed by a came a connec ed o a PC; hese images allow he analysis o 66 oli es pe ma ix. In ha pape , he au ho used a Bayesian ma h model o p e-classi ica ion o pe o m his p ocess. Neu al ne wo k so wa e was used, wi h 15 so ing pa ame e s and a hidden laye . The esul was success ul. The ne wo k was able o classi y mo e han ou ypes o oli es. The esul s, howe e , could be imp o ed by using high- esolu ion images. P e ious examples employed so wa e-based neu al ne wo ks, whe e he p ocessing speed is no ele an . Thus, al hough he de ec ion and classi ica ion echnology in hese s udies is applicable, i is necessa y o look o echnology ha achie es high p ocessing speeds. In addi ion, i is p e e able o use sys ems wi h limi ed physical implemen a ion ha can be ins alled wi h low in usion a lowe cos s, ins ead o adi ional ha dwa e, such as indus ial PC’s. The e o e, neu al ne wo k echnology based on a physical chipse ensu es i s success ul implemen a ion and low cos . The possible applica ions o his ype o chip a e e y b oad. The ollowing a icle is in e es ing because CM1K neu al chip o Gene al Vision Inc. (Pe aluma, CA 94952 USA) was implemen ed in a common indus ial p ocess used o ill bo les in ac o ies. This p ocess is qui e simple; howe e , i may be necessa y o use an in elligen de ice o inspec he p ocess and ensu e ood secu i y [ 17 ]. In ano he example, pa allel neu al ne wo k chips we e used o sho e o ish inspec ion be o e ille ing [ 18 ]. Each ne wo k chip sys em uses ou neu al ne wo k chips (accoun ing o 312 neu ons) based on a na i ely pa allel, ha d-wi ed a chi ec u e ha pe o ms eal- ime lea ning and nonlinea classi ica ion (RBF). The use o CM1K in eal ime e i ies i s use o he pu poses o he p esen s udy. The combina ion o compu e ision and neu al ne wo ks o e s a way o pe o m asks ha could be mo e complex. Liu e al. [ 19 ] p oposed a neu al ne wo k chip o license pla e ecogni ion. This chip combines a ideo image-p ocessing module wi h a neu al ne wo k module by using equalized image p ocessing algo i hms and ne wo k classi ica ion algo i hms. San u Sa da e al. [ 20 ] published a pape based on au oma ed acial ecogni ion, which is a echnique employed in a wide ange o p ac ical applica ions, including pe sonnel access con ol and iden i ica ion sys ems. Image ecogni ion is simple han image p ocessing me hods o acial ecogni ion, mainly due o he lack o a ixed pa e n o compa ison pu poses; hese applica ions e eal he emendous possibili ies o CM1K. La e CM1K chipse s ha e deli e ed sa is ac o y esul s. An in e es ing aspec o conside in u u e esea ch would be o compa ison ou chosen chip wi h hese o he echnologies desc ibed in he ollowing s udies [21–30]. The use o he IoT is widely ex ended nowadays and he e a e nume ous cases, o example [ 31 – 35 ] in p ecision ag icul u e, i iga ion, empe a u e con ol, moni o ing o he ag icul u al p oduc ion p ocess o au oma ed-oli e-chain. In he case p esen ed in his pape , we a e going o be able o moni o and analyze he ope a ion o DRR machines h ough he in e ne . 2. Ma e ials and Me hods In his sec ion, he in e nal s uc u e o he CM1K chip (Sec ions 2.1 and 2.2) and he implemen a ion o Ma lab o neu al ne wo ks wi h cha ac e is ics simila o hose o he CM1K chip a e analyzed (Sec ion 2.3). A chip such as he CM1K has a limi a ion in ha he en y ec o i can suppo is educed in his case (256 by es), meaning ha i i is going o p ocess an image, ha image canno be la ge in pixel Senso s 2020,20, 1541 4 o 22 size han his alue (256 by es) and ha i could only be B/W wi h 256 le els o g ay. In o de o assu e ha a neu al ne wo k wi h hese cha ac e is ics (equipped wi h high speed as i is pu e ha dwa e, bu o limi ed ec o en y) can be accep able o he pu pose o classi ica ion in oli e pi ing machines, a neu al ne wo k is simula ed in Ma lab which wo ks wi h a delay and allows us o assu e ha co ec classi ica ion will be ca ied ou in a ious ca ego ies wi h images o jus 16 ×16 pixels in B/W. In Sec ions 2.4 and 2.5 he physical implemen a ion o he classi ica ion sys em is shown wi h a eal oli e pi ing machine and in Sec ion 2.6 he comme cial e sion o he CM1K chip ha has been used in he ials is analyzed. Fo hese sec ions, he desc ip ion supplied in [ 36 ] will be used by he au ho s o his pape . In Sec ion 2.7 he communica ion ha dwa e o he CM1K-PC o classi ica ion in eal ime is desc ibed and las ly, in Sec ion 2.8, he IoT sys em o con ol he oli e pi ing, slicing and s u ing machine (DRR in Spanish) and da a analysis is analyzed. 2.1. Neu al Ne wo k In his wo k a supe ised neu al ne wo k wi h backp opaga ion has been chosen because he e a e no ele an di e ences wi h espec o he calcula ion p ocesses used by o he ne wo ks. Mo eo e , his sys em is one o he mos widesp ead sys ems used o image iden i ica ion pu poses, and i is widely used in he MATLAB Neu al Ne wo k Toolbox employed in his pape . Backp opaga ion is he mos e icien way o se his alue [ 37 ]. Fi s , e o s a e calcula ed in ou pu uni s by conside ing he di e ence be ween he desi ed and p ede e mined alues. Nex , hey p opaga e h ough he ne wo k using he weigh s and ob ain he minimum alue in he mos op imal way. The so wa e emula ion o a neu al ne wo k is compu a ionally demanding, which indica es ha i s ope a ions will ake a long pe iod o ime o be ca ied ou . The use o a chip ha physically implemen s a neu al ne wo k (as in his pape wi h he CM1K chip o In el Cu ie chip) will speed up he aining and esponse p ocesses o he neu al ne wo k. A compa ison wi h o he exis ing ha dwa e neu al ne wo ks like Google Co al Edge TPU [ 38 ], In el ® Mo idius ™ Neu al Compu e S ick 2 [ 39 ] o N idia-Je son-Nano [40] and op imized esul s will be p o ided in u u e pape s. 2.2. Ope a ion o he CM1K Chip The CM1K is a pa e n ecogni ion accele a o chip ha is ainable in eal- ime by lea ning examples [ 41 ]. I is a