基于自适应多特征融合的双重图像退化修复网络

Dual image degradation repair network based on adaptive multi-feature fusion

当前图像修复方法大多局限于处理某个单一特定任务,如超分辨率、去噪、着色等,很少有网络模型同时具备处理双重退化的能力.而现存可以解决多重图像退化问题的算法普遍结构复杂、训练时间长、人力成本较大.本文提出一种基于自适应多特征融合的双重退化修复网络(adaptive multi-feature fusion dual degradation restoration network,AMFNet),利用自引导模块(SGM)融合图像的多尺度信息,有效去除了图像中的部分缺陷;使用带有空洞卷积的编码解码器模块巩固图像的语义信息,实现了中间图像的着色;引入带有自适应多特征融合模块(AMF)的中间信息传输机制(ITM)链接以上两大结构,自适应选择保留网络递进过程的图像特征以避免有用信息的丢失.实验结果表明,基于自适应多特征融合的双重图像退化修复网络模型视觉生成效果最优,通过在CelebA和Landscape数据集上的测试分析,其结构相似度(SSIM)与感知图像补丁相似度(LPIPS)优于同类方法,而峰值信噪比(PSNR)则远超同类方法高达5 dB.

Most current image restoration methods are limited to a single task,such as super-resolution,denoising,and shading,and few network models can handle dual image degradations.The existing algorithms that can solve the problem of multiple image degradation generally have complex structures,long training times,and large labor costs.This paper proposes a dual image degradation repair network based on an adaptive multi-feature fusion-adaptive multi-feature fusion dual degradation restoration network.The proposed model uses a self-guided module to fuse the multiscale information of an image,effectively removing some defects in the image.An encode-decode module with dilated convolution is used to consolidate the semantic information of the image and complete the coloring of the intermediate image.The intermediate information transmission mechanism with an adaptive multi-feature fusion module is introduced to link the above two structures,which can adaptively preserve the image features of the network progression process to avoid losing useful information.The experimental results show that the dual image degradation repair network model based on adaptive multi-feature fusion has the best visual effect,and through a test analysis on the Celeb A and Landscape datasets,a better learned perceptual image patch similarity and structural similarity index are obtained compared to similar methods,and the peak signal-to-noise ratio is up to 5 d B ahead of comparable methods.