Advanced Binary Neural Network for Single Image Super Resolution

Binary neural network (BNN) is an effective approach to accelerate the model inference and has been initially applied in the field of single image super resolution (SISR). However, the optimization of efficiency and accuracy remains a major challenge for achieving further improvements. While existing BNN-based SR methods solve the SISR problems by proposing a residual block-oriented quantization mechanism, the quantization process in the up-sampling stage and the representation tendency of binary super resolution networks are ignored. In this paper, we propose an Advanced Binary Super Resolution (ABSR) method to optimize the binary generator in terms of quantization mechanism and up-sampling strategy. Specifically, we first design an excitation-selection mechanism for binary inference, which could distinctively implement self-adjustment of activation and significantly reduce inference errors. Furthermore, we construct a binary up-sampling strategy that achieves performance almost equal to that of real-valued up-sampling modules, and fully frees up the inference speed of the binary network. Extensive experiments show that the ABSR not only reaches state-of-the-art BNN-based SR performance in terms of objective metrics and visual quality, but also reduces computational consumption drastically.