ully pa allel silicon neu al ne wo k o ei he lea ning o ecogni ion [ 42 ], ha can s o e and p ocess in o ma ion simul aneously. I is composed o ou modules [ 43 ], has wo possible classi ie s: K-nea es neighbo (KNN) o adial basis unc ion (RBF) [ 44 ] and uses he “Winne -Takes-All” s a egy [ 45 ]. I has 1024 neu ons o 256 by es. In el Cu ie chip is a diminished e sion wi h only 128 neu ons o 128 by es. 2.3. The MATLAB Neu al Ne wo k Toolbox In his wo k we a e using he MATLAB Neu al Ne wo k [ 46 ] by using a speci ic ne wo k called ‘au oencode ’ o classi ica ion pu poses [ 47 , 48 ], i will allow us o analyze he e ec o image esolu ion on he so ing capaci y o he ne wo k. 2.3.1. P elimina y Tes s: Maximum Resolu ion A ailable The maximum esolu ion implemen ed unde came a boa d is 144 × 176 mee ing he size equi emen s o he pocke acco ding o used lenses. A e he i s es using he abo e-men ioned esolu ion, a 25,344 pixels a 1-by e colo dep h g ayscale image is ob ained, which implies a high p ocessing a e (nea ly wo minu es using a 2.5 GHz CPU, RAM 8 GB, Mic oso Windows 10, Redmond, Washing on, USA, and Ma hWo ks MATLAB 2019, Massachuse s, USA). Senso s 2020,20, 1541 5 o 22 P elimina y es s sugges ha p ocessing op imiza ion measu es need o be ca ied ou . These measu es will ocus on wo main aspec s: •Es ablishing a egion o in e es (ROI) on he image. • Tes ing di e en esolu ions ha allow he iden i ica ion o he image a he minimum p ocessing a e. In addi ion o hese aspec s, i is impo an o conside he p ocessing a e o he physical chips: In el Cu ie (128 neu ons wi h a 128-by e inpu ec o ) and Neu oMem CM1K (1024 neu ons wi h a 256-by e inpu ec o ). Conside ing he abo e-men ioned es ic ions, he maximum esolu ion o he image mus be be ween 128 and 256 pixels o a squa e image, as inally used: 16x16 pixels o Neu oMem CM1K 256-by e chip and 11 ×11 pixels (121 by es <128 by es) o In el Cu ie chip. 2.3.2. P elimina y Tes s: Minimum Resolu ion Se e al es s o iden i y he minimum esolu ion accep ed by physical chips a e ca ied ou wi h he pu pose o se ing he lowes p ocessing a e. To do so, eal images a e o be p ocessed in MATLAB and i should be es ima ed i he sys em is able o iden i y hem. The i s s ep is o se he ROI, which mus be a pa o he image wi h enough in o ma ion in o de o iden i y he posi ion o he oli e in he pocke . The e o e, he oli e mus be comple ely displayed on he image. The ollowing Figu es 1and 2show some examples o 176 × 144 pixel-images wi h he (X,Y) e e ence sys em. As shown in he igu es, he oli e is es ic ed in he X-axis ( he same pocke con ines he oli e o i s space) and, he e o e, he selec ion o he pa ame e s o ROI in ha axis is cons an . The e is a possibili y o he oli e o mo e in Y-axis, consequen ly leading o a de ia ion o he dimensions o ROI. Howe e , i is empi ically p o en ha , due o he inclina ion o he chain and he upwa d mo emen o he oli e, g a i y, ine ia and some elemen s o he machine (b ushes, o example) o ces he oli e o s ay in a simila posi ion. Va ia ions may be iden i ied in o ien a ion, which is he subjec o his pape . Senso s 2020, 20, x FOR PEER REVIEW 5 o 22  Tes ing di e en esolu ions ha allow he iden i ica ion o he image a he minimum p ocessing a e. In addi ion o hese aspec s, i is impo an o conside he p ocessing a e o he physical chips: In el Cu ie (128 neu ons wi h a 128-by e inpu ec o ) and Neu oMem CM1K (1024 neu ons wi h a 256-by e inpu ec o ). Conside ing he abo e-men ioned es ic ions, he maximum esolu ion o he image mus be be ween 128 and 256 pixels o a squa e image, as inally used: 16x16 pixels o Neu oMem CM1K 256-by e chip and 11 × 11 pixels (121 by es < 128 by es) o In el Cu ie chip. 2.3.2. P elimina y Tes s: Minimum Resolu ion Se e al es s o iden i y he minimum esolu ion accep ed by physical chips a e ca ied ou wi h he pu pose o se ing he lowes p ocessing a e. To do so, eal images a e o be p ocessed in MATLAB and i should be es ima ed i he sys em is able o iden i y hem. The i s s ep is o se he ROI, which mus be a pa o he image wi h enough in o ma ion in o de o iden i y he posi ion o he oli e in he pocke . The e o e, he oli e mus be comple ely displayed on he image. The ollowing Figu es 1 and 2 show some examples o 176 × 144 pixel-images wi h he (X,Y) e e ence sys em. As shown in he igu es, he oli e is es ic ed in he X-axis ( he same pocke con ines he oli e o i s space) and, he e o e, he selec ion o he pa ame e s o ROI in ha axis is cons an . The e is a possibili y o he oli e o mo e in Y-axis, consequen ly leading o a de ia ion o he dimensions o ROI. Howe e , i is empi ically p o en ha , due o he inclina ion o he chain and he upwa d mo emen o he oli e, g a i y, ine ia and some elemen s o he machine (b ushes, o example) o ces he oli e o s ay in a simila posi ion. Va ia ions may be iden i ied in o ien a ion, which is he subjec o his pape . Y X Y X Figu e 1. Example o he use o egion o in e es (ROI) in an oli e in a no mal posi ion. The ed poin co esponds o he o igin o coo dina es o pixels. The g een ec angle comp ises he pocke wi hin i s walls and dep h and limi ed by he p e ious and he nex pocke o he eed chain. The ed ec angle is he ROI. Figu e 1. Example o he use o egion o in e es (ROI) in an oli e in a no mal posi ion. The ed poin co esponds o he o igin o coo dina es o pixels. The g een ec angle comp ises he pocke wi hin i s walls and dep h and limi ed by he p e ious and he nex pocke o he eed chain. The ed ec angle is he ROI. Senso s 2020,20, 1541 6 o 22 Senso s 2020, 20, x FOR PEER REVIEW 6 o 22 Figu e 2. Example o he use o ROI in an oli e in a “boa ” posi ion be ween [80°, 90°] and [−80°, −90°]. A e p e ious conside a ions, he ollowing condi ions a e se ou (measu ed in pixels and e e ed o he o igin o coo dina es, Figu e 3): Figu e 3. Selec ed ROI. Once he ROI pa ame e s in X and Y a e se , se e al es s a e ca ied ou in