纹理细节恢复的图像超分辨率重建算法

Image super-resolution reconstruction algorithm for texture detail recovery

针对现有单图像超分辨率方法在重建过程中容易忽略原图像中不同结构-纹理的差异与联系,导致生成的高分辨率图像缺乏纹理细节并存在伪影的问题,提出了纹理细节恢复的图像超分辨率重建算法。该方法由梯度分支、纹理分支和图像超分辨率分支组成。其中,在梯度分支和纹理分支之间使用了类注意力模块处理二者的特征混淆问题,并通过双向特征融合模块实现了对结构特征与纹理特征的相互促进,作为先验信息以达到纹理细节信息增强的目的。此外,在图像超分辨率分支还通过构建特征恢复模块,利用浅层和深层信息帮助网络保留了图像中更丰富的上下文信息和纹理细节。该方法通过在DIV2K数据集上进行了网络训练,并在5个基准测试集Set5、Set14、BSD100、Urban100和MANGA109上进行了实验,峰值信噪比(PSNR, Peak Signal to Noise Ratio)分别:37.88dB、33.28dB、32.0781dB、31.89dB、38.39dB,相比现有方法均有显著提升。实验结果表明,本文方法获得了有效的重建图像并且保留更多的图像细节,生成具有边缘清晰和逼真细节的超分辨率图像。

Aiming at the problem that the existing single-image super-resolution methods tend to ignore the differences and relations between different structures and textures in the original image during the reconstruction process,resulting in the lack of texture details and artifacts in the generated high-resolution image,an Image Super-resolution Reconstruction Algorithm for Texture Details Recovery(TDRSR)is proposed.The method consists of gradient branch,texture branch and image super-resolution branch.Among them,the class attention module is used between the gradient branch and the texture branch to deal with the feature confusion problem of the two,and the mutual promotion of structural features and texture features is realized through the bidirectional feature fusion module,which is used as prior information to achieve texture details.purpose of information enhancement.In addition,the image super-resolution branch also helps the network to retain richer contextual information and texture details in the image by building a feature recovery module that utilizes both shallow and deep information.The method trains the network on the DIV2K dataset,and conducts test experiments on 5 benchmark test sets Set5,Set14,B100,Urbanl00 and MANGA109,the peak signal-to-noise ratio(PSNR):37.88dB,33.28dB,32.0781dB,31.89dB and 38.39dB,which are significantly improved compared to the results of other methods.The experimental results show that the method obtains effective reconstructed images and preserves more image details,generating super-resolution images with sharp edges and realistic details.