Supporting a Celtic coin die study with image matching and hierarchical clustering for creating a social networks analysis

Supporting a Celtic coin die study with 1 image matching and hierarchical 2 clustering for creating a social 3 networks analysis 4 5 Sebastian Gampe*1, Karsten Tolle1, 6 Caroline von Nicolai2, Markus Möller2 7 8 1 Big Data Lab, Goethe-Universität – Frankfurt am Main, Germany 9 2 Juniorprofessur für Digitale Archäologie Mitteleuropas, Universität Leipzig – Leipzig, Germany 10 11 *Corresponding author 12 Correspondence: [email protected] 13 14 15 ABSTRACT 16 Studying coin dies, particularly Celtic ones, is a challenging and time-consuming task 17 for numismatists. To support them in this task, we implemented a coin studies pipeline 18 in ClaReNet, which we have further improved in our current DeReNum project. The 19 aim is to maintain human involvement and oversight, which is particularly important 20 given the challenges posed by Celtic coins for die studies, as detailed in the paper. 21 Notably, our focus is on silver coins, which are often in worse condition than gold coins. 22 Extremely poorly preserved coins can be eliminated by a pre-sorting step supported 23 by the Deepcluster method. The numismatist will also be involved in this step and will 24 have the final decision. Pre-sorting with Deepcluster is particularly useful when the 25 number of coins is high or the state of preservation varies greatly. As a result of the 26 pipeline, which will be explained in detail in this paper, the user will receive a 27 dendrogram ordering the coin images (one image per coin side) and proposing 28 clusters of coins likely to have been produced from the same die. Using the open-29 source tool Orange Data Mining, we also propose a way to visualise the dendrogram 30 and provide a working environment. 31 Within the pipeline, we are currently comparing two image-matching algorithms: ORB 32 and DISK. Based on tests with a dataset using a ground truth from the ClaReNet 33 project, DISK, which is based on convolutional neural networks, performed better than 34 ORB. However, another test showed that ORB can handle coins with different 35 orientations more effectively. 36 In this paper, we will present examples of new dies identified using our pipeline. 37 However, we are still in the early stages of the project and are currently collecting 38 material for analysis. 39 Keywords: Machine Learning, Image Matching, Convolutional Neural Networks, Clustering, 40 Celtic Coins 41 42 Introduction 43 In our new project, ‘De Retibus Nummorum’ (DeReNum)1, we are dedicated to researching the 44 Celtic coinage systems and the associated coins, which often feature highly abstract, ornamental 45 patterns, using traditional numismatic and computer-assisted methods. 46 47 The objectives of our project are: 48 49 1. Reconstruction of the origin and development of the Celtic monetary economy; 50 2. Examination of the typological distribution and die links of selected coin series; 51 3. Research into the connections and dependencies in the late Celtic communication 52 networks and economic areas. 53 54 In order to achieve these goals, we will examine selected and representative celtic coin series 55 from central Europe. To this end, we must first collect as many data sets as possible on these 56 coin series from archaeological institutions and, with the appropriate permission, integrate them 57 into the virtual union catalogue “Online Celtic Coinage” 2 (Wigg-Wolf et al. 2025), that was 58 developed within the previous project ClaReNet3. A die study will then be carried out on the 59 collected coin images, which we intend to support with algorithmic and AI-based methods such as 60 the pipeline presented here. The collected data and the dies will hereafter be incorporated into a 61 social network analysis, which can provide further insights into the trading networks of the Celts in 62 Central Europe. 63 In the following, the individual data sets from the various institutions are combined into one 64 large data set, each covering a specific series of coins, such as Bushel quinarii (Büschelquinare), 65 the Mussle staters (Muschelstatere), or the so-called Celtic small silver. In addition to image rights, 66 we have to deal with varying image qualities of the coin images and different types of databases. 67 The data set of Büschelquinare coins used in this paper currently comprises around 1,800 coins 68 from approximately 80 different sites, mainly distributed across southern Germany, Austria, 69 Switzerland, the Czech Republic, and Slovakia. The coins were minted between approximately 70 150 and 25 BC and feature the eponymous bushels on the front, and a stylised horse in motion on 71 the back (Fig. 1 right). 