jiny2001/dcscn-super-resolution

Implements residual learning with skip connections and 1×1 convolutions (Network in Network) as the core architecture, enabling efficient feature extraction across local and global image regions. Supports multiple upsampling strategies including sub-pixel convolution (Pixel Shuffler), transposed convolutions, and self-ensemble techniques, with optional depthwise separable convolutions for reduced model complexity. Built on TensorFlow 2.0+ with flexible training pipelines that allow custom dataset integration, configurable model scaling (4-12 layers, 32-196 filters), and pre-trained weights for immediate evaluation on standard benchmarks (Set5, Set14, BSD100).

714 stars. No commits in the last 6 months.