Back-Projection Lightweight Network for Accurate Image Super Resolution

Abstract

Recently, deep neural networks have led to tremendous advances in image super-resolution and achieved remarkable results. However, most deep learning methods attain appreciable performance by increasing model parameters and complexity. Consequently, they are difficult to utilize in common devices despite significant results. With respect to application in resource-limited devices, the goal of the present study involves designing a compact network that exhibits good scalability. We proposed a back projection network (BPnet) as we were inspired by a traditional image super-resolution technique, namely iterative back projection (IBP). Our proposed network is composed of different modules. Each module performs the enlarging function based on the result of the previous module, which is similar to an iteration in IBP. The input of each module network corresponds to the previous down-sampled output minus the low-resolution input image. Additionally, the output of the network is the residual between the ground truth high-resolution image and previous output. The non-linear property of a neural network is maximized through the sparsity of residual input/output. Thus, we can achieve a lightweight network without sacrificing the quality of the results. The experiments indicate that the number of parameters of the proposed BPnet can be less than quarter of those proposed in the state-of-the-art papers and still have comparable results.