o de o es ablish i he sys em is s ill able o iden i y images using he selec ed esolu ions. A se o 10,000 images o 176 × 144 pixels in B/W o he oli es in no mal posi ion and boa and emp y bucke s a e a ailable, he e a e no anomalous cases (i.e., lea es, small b anches, s ones, double oli es, small pa s) in his se . This se is a ailable o o he esea che s [49]. Wi h an applica ion in Ma lab, each image is bina ized and he o ien a ion o each oli e is calcula ed. Fo he es s desc ibed in his wo k, hey a e andomly selec ed in g oups o 300 oli es. The o ien a ion o each oli e (no mal, in e media e posi ion o boa ) is de ined by he posi ion ela i e o he punches (Figu e 4). Figu e 2. Example o he use o ROI in an oli e in a “boa ” posi ion be ween [80 ◦ , 90 ◦ ] and [ − 80 ◦ , − 90 ◦ ]. A e p e ious conside a ions, he ollowing condi ions a e se ou (measu ed in pixels and e e ed o he o igin o coo dina es, Figu e 3): Senso s 2020, 20, x FOR PEER REVIEW 6 o 22 Figu e 2. Example o he use o ROI in an oli e in a “boa ” posi ion be ween [80°, 90°] and [−80°, −90°]. A e p e ious conside a ions, he ollowing condi ions a e se ou (measu ed in pixels and e e ed o he o igin o coo dina es, Figu e 3): Figu e 3. Selec ed ROI. Once he ROI pa ame e s in X and Y a e se , se e al es s a e ca ied ou in o de o es ablish i he sys em is s ill able o iden i y images using he selec ed esolu ions. A se o 10,000 images o 176 × 144 pixels in B/W o he oli es in no mal posi ion and boa and emp y bucke s a e a ailable, he e a e no anomalous cases (i.e., lea es, small b anches, s ones, double oli es, small pa s) in his se . This se is a ailable o o he esea che s [49]. Wi h an applica ion in Ma lab, each image is bina ized and he o ien a ion o each oli e is calcula ed. Fo he es s desc ibed in his wo k, hey a e andomly selec ed in g oups o 300 oli es. The o ien a ion o each oli e (no mal, in e media e posi ion o boa ) is de ined by he posi ion ela i e o he punches (Figu e 4). Figu e 3. Selec ed ROI. Once he ROI pa ame e s in X and Y a e se , se e al es s a e ca ied ou in o de o es ablish i he sys em is s ill able o iden i y images using he selec ed esolu ions. A se o 10,000 images o 176 × 144 pixels in B/W o he oli es in no mal posi ion and boa and emp y bucke s a e a ailable, he e a e no anomalous cases (i.e., lea es, small b anches, s ones, double oli es, small pa s) in his se . This se is a ailable o o he esea che s [ 49 ]. Wi h an applica ion in Ma lab, each image is bina ized and he o ien a ion o each oli e is calcula ed. Fo he es s desc ibed in his wo k, hey a e andomly selec ed in g oups o 300 oli es. The o ien a ion o each oli e (no mal, in e media e posi ion o boa ) is de ined by he posi ion ela i e o he punches (Figu e 4). Senso s 2020,20, 1541 7 o 22 Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1): •[0◦, 10◦] and [0◦,−10◦] (No mal) •[10◦, 20◦] and [−10◦,−20◦] (In e media e) •[20◦,−30◦] and [−20◦,−30◦] (In e media e) •[30◦,−40◦] and [−30◦,−40◦] (In e media e) •[40◦,−50◦] and [−40◦,−50◦] (In e media e) •[50◦,−60◦] and [−50◦,−60◦] (In e media e) •[60◦,−70◦] and [−60◦,−70◦] (In e media e) •[70◦,−80◦] and [−70◦,−80◦] (In e media e) •[80◦,−90◦] and [−80◦,−90◦] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0◦, 10º] ANGLE [0◦,−10◦] ANGLE [10◦,20◦] ANGLE [−10◦,−20◦] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] ANGLE [20◦,30◦] ANGLE [−20◦,−30◦] ANGLE [30◦,40◦] ANGLE [-30◦,-40◦] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020, 20, x FOR PEER REVIEW 7 o 22 Punch needle Oli e in no mal posi ion Oli e in boa posi ion Ha o deboning Figu e 4. Punch needles in an oli e pi ing machine. The classi ica ion by angles is as ollows (see Table 1):  [0°, 10°] and [0°, −10°] (No mal)  [10°, 20°] and [−10°, −20°] (In e media e)  [20°, −30°] and [−20°, −30°] (In e media e)  [30°, −40°] and [−30°, −40°] (In e media e)  [40°, −50°] and [−40°, −50°] (In e media e)  [50°, −60°] and [−50°, −60°] (In e media e)  [60°, −70°] and [−60°, −70°] (In e media e)  [70°, −80°] and [−70°, −80°] (In e media e)  [80°, −90°] and [−80°, −90°] (Boa ) Table 1. Images (bo h o iginal and bina ized) classi ied acco ding o angula in e als. ANGLE [0°, 10º] ANGLE [0°,−10°] ANGLE [10°,20°] ANGLE [−10°,−20°] ANGLE [20°,30°] ANGLE [−20°,−30°] ANGLE [30°,40°] ANGLE [-30°,-40°] Senso s 2020,20, 1541 8 o 22 Table 1. Con . ANGLE [40◦,50◦] ANGLE [−40◦,−50◦] ANGLE [50◦,60◦] ANGLE [−50◦,−60◦] Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). ANGLE [60◦,70◦] ANGLE [−60◦,−70◦] ANGLE [70◦,80◦] ANGLE [−70◦,−80◦] Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). ANGLE [80◦,90◦] o [−80◦,−90◦] Senso s 2020, 20, x FOR PEER REVIEW 8 o 22 ANGLE [40°,50°] ANGLE [−40°,−50°] ANGLE [50°,60°] ANGLE [−50°,−60°] ANGLE [60°,70°] ANGLE [−60°,−70°] ANGLE [70°,80°] ANGLE [−70°,−80°] ANGLE [80°,90°] o [−80°,−90°] 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions:  Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e .  Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s:  The ha dwa e ha le s he came a [50] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5 and 6). 2.4. Ha dwa e Used in Image Cap u e The sys em de eloped has wo unc ions: • Ob ain images o de e ed analysis wi h Ma lab, he In el Cu ie and CM1K neu al chips o e alua e he ope a ion o he la e . • Cha ac e ize he eal- ime ope a ion h ough IoT o he pi ing machines o an oli e ac o y ha each speeds o up o 2500 oli es/min. The sys em consis s o wo pa s: • The ha dwa e ha le s he came a [ 50 ] con ol and ligh ing based on a magne ic senso ha allows o ob ain images o each bucke acing he came a (Figu es 5and 6). • A pe sonal compu e (PC) which allows, o he i s case, he s o age o images o de e ed analysis and in he second case oge he wi h a CM1K chip o eal- ime analysis o he same o e he in e ne wi h an applica ion de eloped in Q C ea o [51] (Figu e 7). Senso s 2020,20, 1541 9 o 22 Senso s 2020, 20, x FOR PEER REVIEW 9 o 22 Suppo pla e Magne ic senso PVC adap e Figu e 5. Magne ic senso used o de ec he passage o he pocke s in he chain. Powe supply 24V o ci cui clean ision a ea Ci cui o con ol LED and came a Ci cui o con ol magne ic senso Powe supply 12V o con ol leds Powe supply 5V o oc ocouple ci cui o came a igge Figu e 6. Elec onic ci cui o he ex e nal igge and LED.  