Memory Optimized Deep Dense Network for Image Super-resolution

CNN methods for image super-resolution consume a large number of training-time memory, due to the feature size will not decrease as the network goes deeper. To reduce the memory consumption during training, we propose a memory optimized deep dense network for image super-resolution. We first reduce redundant features learning, by rationally designing the skip connection and dense connection in the network. Then we adopt share memory allocations to store concatenated features and Batch Normalization intermediate feature maps. The memory optimized network consumes less memory than normal dense network. We also evaluate our proposed architecture on highly competitive super-resolution benchmark datasets. Our deep dense network outperforms some existing methods, and requires relatively less computation.