基于小波-残差注意力的超分辨率重建网络OA

To address the issues of insufficient utilization of frequency domain features and low efficiency in the fusion of global and local features in existing super-resolution methods,a wavelet-residual attention network for super-resolution(WRANet)had been proposed.A cross-band feature extraction module(FFB)had been designed,and a Wavelet Attention Convolution(WAConv)module had been proposed,which had employed a two-layer wavelet operation to obtain multi-scale frequency domain features.High-frequency features had been extracted and an extremely large receptive field had been obtained through convolution and a spatial-channel dual-branch attention mechanism(DAB).Residual attention(RAB)had been utilized to focus on high-frequency details.A cross-layer dense connection strategy had been established to dynamically fuse shallow texture and deep semantic features via 1×1 convolution.A feature aggregation module(FAM)had been employed to merge fine-grained local features and coarse-grained global features.Experiments had shown that the proposed method had achieved higher objective metrics compared to other methods,and the visual results had demonstrated that the images reconstructed by this method contained rich texture details.