A pe sonal compu e (PC) which allows, o he i s case, he s o age o images o de e ed analysis and in he second case oge he wi h a CM1K chip o eal- ime analysis o he same o e he in e ne wi h an applica ion de eloped in Q C ea o [51] (Figu e 7). Figu e 5. Magne ic senso used o de ec he passage o he pocke s in he chain. Senso s 2020, 20, x FOR PEER REVIEW 9 o 22 Suppo pla e Magne ic senso PVC adap e Figu e 5. Magne ic senso used o de ec he passage o he pocke s in he chain. Powe supply 24V o ci cui clean ision a ea Ci cui o con ol LED and came a Ci cui o con ol magne ic senso Powe supply 12V o con ol leds Powe supply 5V o oc ocouple ci cui o came a igge Figu e 6. Elec onic ci cui o he ex e nal igge and LED.  A pe sonal compu e (PC) which allows, o he i s case, he s o age o images o de e ed analysis and in he second case oge he wi h a CM1K chip o eal- ime analysis o he same o e he in e ne wi h an applica ion de eloped in Q C ea o [51] (Figu e 7). Figu e 6. Elec onic ci cui o he ex e nal igge and LED. Senso s 2020,20, 1541 16 o 22 Table 4. Resul s o he aining pa e ns using a esolu ion o 1 6 ×16 pixels. Type o Tes Deg ees In e media e Posi ion Simple Neu on 66 Neu ons 24 Neu ons 4 Boa d Vec o Repea abili y E o Repea abili y E o Repea abili y E o B ainca d (CM1K) 256 18%13%11% 29%24%20% 3 10% 3 4% 3 0% 4 12% 4 3% 4 2% 5 13% 5 4% 5 1% 6 13% 6 5% 6 1% 7 15% 7 2% 7 0% 8 16% 8 5% 8 1% 9 16% 9 8% 9 1% 10 13% 10 2% 10 0% To als 12.43% 4% 0.63% The op imal op ion is B ainca d 256 chip (0.63% e o a e <1%) and only needs ou neu ons o classi y be ween no mal, emp y and boa oli es. I we use an In el cu ie chip e o a e is 4.80% <5% and only needs ou neu ons oo. This means ha he sugges ed sys em is capable o de ec ing up o a 95–99% o he de iciencies (boa s and emp y) p esen ed by he machine. 3.3. Analysis o he Resul s o he Ope a ion o he DRR Machine An applica ion (“Logge DRR”) has been de eloped o analyze he ope a ion o he oli e pi ing machine wi h GUI by Ma lab (Figu e 15). The mos ele an pa ame e s o his applica ion will be indica ed below: Senso s 2020, 20, x FOR PEER REVIEW 16 o 22 Table 4. Resul s o he aining pa e ns using a esolu ion o 1 6 × 16 pixels. Type o Tes Deg ees In e media e Posi ion Simple Neu on 66 Neu ons 24 Neu ons 4 Boa d Vec o Repea abili y E o Repea abili y E o Repea abili y E o B ainca d (CM1K) 256 1 8% 1 3% 1 1% 2 9% 2 4% 2 0% 3 10% 3 4% 3 0% 4 12% 4 3% 4 2% 5 13% 5 4% 5 1% 6 13% 6 5% 6 1% 7 15% 7 2% 7 0% 8 16% 8 5% 8 1% 9 16% 9 8% 9 1% 10 13% 10 2% 10 0% To als 12.43% 4% 0.63% The op imal op ion is B ainca d 256 chip (0.63% e o a e < 1%) and only needs ou neu ons o classi y be ween no mal, emp y and boa oli es. I we use an In el cu ie chip e o a e is 4.80% < 5% and only needs ou neu ons oo. This means ha he sugges ed sys em is capable o de ec ing up o a 95–99% o he de iciencies (boa s and emp y) p esen ed by he machine. 3.3. Analysis o he Resul s o he Ope a ion o he DRR Machine An applica ion (“Logge DRR”) has been de eloped o analyze he ope a ion o he oli e pi ing machine wi h GUI by Ma lab (Figu e 15). The mos ele an pa ame e s o his applica ion will be indica ed below: Ins an aneous o accumula ed da a scale Selec ed pe iod, da e and ime ange. Ins an aneous o accumula ed g aph: Boa s and Doubles Figu e 15. G aphical Use In e ace (GUI) c ea ed in Ma lab o analyze da a om he “DRR machines”. The ep esen a i e da a is indica ed in he ollowing box (Figu e 16:  Speed: Indica ed in pocke s pe minu e  P oduc ion: Real alue in oli es pe minu e (wi hou emp y pocke )  “Boa ” oli es, no mal, double oli es, emp y, small pieces, anomalies: Pe cen age o he o al p oduc ion.  Accumula ed alues: Da a added in he selec ed pe iod. Figu e 15. G aphical Use In e ace (GUI) c ea ed in Ma lab o analyze da a om he “DRR machines”. The ep esen a i e da a is indica ed in he ollowing box (Figu e 16: •Speed: Indica ed in pocke s pe minu e •P oduc ion: Real alue in oli es pe minu e (wi hou emp y pocke ) • “Boa ” oli es, no mal, double oli es, emp y, small pieces, anomalies: Pe cen age o he o al p oduc ion. •Accumula ed alues: Da a added in he selec ed pe iod. Senso s 2020,20, 1541 17 o 22 Senso s 2020, 20, x FOR PEER REVIEW 17 o 22 Figu e 16. Se ings dialogue box o he GUI om Ma lab. In Figu e 17, a p ac ical example o use is shown. The e a e wo accumula ed g aphs, he oli e g een one co esponding o “boa ” oli es de ec ed and he och e colo co esponding o double oli es en e ing in he pocke . Each da a upda e poin co esponds o a coun o 1000 oli es. Machine s op (1) Recalib a ion s op (3) Figu e 17. GUI o “boa ” oli es and doubles o he “DRR machine” diagnosis. The spaces co espond o pe iods when he machine was s opped (1), his can be due o changes in he eed chain (change o con aine whe e he oli es come om) o s ops o adjus he machine o pe sonnel b eaks. In (2), he machine is obse ed o ha e a misalignmen (due o he misalignmen o he b ush which s aigh ens he oli es in he eeding chain), ollowed by a s op o ecalib a ion (3). Immedia ely a e ha , we can obse e how he slope dec eases (4)→(5) which means he numbe o “boa ” oli es has dec eased. Addi ionally, he numbe o double oli es ( a ious in he same pocke ) inc eases (6), meaning ha he ecalib a ion has no been op imal and he eed pla e has been poo ly adjus ed. Figu e 18 shows an example o a ious combined pa ame e s (boa accumula es, emp y accumula ed, doubles accumula ed and small pa s accumula ed). Figu e 16. Se