72 73 Figure 1 – Two coins with obverse and reverse from our dataset. The coin on the 74 left still shows a clearly recognisable head and belongs to the prototype in Allen's 75 typology. The coin on the right has been classified as type E and shows the already 76 very abstract bushels. Both coins show a horse with additional markings on the 77 reverse. (Photos left: St. Friedrich, Archäologische Staatssammlung München OCC 78 no. 2467; Photos right: N.N., Staatliche Münzsammlung München OCC no. 3095) 79 They weigh approximately 1.2 to 1.8 g, and have a diameter of approximately 12 mm. The 80 coins are either silver coins or coins that appear to be silver, being plated, which means they have 81 an iron or copper core. The silver content is then concentrated only on the visible surface. One 82 particular challenge posed by these coins is the development of the image on the obverse, which 83 is believed by researchers to have evolved from a head in profile to a highly abstract representation 84 1 DeReNum project blog: https://derenum.hypotheses.org/. 2 Online Celtic Coinage catalogue: https://occ.dainst.org/. 3 ClaReNet project blog: https://clarenet.hypotheses.org/. of various articulated bushels (Fig. 1). This is reflected above all in the five different typologies, 85 which contain varying numbers of subtypes and associated datings. We currently use Derek F. 86 Allen's typology (Allen 1978), because other typologies are based on it and can be subordinated 87 to it, whereas the other typologies ((Kellner 1990, Ziegaus 2000, Brandt & Irlinger 2002; Nick 2012) 88 are not compatible with each other. Due to the typical characteristics of archaeological material, 89 this classification naturally does not result in subtypes of equal size in terms of the number of coins. 90 However, since we do not want to train a classification model at this stage, an underrepresentation 91 of some types is currently not problematic (c.f. Gampe & Tolle 2024). In addition, these coins also 92 present other challenges, such as varying degrees of preservation and restoration, as well as 93 incomplete die impressions on many coins. Since these partial impressions are a common feature 94 of Celtic coins, and were caused not only by die slipping but also by blanks that were too small for 95 the die, this incompleteness did not seem to be a reason for the mint masters at the time not to put 96 these coins into circulation. Nonetheless, these factors not only make it difficult for numismatists 97 to find die links, but also for our coin die pipeline to deal with. However, even poorly preserved 98 coins are relevant to our project, as we need a data set as large as possible to test our tools and 99 approaches. 100 Related Work 101 In recent years, automated ancient coin die studies have been developed for various coin 102 series. So far, Greek and Roman coins in particular have received the most attention. This was 103 made possible by keypoint-based image matching technology, which allowed coin images to be 104 compared with each other based on specific characteristics, known as keypoints. This technology 105 was developed at the end of the last millennium by D. G. Lowe, who designed the SIFT (Scale 106 Invariant Feature Transform: Lowe 1999) algorithm. These algorithmic methods were the standard 107 in image matching for a long time, and have been further developed over the years by approaches 108 such as SURF (Speeded Up Robust Features: Bay & Van Gool 2006) and ORB (Oriented FAST 109 and Rotated BRIEF: Rublee et. al. 2011). In recent years, machine learning methods have also 110 been incorporated to further increase the performance of these matchers. Deep learning 111 approaches such as convolutional neural networks (DISK: Tyszkiewicz et al. 2020) and 112 transformer models (SuperGlue: Sarlin et al. 2019) are used here. 113 Automated die studies on ancient coins have not yet been attempted on a large scale. The 114 CADS approach (Computer-Aided Die Study) was the first attempt in 2020 to support numismatists 115 in this challenging work (Taylor 2020). The pipeline was very similar to ours, as it used ORB and 116 hierarchical clustering. However, this approach was developed and tested on very well-preserved 117 Greek coins. A year later, Heinecke et al. published a study that used a similar system with 118 Gaussian Process (GP) keypoints and Bayesian microclustering on Neronian silver denarii 119 (Heinecke et al. 2021). In 2024, an approach was developed using a deep learning-based XFeat 120 model for keypoint detection and subsequent clustering with the label propagation clustering 121 method (Cornet et al. 2024). This approach was also tested on Greek coins. 122 A new study was recently published that also uses a keypoint system to create an overlay from 123 two coins and then uses an SSIM score (Structural Similarity) to find similarities on both surfaces 124 (Labedan et al. 2025). The numismatic basis for this study was a hoard of Roman coins. It is 125 therefore clear that keypoint-based image matching is not a new approach for ancient coins. 126 However, the mentioned studies cover well preserved Greek and Roman coins. Celtic coins, which 127 differ significantly in design from Greek and Roman coins often possess incomplete impressions, 128 thus posing a completely different challenge, were not tested. Therefore, our pipeline focuses on 129 the numismatic legacy of the Celts. 