Edge-Preserving Generative Adversarial Networks for Enhanced Underwater Image Restoration

Edge-Preserving Generative Adversarial Networks for Enhanced Underwater Image Restoration Algorithms of each module of our proposed work is as follows: Algorithm 1 Training Procedure 1. Start with paired training samples. 2. Initialize generator , multi-scale discriminator , and set hyperparameters . 3. Set optimizer: and and learning rate and parameters . 4. Set noise . 5. Initialize . Training Loop 6. For epoch = 1 to N do 7. Set generator and discriminator to training mode. Initialize epoch generator loss . 8. For each minibatch do Discriminator Update (every 25 batches) 9. If (batch index mod 25) = 0 then, 10. Freeze generator: compute fake image without gradient: (no gradient). 11. Add Gaussian noise to real and fake images: . 12. Compute discriminator outputs (multi-scale): . Each branch returns a probability in . 13. Binary cross-entropy losses: 14. Total discriminator loss: . 15. Update discriminator parameters via gradient descent: . 16. End if. Generator Update (every batch) 17. Compute generated image with gradient: . 18. Content loss: 19. Adversarial loss (generator tries to fool ): 20. Perceptual loss (simplified MSE): . 21. Edge loss (edge consistency via Canny): , where is the same Canny operator used for input. 22. Weighted total loss: . 23. Update generator parameters: . 24. End for (training minibatch loop). Validation Phase 25. Set generator and discriminator to evaluation mode. 26. Initialize empty lists for PSNR and SSIM values. 27. For each validation pair do: i. Compute enhanced output: . ii. Compute . iii. Compute SSIM per standard formula: iv. Append PSNR and SSIM to their respective lists. 28. End for loop. 29. Compute mean PSNR and mean SSIM: . Model Checkpointing 30. If validation PSNR > best PSNR then, save as best weights and set best PSNR to current PSNR. 31. End if 32. End for (epoch loop). 33. End. Algorithm 2: Edge Map Extraction 1. Start with RGB image , or 2. Detach tensor from autograd graph to avoid gradient tracking: 3. Determine if input is batched: a. If has 4 dimensions batched input b. Else single image. Case 1: Batched Input (B, 3, H, W) 4. Initialize empty list for edge maps: . 5. For each image in batch : 6. Convert PyTorch tensor to NumPy array and reorder channels: 7. Undo normalization from : 8. Convert RGB to grayscale using simple averaging: 9. Apply Canny edge detector with Gaussian smoothing : 10. Convert edge map to PyTorch tensor on original device: 11. Append to list . 12. End for. 13. Stack edge maps into batch form and add channel dimension: 14. Return . Case 2: Single Image (3, H, W) 15. Convert tensor to NumPy array and reorder channels: 16. Undo normalization from to : 17. Convert to grayscale: 18. Apply Canny edge detector: 19. Convert to PyTorch tensor and add channel dimension: 20. Return . 21. End Algorithm 3: Deformable Convolution Block 1. Start with input feature map . 2. Compute learned sampling offset using a convolution: , where represents offsets for each of the kernel positions. 3. Apply deformable convolution using offsets: , which samples spatially shifted positions where . 4. Apply Group Normalization: . 5. Apply LeakyReLU activation . 6. Return output feature map . 7. End. Algorithm 4: Residual Block 1. Start with input feature map . 2. Apply identity skip connection: . 3. Apply first convolution with kernel and padding = 1: . 4. Apply Group Normalization: . 5. Apply LeakyReLU activation with slope 0.2: . 6. Apply the second convolution with kernel and padding = 1: . 7. Apply Group Normalization: . 8. Add skip connection (residual addition): . 9. Return output feature map . 10. End. Algorithm 5: Attention Module 1. Start with input feature map , where is the number of the channels. 2. Apply convolution to compute attention logits:. 3. Apply element-wise sigmoid activation to form attention weights: , where . 4. Apply channel-wise and spatial-wise modulation by multiplying input and attention map element-wise: . Where denotes element-wise multiplication. 5. Return attention-refined feature map . 6. End. Algorithm 6: Multi-Scale Discriminator Forward Pass 1. Start with input image . 2. Apply first convolution with stride 2: 3. Apply second convolution: 4. Apply final convolution to reduce to 1 channel: 5. Output of branch 1 is . 6. Downsample input using average pooling:. 7. Apply first convolution on downsampled input: 8. Apply second convolution: 9. Apply final convolution to reduce to 1 channel:. 10. Output of branch 2 is . 11. Return the list of discriminator outputs at two scales: . 12. End. Algorithm 7: Enhanced Generator Architecture 1. Start with input tensor formed by concatenating RGB distorted image with its edge map. Encoder Path 2. Apply first convolution (downsample to 128×128): 3. Apply Deformable Block: 4. Apply third encoder convolution (downsample to 64×64): Bottleneck 5. Initialize bottleneck representation: . 6. Pass through 6 Residual Blocks: 7. Apply Attention Module to refine spatial focus: Decoder Path 8. Apply first transposed convolution (upsample to ): 9. Apply Deformable Block: Output Layer 10. Apply activation to constrain output to the range . 11. Return enhanced output image ŷ. 12. End.