ings dialogue box o he GUI om Ma lab. In Figu e 17, a p ac ical example o use is shown. The e a e wo accumula ed g aphs, he oli e g een one co esponding o “boa ” oli es de ec ed and he och e colo co esponding o double oli es en e ing in he pocke . Each da a upda e poin co esponds o a coun o 1000 oli es. Senso s 2020, 20, x FOR PEER REVIEW 17 o 22 Figu e 16. Se ings dialogue box o he GUI om Ma lab. In Figu e 17, a p ac ical example o use is shown. The e a e wo accumula ed g aphs, he oli e g een one co esponding o “boa ” oli es de ec ed and he och e colo co esponding o double oli es en e ing in he pocke . Each da a upda e poin co esponds o a coun o 1000 oli es. Machine s op (1) Recalib a ion s op (3) Figu e 17. GUI o “boa ” oli es and doubles o he “DRR machine” diagnosis. The spaces co espond o pe iods when he machine was s opped (1), his can be due o changes in he eed chain (change o con aine whe e he oli es come om) o s ops o adjus he machine o pe sonnel b eaks. In (2), he machine is obse ed o ha e a misalignmen (due o he misalignmen o he b ush which s aigh ens he oli es in he eeding chain), ollowed by a s op o ecalib a ion (3). Immedia ely a e ha , we can obse e how he slope dec eases (4)→(5) which means he numbe o “boa ” oli es has dec eased. Addi ionally, he numbe o double oli es ( a ious in he same pocke ) inc eases (6), meaning ha he ecalib a ion has no been op imal and he eed pla e has been poo ly adjus ed. Figu e 18 shows an example o a ious combined pa ame e s (boa accumula es, emp y accumula ed, doubles accumula ed and small pa s accumula ed). Figu e 17. GUI o “boa ” oli es and doubles o he “DRR machine” diagnosis. The spaces co espond o pe iods when he machine was s opped (1), his can be due o changes in he eed chain (change o con aine whe e he oli es come om) o s ops o adjus he machine o pe sonnel b eaks. In (2), he machine is obse ed o ha e a misalignmen (due o he misalignmen o he b ush which s aigh ens he oli es in he eeding chain), ollowed by a s op o ecalib a ion (3). Immedia ely a e ha , we can obse e how he slope dec eases (4) → (5) which means he numbe o “boa ” oli es has dec eased. Addi ionally, he numbe o double oli es ( a ious in he same pocke ) inc eases (6), meaning ha he ecalib a ion has no been op imal and he eed pla e has been poo ly adjus ed. Figu e 18 shows an example o a ious combined pa ame e s (boa accumula es, emp y accumula ed, doubles accumula ed and small pa s accumula ed). Senso s 2020,20, 1541 18 o 22 Senso s 2020, 20, x FOR PEER REVIEW 18 o 22 Fi s misalignmen o he eed pla e (1) Second misalignmen o he eed pla e (2) Thi d misalignmen o he eed pla e (3) Lack o p oduc in he eede while he DRR machine is ope a ing (4) Figu e 18. Mul iple accumula ed alues (boa s, emp y, doubles and small pa s). This ype o g aph allows us o obse e i he calib a ion p ocess en ails a comp ehensi e imp o emen in pe o mance o , con e sely, whe he an imp o emen in a pa ame e a e an adjus men en ails a educ ion in goodness o i in ano he . In (1), (2) and (3) i can be obse ed how he eeding pla e goes ou o adjus men a li le mo e each ime, which has an impac on he numbe o doubles and small pa s. Du ing all o his ime, he g ow h in he numbe o “boa ” oli es is mono onous which means ha he e a e no misalignmen s in he s aigh ening b ush. Fo i s pa , he numbe o emp y (4) inc eases ab up ly each ime a lack o p oduc in he eede when he DRR machine is wo king. A imes, he da a analysis wi hou accumula ing da a can be in e es ing in o de o see he e ec o speci ic ac ions. In Figu e 19, he peaks in he “boa ” oli es/doubles (1), o only doubles (2), emp y (3) a e obse ed due o he ac ion o isola ed adjus men s o ope a ion o he machine. In all o he cases shown, hese ac ions a e un a o able (ei he calib a ion o ope a ion). (1) (2) (3) (2) Figu e 19. Ins an aneous alues o boa s, doubles, small pa s and emp y. Figu e 18. Mul iple accumula ed alues (boa s, emp y, doubles and small pa s). This ype o g aph allows us o obse e i he calib a ion p ocess en ails a comp ehensi e imp o emen in pe o mance o , con e sely, whe he an imp o emen in a pa ame e a e an adjus men en ails a educ ion in goodness o i in ano he . In (1), (2) and (3) i can be obse ed how he eeding pla e goes ou o adjus men a li le mo e each ime, which has an impac on he numbe o doubles and small pa s. Du ing all o his ime, he g ow h in he numbe o “boa ” oli es is mono onous which means ha he e a e no misalignmen s in he s aigh ening b ush. Fo i s pa , he numbe o emp y (4) inc eases ab up ly each ime a lack o p oduc in he eede when he DRR machine is wo king. A imes, he da a analysis wi hou accumula ing da a can be in e es ing in o de o see he e ec o speci ic ac ions. In Figu e 19, he peaks in he “boa ” oli es/doubles (1), o only doubles (2), emp y (3) a e obse ed due o he ac ion o isola ed adjus men s o ope a ion o he machine. In all o he cases shown, hese ac ions a e un a o able (ei he calib a ion o ope a ion). Senso s 2020, 20, x FOR PEER REVIEW 18 o 22 Fi s misalignmen o he eed pla e (1) Second misalignmen o he eed pla e (2) Thi d misalignmen o he eed pla e (3) Lack o p oduc in he eede while he DRR machine is ope a ing (4) Figu e 18. Mul iple accumula ed alues (boa s, emp y, doubles and small pa s). This ype o g aph allows us o obse e i he calib a ion p ocess en ails a comp ehensi e imp o emen in pe o mance o , con e sely, whe he an imp o emen in a pa ame e a e an adjus men en ails a educ ion in goodness o i in ano he . In (1), (2) and (3) i can be obse ed how he eeding pla e goes ou o adjus men a li le mo e each ime, which has an impac on he numbe o doubles and small pa s. Du ing all o his ime, he g ow h in he numbe o “boa ” oli es is mono onous which means ha he e a e no misalignmen s in he s aigh ening b ush. Fo i s pa , he numbe o emp y (4) inc eases ab up ly each ime a lack o p oduc in he