Single Image Super Resolution with Deep Residual Networks

Engineered high-performance convolutional architectures for high-fidelity single image super-resolution (SISR), with applications ranging from photographic benchmarks to satellite earth observation.

Technical Implementation & Results

• Deep Residual Network Architecture (EDSR):
  • Constructed an Enhanced Deep Super-Resolution model utilizing 32 deep residual blocks with removed batch normalization layers to preserve range flexibility and reduce memory overhead.
  • Employed sub-pixel convolution (pixel-shuffle) layers for efficient spatial feature upsampling.
• Mixed-Precision Training & Memory Optimization:
  • Leveraged PyTorch Automatic Mixed Precision (AMP) (FP16/FP32) to double training throughput while maintaining numerical stability.
  • Implemented dynamic spatial patch tiling during backpropagation and inference, effectively bounding peak VRAM allocation to prevent out-of-memory errors on large input tensors.
• Benchmark Performance:
  • Trained over 290,000 iterations using an (L_1) pixel-loss objective, attaining a 34.84 dB PSNR and 0.939 SSIM on the standard DIV2K validation dataset.
  • Implemented and benchmarked an SRCNN on the Sen2Venus high-resolution satellite imagery dataset, achieving 37.47 dB PSNR in reconstructing non-linear high-frequency spatial features from low-resolution multi-spectral bands.