Analysis and Optimization of CNN-based Super Resolution with Filter Pruning

Recent studies using convolutional neural networks (CNN) have been conducted in the field of super resolution. These studies include Super-Resolution Convolution Neural Networks (SRCNN) and Very Deep Convolution Neural Networks Super-Resolution (VDSR). SRCNN and VDSR are known to be heavy in terms of parameter size and the inference speed is low to be applied in real-time. In order to overcome these issues, we apply filter pruning to optimize the network. Experimental results show that when the network is 71% pruned, the parameter size was reduced to 71%, the inference speed is enhanced by about 60%, and the inference quality is maintained at 94% compared to the original VDSR network.