eede when he DRR machine is wo king. A imes, he da a analysis wi hou accumula ing da a can be in e es ing in o de o see he e ec o speci ic ac ions. In Figu e 19, he peaks in he “boa ” oli es/doubles (1), o only doubles (2), emp y (3) a e obse ed due o he ac ion o isola ed adjus men s o ope a ion o he machine. In all o he cases shown, hese ac ions a e un a o able (ei he calib a ion o ope a ion). (1) (2) (3) (2) Figu e 19. Ins an aneous alues o boa s, doubles, small pa s and emp y. Figu e 19. Ins an aneous alues o boa s, doubles, small pa s and emp y. Senso s 2020,20, 1541 19 o 22 Finally, Figu e 20 shows ins ances in which he pe o mance dec eases when wo king wi h a machine wi h a poo ly adjus ed eeding pla e and s aigh ening b ush (1), only wi h he eed pla e misaligned (2) o due o he lack o p oduc in he eede (3). Senso s 2020, 20, x FOR PEER REVIEW 19 o 22 Finally, Figu e 20 shows ins ances in which he pe o mance dec eases when wo king wi h a machine wi h a poo ly adjus ed eeding pla e and s aigh ening b ush (1), only wi h he eed pla e misaligned (2) o due o he lack o p oduc in he eede (3). (1) (2) (2) (3) Figu e 20. Ins an aneous alues o speed and p oduc ion. 4. Conclusions Wi h he cu en pape , an IoT con ol sys em o a DRR machine has been de eloped which allows i s ope a ion o be imp o ed: An IoT sys em has been de eloped which allows o he ecep ion o use ul ope a ional in o ma ion o he DRR machine emo ely in o de o do an ope a ional diagnosis based on his in o ma ion, and, in his case, ca y ou calib a ion ope a ions. This IoT sys em allows o con igu a ion o he equipmen ia a emo e desk op. In he second place, he classi ica ion sys em has been made o wo k in an indus ial se ing a a nominal speed o he DRR machine (up o 2500 oli es pe minu e) h ough he use o a PC-CM1K connec ion which uses a mic ocon olle wi h an In el Cu ie o Teensy “ i ual po ” as an in e ace. In he hi d place, he aining and es ing o a neu al ne wo k based on a physical chip o classi y oli es in he eed chain o a pi ing machine has been success ully comple ed. I has been p o en ha he use o In el Cu ie o CM1K chips a e sui able o simple so ing (no mal, emp y, boa s), wi h e o a es o less han 5% wi h he In el Cu ie chip and less han 1% wi h he CM1K chip. An applica ion de eloped in Q C ea o has been used ha p e iously de ec s o i s s udy he cases wi h small pa s o oli es o double oli es p esen in he pocke and sepa a es hem om hose ha a e sen o he neu al chip ha co espond o no mal, emp y pocke s and boa s. One limi a ion o he sys em cu en ly de eloped is ha i only allows o he diagnosis o he DRR machine so ha he ope a o can ecalib a e i . In a u u e pape , he possibili y o ejec ing hose oli es which a e badly posi ioned (“boa ” oli es o doubles) o which ha e de ec s (small pieces) in he eeding chain will be p esen ed. In his way, “boa ” oli es o doubles being badly de-s oned will be a oided. Likewise, i will assu e ha he oli es which b eak upon en e ing in he eed chain causing small pieces do no ge o he de-s oning a ea, hus a oiding, on he one hand, he loss o p oduc (1–2%), and on he o he hand, he accele a ed de e io a ion o he punch needles and ha s o deboning. Au ho Con ibu ions: Concep ualiza ion, A.M.L.; alida ion, A.M.L., J.M.M.M., A.R.C., M.d.J.L., J.M.M.L., M.J.S. and A.L.P.; supe ision: A.M.L., J.M.M.M. and A.R.C.; in es iga ion, w i ing—o iginal d a p epa a ion, w i ing— e iew and edi ing, concep ualiza ion, so wa e, o mal analysis, esou ces, da a cu a ion, Figu e 20. Ins an aneous alues o speed and p oduc ion. 4. Conclusions Wi h he cu en pape , an IoT con ol sys em o a DRR machine has been de eloped which allows i s ope a ion o be imp o ed: An IoT sys em has been de eloped which allows o he ecep ion o use ul ope a ional in o ma ion o he DRR machine emo ely in o de o do an ope a ional diagnosis based on his in o ma ion, and, in his case, ca y ou calib a ion ope a ions. This IoT sys em allows o con igu a ion o he equipmen ia a emo e desk op. In he second place, he classi ica ion sys em has been made o wo k in an indus ial se ing a a nominal speed o he DRR machine (up o 2500 oli es pe minu e) h ough he use o a PC-CM1K connec ion which uses a mic ocon olle wi h an In el Cu ie o Teensy “ i ual po ” as an in e ace. In he hi d place, he aining and es ing o a neu al ne wo k based on a physical chip o classi y oli es in he eed chain o a pi ing machine has been success ully comple ed. I has been p o en ha he use o In el Cu ie o CM1K chips a e sui able o simple so ing (no mal, emp y, boa s), wi h e o a es o less han 5% wi h he In el Cu ie chip and less han 1% wi h he CM1K chip. An applica ion de eloped in Q C ea o has been used ha p e iously de ec s o i s s udy he cases wi h small pa s o oli es o double oli es p esen in he pocke and sepa a es hem om hose ha a e sen o he neu al chip ha co espond o no mal, emp y pocke s and boa s. One limi a ion o he sys em cu en ly de eloped is ha i only allows o he diagnosis o he DRR machine so ha he ope a o can ecalib a e i . In a u u e pape , he possibili y o ejec ing hose oli es which a e badly posi ioned (“boa ” oli es o doubles) o which ha e de ec s (small pieces) in he eeding chain will be p esen ed. In his way, “boa ” oli es o doubles being badly de-s oned will be a oided. Likewise, i will assu e ha he oli es which b eak upon en e ing in he eed chain causing small pieces do no ge o he de-s oning a ea, hus a oiding, on he one hand, he loss o p oduc (1–2%), and on he o he hand, he accele a ed de e io a ion o he punch needles and ha s o deboning. Senso s 2020,20, 1541 20 o 22 Au ho Con ibu ions: Concep ualiza ion, A.M.L.; alida ion, A.M.L., J.M.M.M., A.R.C., M.d.J.L., J.M.M.L., M.J.S. and A.L.P.; supe ision: A.M.L., J.M.M.M. and A.R.C.; in es iga ion, w i ing—o iginal d a p epa a ion, w i ing— e iew and edi ing, concep ualiza ion, so wa e, o mal analysis, esou ces, da a cu a ion, isualiza ion, p ojec adminis a ion: A.L.P. All au ho s ha e ead and ag eed o he published e sion o he manusc ip . Funding: This esea ch ecei ed no ex e nal unding. Con lic s o In e es : The au ho s decla e no con lic o in e es . Re e ences 1. San os, F.J. Siles New echnologies in able oli e p ocessing. G asas Acei es 1999,502, 131–140. [C ossRe ] 2. Madueño, A.; Line os, M.; Madueño, J. Sys em and P ocedu e Based on a Synch onism Senso o he De ec ion o Mal unc ions in Pi ing Machines Oli e and Filling Machines, Quan i ica ion and Op imiza ion o Pe o mance, Signaling, Moni o ing and Remo e Con ol. ES2529816A2. A ailable online: h ps: //pa en s.google.com/pa en /ES2529816A2/en (accessed on 2 Feb ua y 2020). 