【摘要】 传统的基于CNN的方法在对低分辨率图像进行重构处理过程中,并未将图像中的低频结构信息和高频细节信息进行区别处理,且网络的层与层之间缺乏信息交流,从而造成高分辨率重建图像结果中出现信息缺失。为获取更多图像各层次特征的结构与细节信息,本文构建了基于小波域的残差密集网络(WRDSR)。该网络在二维离散小波变换形成的小波域内,利用密集连接和残差连接对图像不同频率的信息进行充分提取后,将融合后的特征输入到亚像素卷积层生成高分辨率图像的小波子带图像,最后通过二维离散小波逆变换生成高分辨率图像。与Bicubic、SRCNN、VDSR、LapSRN、DWSR、SDSR等算法相比,WRDSR在评价指标PSNR/SSIM上平均提高了2.824dB/0.059 5、0.747dB/0.016 8、0.016dB/0.002 4、0.025dB/0.004 3、0.21dB/0.004 7和0.20dB/0.005 7,在更高效地利用原始图像信息的同时,解决了信息缺失的问题,使得重建图像的纹理更清晰,细节更丰富,视觉效果更佳。更多还原

【Abstract】 The traditional CNN-based method can not distinguish between low-frequency structural information and high-frequency detailed information in the process of reconstructing low-resolution images.And there is a lack of information communication between layers of the network,which leads to the problem of missing information in the high-resolution reconstruction image.In order to obtain more information about the structure and details of each level of image features a residual dense network is constructed based on the wavelet domain(WRDSR).In the wavelet domain formed by the two-dimensional discrete wavelet transform,the network uses dense connections and residual connections to fully extract the information of different frequencies of the image,and then generates highresolution image wavelets by inputting the fused features into sub-pixel convolution.Finally,the final high-resolution image is generated by two-dimensional discrete wavelet inversion.Compared with Bicubic,A+,SRCNN,VDSR,LapSRN,DWSR,SDSR etc.,WRDSR improves 2.824 dB/0.059 5,0.896 dB/0.018 2,0.747 dB/0.016 8,0.016 dB/0.002 4,0.025 dB/0.004 3,0.21 dB/0.004 7 and 0.20 dB/0.0057 on average on PSNR/SSIM,respectively.While making more efficient use of the original image information,WRDSR solves the drawback of missing information,making the reconstructed image texture clearer,richer in details and better in visual effect.更多还原