Paper Review: Efficient Sub-Pixel Convolutional Neural Network

 

Paper arxiv link: Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network

Efficient Sub-Pixel Convolutional neural Network (ESPCN)

Efficient_Sub_Pixel_Convolutional_Neural_Network/networkstructure.jpg

Figure 1. The proposed efficient sub-pixel convolutional neural network (ESPCN), with two convolution layers for feature maps extraction, and a sub-pixel convolution layer that aggregates the feature maps from LR space and builds the SR image in a single step.

Contrast to previous works

  • super-resolve only at the end of the network (efficient sub-pixel convolution layer) → Eliminate the need to perform super-resolve in high resolution

Efficient sub-pixel convolution layer

  • Last layer of ESPCN
  • Rearrange tensor of \(H \times W \times C \cdot r^2\) to \(rH \times rW \times C\) tensor like Figure 1.
  • Implemented at TensorFlow as depth to space and PyTorch as PixelShuffle

Advantages

  • Upscaling at the last layer.
    • Network operations like feature extracting are done at low resolution space. This means that this network needs less computational resource than operating network at high resolution.
  • Learnable \(n\) upscaling filters
    • Network learn \(n\) upscaling filters for feature maps at last layer rather than one upscaling filter for input image.
    • This layer is not explicit interpolation filter, so network can implicitly learn processing necessary for super resolution.

    → Network can learn better and more complex low resolution to high resolution mapping than single fixed filter upscaling at first layer.