3. Tang, Y.; Li, L.; Wang, C.; Chen, M.; Feng, W.; Zou, X.; Huang, K. Real- ime de ec ion o su ace de o ma ion and s ain in ecycled agg ega e conc e e- illed s eel ubula columns ia ou -ocula ision. Robo . Compu . In eg . Manu . 2019,59, 36–46. [C ossRe ] 4. Chen, M.; Tan, Y.; Zou, X.; Huang, K.; Li, L.; He, Y. High-accu acy mul i-came a econs uc ion enhanced by adap i e poin cloud co ec ion algo i hm. Op . Lase s Eng. 2019,122, 170–183. [C ossRe ] 5. Nie, M.; Zhao, Q.; Xu, Y.; Shen, T. Machine Vision-based Apple Ex e nal Quali y G ading. In P oceedings o he Chinese Con ol and Decision Con e ence, Nanchang, China, 3–5 June 2019. 6. Lucas, A.; Madueño, A.; De J ó da , M.; Molina, J.; Ruiz, A. Cha ac e iza ion o he pe cen age o poo ly posi ioned oli es in pi ing, olling and illing machines o able oli es (DRR). In P oceedings o he X Cong esso Ibé ico de Ag oengenha ia, Huesca, Spain, 3–6 Sep embe 2019. 7. Lin, G.; Tang, Y.; Zou, X.; Li, J.; Xiong, J. In- ield ci us de ec ion and localisa ion based on RGB-D image analysis. Biosys . Eng. 2019,186, 34–44. [C ossRe ] 8. Lin, G.; Tang, Y.; Zou, X.; · Xiong, J.; Fang, Y. Colo -, dep h-, and shape-based 3D ui de ec ion. P ec. Ag ic. 2020,21, 1–17. [C ossRe ] 9. Lin, G.; Tang, Y.; Zou, X.; Xiong, J.; Li, J. Gua a. De ec ion and Pose Es ima ion Using a Low-Cos RGB-D Senso in he Field. Senso s 2019,19, 428. [C ossRe ] 10. Yang, F. Classi ica ion o apple su ace ea u es using machine ision and neu al ne wo ks. Compu . Elec on. Ag ic. 1993,9, 1–12. [C ossRe ] 11. Naga a, M.; Ba o, P.; Mi a ia, M.; Cao, Q.; Ki aha a, T. S udy on So ing Sys em o S awbe y Using Machine Vision (Pa 1). Jap. Soc. Ag ic. Mach. 2000,62, 100–110. 12. Beh oozi, N.; Ta akoli, T.; Ghassemian, H.; Hadi, M.; Banaka , A. Applied machine ision and a i icial neu al ne wo k o modeling and con olling o he g ape d ying p ocess. Compu . Elec on. Ag ic. 2013 ,98, 205–213. [C ossRe ] 13. Ga ica, G.; Bes b, S.; Ce onic, J.; Le anc, G. Oli e F ui s Recogni ion Using Neu al Ne wo ks. P oc. Compu . Sci. 2013,17, 412–419. [C ossRe ] 14. Mancuso, S.; Nicese, F.P. Iden i ying Oli e (Olea eu opaea) Cul i a s Using A i icial Neu al Ne wo ks. Am. Soc. Ho ic. Sci. 1999,124, 527–531. [C ossRe ] 15. Sun, D. Compu e Vision Technology o Food Quali y E alua ion, 2nd ed.; Academic P ess: Wal ham, MA, USA, 2016; pp. 273–350. 16. Diaz, R. Compu e Vision Technology o Food Quali y E alua ion, 2nd ed.; Academic P ess: Wal ham, MA, USA, 2016; pp. 351–367. 17. Bo le Inspec ion. Gene al Visions. 2013. A ailable online: h ps://www.gene al- ision.com/appno es/AN_ Bo leInspec ion.pd (accessed on 2 Feb ua y 2020). 18. Menendez, A.; Paille , G. Fish Inspec ion Sys em Using a Pa allel Neu al Ne wo k Chip and he Image Knowledge Builde Applica ion. AI Mag. 2008,29, 21. 19. Liu, Y.; Wei, D.; Zhang, N. Vehicle-license-pla e ecogni ion based on neu al ne wo ks. In P oceedings o he IEEE on In o ma ion and Au oma ion, Shenzhen, China, 6–8 June 2011. Senso s 2020,20, 1541 21 o 22 20. Sa da , S.; Tewa i, G.; Babu, K.A. A ha dwa e/so wa e co-design model o ace ecogni ion using Cognimem Neu al Ne wo k chip. In P oceeding o he IEEE on Image In o ma ion P ocessing, Shimla, India, 3–5 No embe 2011. 21. Da ies, M.; S ini asa, N.; Lin, T.; Chinya, G.; Cao, Y.; Choday, S.; Dimou, G.; Joshi, P.; Imam, N.; Jain, S.; e al. Loihi: A Neu omo phic Manyco e P ocesso wi h On-Chip Lea ning. IEEE Mic o 2018 ,38, 82–99. [C ossRe ] 22. Mo an, S.; Gaonka , B.; Whi ehead, W.; Wolk, A.; Macyszyn, L.S.; Iye , S. Deep lea ning o medical image segmen a ion—Using he IBM T ueNo h neu osynap ic sys em. In P oceedings o he SPIE Medical Imaging, Hous on, TX, USA, 6 Ma ch 2018. 23. Mo adi, S.; Qiao, N.; S e anini, F.; Indi e i, G. A scalable mul i-co e a chi ec u e wi h he e ogeneous memo y s uc u es o dynamic neu omo phic asynch onous p ocesso s (dynaps). IEEE T ans. Biomed. Ci cui s Sys . 2018,12, 106–122. [C ossRe ] [PubMed] 24. F enkel, C.; Le eb e, M.; Lega , J.; Bol, D. A 0.086-mm 2 12.7-pJ/SOP 64k-Synapse 256-Neu on Online-Lea ning Digi al Spiking Neu omo phic P ocesso in 28-nm CMOS. IEEE T ans. Biomed. Ci cui s. Sys . 2019 ,13, 145–158. 25. F ied, L. Making machine lea ning a duino compa ible: A gaming handheld ha uns neu al ne wo ks-[Resou ces_Hands On]. IEEE Spec . 2019,56, 14–15. [C ossRe ] 26. Lobache , I.; Male yk, R.; An oschuk, S.; Filiahin, D.; Lobache , M. In eg a ion o neu al ne wo ks in o sma senso ne wo ks. In P oceedings o he IEEE Xplo e, Kie , Uk aine, 24–27 May 2018. 27. Mi al, S. A Su ey on op imized implemen a ion o deep lea ning models on he NVIDIA Je son pla o m. J. Sys . A ch. 2019,97, 428–442. [C ossRe ] 28. Kim, J. New Neu omo phic AI NM500 and I s ADAS Applica ion. In AETA-2018 Recen Ad ances in Elec ical Enginee ing and Rela ed Sciences: Theo y and Applica ion; Lec u e No es in Elec ical, Enginee ing; Zelinka, I., B ands e e , P., T ong Dao, T., Hoang Duy, V., Kim, S., Eds.; Sp inge : Cham, Swi ze land, 2019; Volume 554, pp. 3–12. 29. CogniPa SDK o Ma lab. Gene al Visions. 2018. A ailable online: h ps://www.gene al- ision.com/ download/cp_sdk_ml/(accessed on 2 Feb ua y 2020). 