depth-model-final

Residual Channel Attention Network (RCAN)

Pretrained deep residual network leveraging channel attention modules (RCAB) and pixel-shuffle upscaling for image restoration and 4x super-resolution under degradations.

Architecture Details

• Head (Feature Extraction): 3x3 Conv layer mapping 3 input RGB channels to 64 feature dimensions.
• Body (Deep Residual Attention): 4 Residual Groups (ResidualGroup), each containing 4 Residual Channel Attention Blocks (RCAB) equipped with squeeze-and-excitation channel attention (reduction ratio = 16).
• Upsampling: Sub-pixel convolution via nn.PixelShuffle(scale=4).
• Tail (Reconstruction): Final 3x3 Conv layer outputting 3-channel RGB images.

Training Configuration

{
  "model_architecture": "RCAN",
  "scale_factor": 4,
  "num_channels": 3,
  "num_features": 64,
  "num_groups": 4,
  "num_blocks_per_group": 4,
  "reduction_ratio": 16,
  "patch_size": 48,
  "batch_size": 16,
  "epochs": 100,
  "loss_function": "L1Loss",
  "optimizer": "Adam",
  "initial_learning_rate": 0.0001,
  "scheduler": "CosineAnnealingLR (T_max=30, eta_min=1e-6)",
  "target_metric": "PSNR"
}