130 Deepcluster for pre-sorting 131 The first step in our pipeline consists of pre-sorting the coin images using the unsupervised 132 Deepcluster approach (Caron et al. 2019). As we had already had good experiences with 133 Deepcluster for pre-sorting coin images in the ClaReNet project, it made sense to use it again 134 (Deligio et al. 2024). We would like to remove severely damaged coins whose images are heavily 135 worn or corroded from the dataset, as we want to increase the chance of linking very similar coins 136 with each other in the subsequent pipeline steps. However, we want to proceed carefully so that 137 we do not remove too many coins in our already limited data set. 138 139 140 Figure 2 - Sections from two clusters after applying Deecluster. The upper cluster 141 contains better-preserved coins of various types, while the lower cluster consists of 142 coins that are difficult to identify. (Photos top: N.N., Archäologisches Museum 143 Kehlheim OCC No.2267; St. Friedrich, Archäologische Staatssammlung München 144 OCC nos. 2409, 2412, 2462, 2463, 2502, 2508, 2654; C. von Nicolai, Stadtarchiv 145 Ingolstadt OCC nos. 2766, 2768, 2789, 2228, 2849, 2852, 2864, 2882, 2920, 2927. 146 Photos bottom: St. Friedrich, Archäologische Staatssammlung München OCC nos. 147 2510, 2550, 2555, 2563, 2564, 2566, 2579, 2590, 2592, 2597, 2617, 2632, 2657, 148 2686, 2710, 2722; C. von Nicolai, Stadtarchiv Ingolstadt OCC no. 2862) 149 The Deepcluster model uses components of a convolutional neural network (CNN) to extract 150 features from the coin images in the first step. This creates a feature vector for each image in the 151 dataset. These are then clustered using a k-means algorithm. The individual clusters are 152 subsequently given a pseudo label, which is used in the next step as a class for training the CNN. 153 This improves the selection of important features and the classification of clusters with each step 154 of the training (Caron et al. 2019). 155 We applied Deepcluster separately to the front and back of our bushel series data set of 1.800 156 coins. The number of clusters (k) was set in advance at 45 for the obverses and 40 for the reverses. 157 This ensured that the number of images in the cluster did not become too large or too small. 158 Deepcluster itself attempts to make the clusters the same size anyway. Due to this behaviour and 159 the fact that the clusters have to be evaluated manually after the run, it is not necessary to find an 160 optimal number for k. The training was carried out with a learning rate of 0.05, a batch size of 128, 161 and a duration of 200 epochs for both sides. As a result, we obtained many coin image clusters 162 whose surface had a very similar appearance. However, it contained many coins of different 163 subtypes of the Allen typology, both on the obverse and reverse sides. (Fig. 2 top). 164 After manually reviewing the entire result, we were also able to identify clusters that contained 165 only coins in very poor condition (Fig. 2 below). This meant that a total of 129 coin obverse images 166 from five clusters and 79 reverse images from four clusters could be removed from the dataset. 167 Searching for poorly preserved coins in very large sets of images with very different looking coins 168 would have made this very tiring for the eyes, and therefore also much more time-consuming. 169 Coin die studies pipeline 170 Our coin die studies pipeline is designed to make it easier for numismatists to find dies in a 171 large dataset of coin images. This means the domain expert would remain in the loop to finally 172 decide. We do not envision generating a 100% automatic approach. In ancient times, coins were 173 produced using two dies, the upper one (reverse) being movable and the lower one (obverse) 174 fixed. This meant that the upper die wore out faster than the lower one, and had to be replaced 175 after a certain number of minting operations. Thus, the lower die could have been used with two 176 or more upper dies. This resulted in die links, which can be reconstructed from the coins found. 177 Reconstructing and understanding these connections enables numismatists to establish a relative 178 chronology for a coin series. Since there are generally no written sources in prehistoric 179 archaeology, this is often the only way to date celtic coins. To do this, however, one first has to 180 identify the dies themselves. This can be a very time-consuming task, as a large number of coins 181 must be compared with each other in terms of their similarity and, above all, the details of the coin 182 image. To speed up this process, our idea for the ClaReNet project was to use keypoint extraction 183 and image matching algorithms that are capable of evaluating the similarity of coin images. At that 184 time, we tested the approaches already available for finding coin dies, which also worked with 185 image matching. However, these methods had only been tested on Roman and Greek coins. Our 186 Celtic coins differ from these in that their motifs often appear very abstract and frequently lack 187 legends. After several tests failed to produce promising results, we decided to develop our own 188 pipeline. 