30. Neu oMem USB Dongle. Gene al Visions. 2019. A ailable online: h ps://www.gene al- ision.com/ ha dwa e/usbdongle/(accessed on 2 Feb ua y 2020). 31. L ó pez Riquelme, J.A.; So o, F.; Sua d í az, J.; S á nchez, P.; Ibo a, A.; Ve a, J.A. Wi eless Senso Ne wo ks o p ecision ho icul u e in Sou he n Spain. Compu . Elec on. Ag ic. 2019,68, 25–35. [C ossRe ] 32. Ga cia, L.; Pa a, L.; Jimenez, J.M.; Llo e , J.; Lo enz, P. IoT-Based Sma I iga ion Sys ems: An O e iew on he Recen T ends on Senso s and IoT Sys ems o I iga ion in P ecision Ag icul u e. Senso s 2020 ,20, 1042. [C ossRe ] 33. U bano, O.; Pe les, A.; Ped aza, C.; Rubio-A aez, S.; Cas ell ó , M.L.; O ola, M.D.; Me cado, R. Cos -Eec i e Implemen a ion o a Tempe a u e T aceabili y Sys em Based on Sma RFID Tags and IoT Se ices. Senso s 2020,20, 1163. [C ossRe ] 34. Escola D í az, S.; Ca e e o P é ez, J.; Calde ó n Ma eos, A.; Ma inescu, M.C.; Be gua Gue a, B. A no el me hodology o he moni o ing o he ag icul u al p oduc ion p ocess based on wi eless senso ne wo ks. Compu . Elec on. Ag ic. 2011,76, 252–265. [C ossRe ] 35. Au oma ed-oli e-chain. The in e ne o Food & Fa m. 2020. A ailable online: h ps://www.io 2020.eu/ ials/ ui s/au oma ed-oli e-chain (accessed on 20 Decembe 2019). 36. De Joda , M.; Madueño, A.; Lucas, A.; Molina, J.M.; C á nales, A.R.; Madueño, J.M.; Jus icia, M.; Baena, M. Deep lea ning in oli e pi ing machines by compu e isi ó n. Compu . Elec on. Ag ic. 2020 ,171, 105304. [C ossRe ] 37. Hech -Nielsen, R. Theo y o he Backp opaga ion Neu al Ne wo k. In P oceedings o he In e na ional 1989 Join Con e ence on Neu al Ne wo ks, Washing on, DC, USA, 18–22 June 1989. 38. Google Co al Edge TPU. Google LLC. 2020. A ailable online: h ps://co al.ai/docs/accele a o /da ashee / (accessed on 15 No embe 2019). 39. In el ® Mo idius ™ Neu al Compu e S ick 2. In el Co po a ion. 2020. A ailable online: h ps: //www.in el.es/con en /www/es/es/design/p oduc s-and-solu ions/boa ds-ki s-and-modules/mo idius- neu al-compu e-s ick-2/ echnical-lib a y.h ml?g ouping= dc%20Con en %20Types&so = i le:asc (accessed on 10 Decembe 2019). Senso s 2020,20, 1541 22 o 22 40. N idia-Je son-Nano. N idia Co po a ion. 2020. A ailable online: h ps://www.n idia.com/es-es/ au onomous-machines/embedded-sys ems/je son-nano/(accessed on 15 Decembe 2019). 41. TM Tes Neu ons SimpleSc ip . Gene al Visions. A ailable online: h p://www.gene al- ision.com/ documen a ion/TM_Tes Neu ons_SimpleSc ip .pd (accessed on 2 Feb ua y 2020). 42. TM Neu oMem Technology Re e ence Guide. Gene al Visions. 2019. A ailable online: h ps://www.gene al- ision.com/documen a ion/TM_Neu oMem_Technology_Re e ence_Guide.pd (accessed on 2 Feb ua y 2020). 43. TM_CM1K_Ha dwa e_Manual. Gene al Visions. 2017. A ailable online: h ps://www.gene al- ision.com/ documen a ion/TM_CM1K_Ha dwa e_Manual.pd (accessed on 2 Feb ua y 2020). 44. Halgamuge, S.; Poechmuelle , W.; Glesne , M. An Al e na i e App oach o Gene a ion o Membe ship Func ions and Fuzzy Rules Based on Radial and Cubic Basis Func ion Ne wo ks. In . J. App ox. Reason. 1995 , 12, 279–298. [C ossRe ] 45. DS_CM1K. Gene al Visions. 2014. A ailable online: h ps://www.gene al- ision.com/da ashee /DS_CM1K. pd (accessed on 2 Feb ua y 2020). 46. Neu al-ne wo k. The Ma hWo ks, Inc. 1994–2017. A ailable online: h ps://es.ma hwo ks.com/solu ions/ deep-lea ning/con olu ional-neu al-ne wo k.h ml?s_ id=s ch i le (accessed on 2 Feb ua y 2020). 47. T ain Au oencode . The Ma hWo ks, Inc. 1994–2017. A ailable online: h p://es.ma hwo ks.com/help/nne / e / ainau oencode .h ml (accessed on 2 Feb ua y 2020). 48. T ain S acked Au oencode s o Image Classi ica ion. The Ma hWo ks Inc. 1994–2019. A ailable online: h ps://es.ma hwo ks.com/help/deeplea ning/examples/ ain-s acked-au oencode s- o -image- classi ica ion.h ml (accessed on 2 Feb ua y 2020). 49. Image Se Reposi o y. A ailable online: h ps://gi hub.com/To as86/Oli e-image-se (accessed on 18 Decembe 2019). 50. IDS Imaging De elopmen Sys ems GmbH. 2016. A ailable online: h ps://es.ids-imaging.com/s o e/ui- 1220se.h ml (accessed on 2 Feb ua y 2020). 51. QT C ea o . The Q Company. 2020. A ailable online: h ps://doc.q .io/(accessed on 18 Decembe 2019). 52. D opbox. D opbox Inc. 2020. A ailable online: h ps://www.d opbox.com/de elope s/documen a ion (accessed on 2 Feb ua y 2020). 53. B ainca d. Gene al Visions. 2017. A ailable online: h ps://www.gene al- ision.com/documen a ion/TM_ B ainCa d.pd (accessed on 2 Feb ua y 2020). 54. NM500 Chip. Gene al Visi ó n. 2019. A ailable online: h ps://www.gene al- ision.com/documen a ion/TM_ Neu oShield_Ge ingS a ed.pd (accessed on 2 Feb ua y 2020). 55. I 2 C (In e In eg a ed Ci cui ), Phillips Semiconduc o , 1982. A ailable online: h ps://en.wikipedia.o g/wiki/ I%C2%B2C (accessed on 2 Feb ua y 2020). 56. FT232RL USB UART IC. Fu u e Technology De ices In e na ional Limi ed. 2018. A ailable online: h ps://www. dichip.com/Suppo /Documen s/Da aShee s/ICs/DS_FT232R.pd (accessed on 20 Decembe 2019). 57. USB o Se ial Chip CH340. Spa kFun Elec onics. 2015. A ailable online: h ps://cdn.spa k un.com/ da ashee s/De /A duino/O he /CH340DS1.PDF (accessed on 19 Decembe 2019). 58. In el Cu ie Module. In el Co po a ion. 2020. A ailable online: h ps://a k.in el.com/con en /www/es/es/a k/ p oduc s/96282/in el-cu ie-module-in el-qua k-se-soc.h ml (accessed on 21 Decembe 2019). 59. Teensy 4.0 USB De elopmen Boa d. PJRC Elec onics P ojec s Componen s A ailable Wo ldwide. A ailable online: h ps://www.pj c.com/ eensy/(accessed on 20 Decembe 2019). 60. Google Remo e Desk op. Google LLC. 2020. A ailable online: h ps://suppo .google.com/ch ome/answe / 1649523?co=GENIE.Pla o m%3DDesk op&hl=es (accessed on 20 Decembe 2019). © 2020 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion (CC BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).