Closed-loop Feedback Network with Cross Back-Projection for Lightweight Image Super-Resolution

Since the development of deep learning, image super-resolution (SR) has made great progress, and become the focus of academic research. Because high-level features are more informative for the reconstruction, most SR networks have a large number of layers and parameters, which restrict their application in resource-constrained devices. Recently, lightweight networks got a lot attention for their broad application prospect. To improve the performance of lightweight networks by informative high-level features, we introduce feedback mechanism into our method, which can feed back high-level features to refine low-level ones. In this paper, we propose a closed-loop feedback network with cross back-projection for lightweight image super-resolution (CCFN), which uses feedback mechanism in three manners. First, based on error feedback, we propose a cross back-projection feedback block (CFB). CFB uses error feedback to correct the features of multi-scale fusion, which also can be viewed as two cross-learning back-projection units. Second, CFB works in a self-feedback manner, which feeds back the output high-level features to refine the low-level ones of the input. Third, we propose a global feedback, which feeds back the degradation results of SR to LR, to guide the learning of mapping functions from LR to HR. Finally, we use attention-based model as the basic block in CFB, and since our method works in an iterative manner, recursive concatenation is more suitable than multi-reconstruction. The final experimental results show that our CCFN has a competitive performance with few parameters.