189 There are currently two different versions of our pipeline, which differ in terms of the algorithms 190 used and the number of preprocessing steps. Both versions share the following basic structure 191 (Fig. 3): First, a five-step preprocessing is applied to the coin images. This involves removing the 192 colour from the image, improving the contrast, reducing the image sharpness and, finally, removing 193 as much of the coin's edge as possible. This serves to minimise bias as far as possible. The bias 194 primarily involves components of the coin, such as corrosion or abrasion damage, and the coin 195 edge, which can distract the image-matching algorithm from the actual coin image. 196 197 Figure 3 – Schematic representation of both coin die studies pipelines. (Graph: S. 198 Gampe, Big Data Lab, Photos top: N.N. Staatliche Münzsammlung München OCC 199 no. 3095; C. von Nicolai, Deggendorf Kreisarchäologie OCC no. 3372. Orange 200 images: Orange Data Mining. Coin photos top cluster: C. von Nicolai, Stadtarchiv 201 Ingolstadt OCC no. 2923, 2884; St. Friedrich, Archäologische Staatssammlung 202 München OCC nos. 2424, 2418, 2450, 2456; C. von Nicolai, Staatliche 203 Münzsammlung München OCC no. 2975. Coin photos bottom cluster: N.N., 204 Salzburg Museum OCC no. 3161; St. Friedrich, Archäologische Staatssammlung 205 München OCC nos. 3602, 2468, 3679; N.N, Landesmuseum Württemberg Stuttgart 206 OCC no. 2277; C. von Nicolai, Stadtmuseum Ingolstadt OCC no. 3502) 207 The image matching algorithm then attempts to extract so-called keypoints on each image in 208 the data set. These are special features of the coin, such as edges or lines on the coin relief. The 209 keypoints are stored as a set so that they are available again in the next step. The key points of 210 two images are then compared in pairs. Each key point and its immediate surroundings are 211 examined for similarities with the key points and their surroundings on other coins. As soon as the 212 algorithm identifies two key points as a match, this is stored. When such matches for two coin 213 images are visualised, the key points are represented by lines connecting them (Fig. 3). These 214 lines should be as horizontal as possible, as only keypoints in the same regions of the coins should 215 be made when the coin image is aligned in the same way. Finally, the matches found for a coin 216 pair are counted and entered into a table with n*n (n = number of coins in the data set) fields. In 217 the end, this table contains the number of matches for each coin pair in the data set. Each row of 218 this table shows the similarity of one coin to all other coins in the data set. Mathematically speaking, 219 such a row forms a vector and can therefore be easily compared with the other vectors in the table. 220 In the penultimate step, the distances between all vectors in the table are now calculated using 221 an algorithm. The result of this calculation is a new table whose individual fields each contain the 222 summed distance values for all entries of a pair of vectors. In the final step, the distance table is 223 divided into different clusters using an agglomerative hierarchical clustering algorithm. First, all 224 coin vectors are treated as individual clusters. Then, the two vectors closest to each other are 225 iteratively merged into a new cluster. The distance table for the new cluster is then updated. There 226 are various ways of calculating distance in this clustering (complete, single, average, weighted, 227 and Ward). Within the Orange Datamining tool, the user (the domain expert) is also free to switch 228 between them in order to experiment and see their effect on the results. The merging is repeated 229 until all coin vectors are combined into a single cluster. The result of this process is ultimately 230 presented in a so-called dendrogram. This is a tree structure that organises all data points 231 hierarchically and can thus represent the sequence of cluster mergers through its subtrees. Each 232 subtree symbolises the merger of two or more coins into a new cluster. The closer the individual 233 coin images are connected to each other through the tree structure, the more similar the coins 234 depicted should be and the higher the chance that these coins were produced from the same die. 235 The tree structure should therefore help the numismatists to find new dies in our data set. 236 Both versions of the pipeline differ in terms of the algorithms used for image matching. The first 237 version from the ClaReNet project (Deligio et al.2024) uses the ORB (Oriented FAST and Rotated 238 BRIEF: Rublee et. al. 2011) algorithm, while the second variant, developed by Markus Fiedler in 239 his master's thesis (Fiedler 2025), uses the DISK (Discrete Keypoints: Tyszkiewicz et al. 2020) 240 algorithm. For both variants, we use standardised preprocessing, in which the following algorithms 241 are executed sequentially: 1. greyscaling, 2. histogram equalisation, 3. denoise, 4. histogram 242 equalisation and 5. circle crop (Fig. 3)4. 243 ORB and DISK differ in the way key points are detected in an image. ORB uses a purely 244 algorithmic approach and primarily searches for noticeable contrast changes and visible corners. 