Binarized Neural Network for Single Image Super Resolution

摘要

Lighter model and faster inference are the focus of current single image super-resolution (SISR) research. However, existing methods are still hard to be applied in real-world applications due to the heavy computation requirement. Model quantization is an effective way to significantly reduce model size and computation time. In this work, we investigate the binary neural network-based SISR problem and propose a novel model binarization method. Specially, we design a bit-accumulation mechanism (BAM) to approximate the full-precision convolution with a value accumulation scheme, which can gradually refine the precision of quantization along the direction of model inference. In addition, we further construct an efficient model structure based on the BAM for lower computational complexity and parameters. Extensive experiments show the proposed model outperforms the state-of-the-art binarization methods by large margins on 4 benchmark datasets, specially by average more than 0.7 dB in terms of Peak Signal-to-Noise Ratio on Set5 dataset.