245 The FAST (Features from Accelerated Segment Test) detector is used for this purpose, which 246 considers a circle of surrounding pixels with a radius of 3 pixels (Bresenham-circle) for each point 247 in the image. This means that these 16 pixels are checked to see whether they are brighter or 248 darker than the centre pixel. If a threshold value (usually 12 pixels) is exceeded, the point is 249 recognised as a corner, for example (Rublee et. al. 2011). DISK, on the other hand, looks for 250 specific structures or patterns in the image. It uses convolutional neural networks (CNN) from the 251 field of machine learning and can therefore also detect key points in areas of an image with 252 smoother transitions which are therefore less visually striking in geometric terms. A CNN pre-253 trained on a large image dataset is used for this purpose. Here, local features are extracted from 254 the image by several convolutional layers and stored as feature maps. These are passed on to the 255 next layer in reduced resolution (pooling). After the last convolution layer, a score is now calculated 256 4 The first version of the pipeline originally also had a five-step preprocessing with a different algorithm (total variation regularisation). However, for better comparability, we decided to only use the variant from Fiedler 2025. for each point on the feature map, indicating the probability of a keypoint being present. If a 257 threshold value is exceeded, a keypoint is then output (Tyszkiewicz et al. 2020). 258 The description of keypoints and their comparison in two different images also differs between 259 the two algorithms. ORB uses BRIEF (Binary Robust Independent Elementary Features) 260 descriptors, which construct a binary string by comparing the brightness of several pixel pairs in 261 an area around the keypoint. These binary chains are then compared when matching two 262 keypoints on different images. The Hamming distance, which counts the number of different 263 positions in the chains, is used for this purpose. The smaller this sum is, the greater the similarity 264 between two keypoints is (Rublee et. al. 2011). DISK uses a so-called feature vector, which is 265 created at the end of the convolutional layers, to describe the keypoints. These vectors, which 266 describe points in high-dimensional space, are then compared for two keypoints on different image 267 pairs using Euclidean distance. The smaller the distance between two keypoints, the more likely it 268 is that they are a match. At the end of the matching process for two images, the paired keypoints 269 are recorded. While ORB simply counts the number of matches, DISK's score is based on the 270 quality of the matches rather than their number (Tyszkiewicz et al. 2020). 271 The table as a csv file with the numbers or matching scores is then loaded into the Orange 272 Data Mining app5. There, widgets with various functions are assembled into a workflow. This 273 workflow looks like this for our coin images: First, the csv file is loaded into the programme and 274 then displayed as a table. In the following, the paths to the image folders can be adjusted if they 275 differ from the paths in the csv file. The table can then be displayed again for checking. Now the 276 features from the table are selected for distance calculation in the next widget. Due to testing and 277 the different ways in which the matching scores are specified by both algorithms, the Pearson 278 distance metric is used for ORB and the Spearman distance metric for DISK (Fiedler 2025). After 279 the calculation, hierarchical clustering is performed and you can now view and select the different 280 subtrees in this widget. The coins from these selected subtrees are then displayed in the last 281 widget. These last two widgets are used by our numismatists to locate the dies. It was therefore 282 necessary to briefly instruct them in how to use this workflow. 283 Both pipelines were implemented using Jupyter Notebook, Python and Orange Data Mining. 284 The first pipeline uses OpenCV's ORB algorithm6, while the second pipeline uses Kornia's DISK 285 implementation7. The running times are as follows: Preprocessing the second pipeline takes just 286 a few minutes. This represents a significant improvement compared to the first pipeline, where this 287 process took several hours. Keypoint detection and matching take about eight hours for our dataset 288 of approximately 1,560 (after Deepcluster pre-sorting) images (both sides of the coins) with a 289 nVidia RTX 3090 GPU. This is mainly due to the fact that both ORB and DISK have to compare 290 the keypoints of each image with every other image in the dataset during the matching step. In the 291 worst case, this results in an asymptotic quadratic runtime (O(n²)). However, optimisations in the 292 respective algorithms help to reduce this runtime. In contrast, distance calculation and hierarchical 293 clustering in Orange are performed in a few seconds for our data set. 294 Table 1 – Number of clusters and coins contained in the respective clusters (#) for 295 70 % and 100% consistency 296 297 In his master's thesis, M. Fiedler was able to show that the second version of the pipeline, using 298 the DISK algorithm, was able to improve performance on a dataset from the ClaReNet project 299 (Tab. 1). For this dataset of Corosolitae staters from the hoard of Le Catillion II, there is a die study 300 by Philip de Jersey with approximately 1,300 coins. In the first version of the pipeline, the number 301 5 Orange Data Mining: https://orangedatamining.com/. 6 OpenCV’s documentation of ORB: https://docs.opencv.org/3.4/d1/d89/tutorial_py_orb.html. 7 Kornia’s documentation of DISK: https://kornia.readthedocs.io/en/latest/feature.html#kornia.feature.DISK. of clusters generated that contained at least 70% coins with the same die was counted. A total of 302 208 such clusters were found (Deligio et al. 2024). The second version of the pipeline increased 303 this number to 226. If the proportion of coins with the same die is increased to 100% per cluster, 304 then the DISK pipeline not only created 26 new clusters, but also more than doubled the number 305 of coins contained in the clusters (Fiedler 2025). 306 Due to these significantly better results, we decided to primarily use only the second pipeline 307 and to use the ORB pipeline only for comparisons. However, during an experiment with the DISK 308 pipeline, we discovered a problem with rotation invariance. ORB is claimed to be rotation invariant 309 (Rublee et al. 2011) while DISK is claimed to be very robust against rotations (Tyszkiewicz et al. 310 2020). In an experiment with golden mussle staters from Manching (Ziegaus 2013), we rotated 311 each of these coins in five-degree increments until a full rotation was achieved. These 72 new 312 images were used together with 286 other mussle staters as input for both pipelines. After running 313 both pipelines, the result was surprising. While the ORB pipeline placed all images of the rotated 314 coin in pairs in a subtree and then combined these subtrees into larger subtrees consisting only of 315 these coin images, this was not the case with the DISK pipeline. Although some coins were also 316 linked in pairs at the lowest levels of the dendrogram, a total of eleven merges of subtrees were 317 required before all rotated coins were in one cluster. However, this cluster then also includes all 318 coin images from the rest of the data set. With the ORB pipeline, only six subtree levels were 319 needed to merge all rotated images in the dendrogram, whereby this cluster then also consists 320 only of these images. This clearly shows that the ORB algorithm can handle rotated coin images 321 much better than DISK. However, since DISK performed better on the dataset with the Coriosolitae 322 staters and the problem with rotated coins is rather negligible due to the manual alignment of all 323 images in our datasets, we will continue to use the DISK pipeline primarily in our project. The 324 experiment shows that even newer approaches, which appear to offer better performance, cannot 325 always handle all eventualities in archaeological material. We are currently planning to have an 326 evaluation of the two image matching algorithms carried out as part of a student's master thesis. 327 The aim will be to find out in which cases it is better to use one algorithm and when the other, in 328 order to optimise our results with regard to the die study. In addition, tests will be carried out to 329 determine whether the two algorithms can be combined in a meaningful way. 330 Preliminary results 331 As we are still in the early stages of the project, we would like to show only a few preliminary 332 results of the pre-sorting with Deepcluster and the coin die study pipeline here. In our opinion, pre-333 sorting with Deepcluster was successful, as over 200 images with coins in very poor condition 334 could be sorted out with manageable effort. The most time-consuming task here was running the 335 Deepcluster algorithm, which took about two hours for each side of the bushel quinar data set. For 336 verification purposes, the coins already sorted out with Deepcluster were used once again as input 337 for our coin studies pipeline, together with the rest of the data set. We found that the sorted-out 338 coins were very often contained in a subtree with the better-preserved coins. However, due to their 339 poor condition, it is not really possible to determine that they originate from the same die. There 340 were also very few subtrees that consisted predominantly of these poorly preserved coins and 341 could therefore have been neglected in the evaluation. This test then confirmed our assessment 342 that sorting out items, which could also have been done manually but would have taken 343 considerably more time, did not have a negative impact on the result. 344 The numismatic assessment of the subtrees of our coin studies pipeline, on the other hand, 345 is still a work in progress. However, some coins with identical dies have already been discovered 346 with the help of the dendrogram. As examples, here are two coins from a front cluster and four 347 coins from a back cluster, which we believe were each produced with the same die (Fig. 4). 348