基于残差网络的结直肠内窥镜图像超分辨率重建方法  

Super-resolution reconstruction for colorectal endoscopic images based on a residual network

作　　者：郑跃坤 葛明锋 常智敏[2] 董文飞 ZHENG Yue-kun;GE Ming-feng;CHANG Zhi-min;DONG Wen-fei(School of Biomedical Engineering(Suzhou),Division of Life Sciences and Medicine,University of Science and Technology of China,Hefei 230026,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Science,Suzhou 215163,China)

机构地区：[1]中国科学技术大学生物医学工程学院(苏州)生命科学与医学部,安徽合肥230026 [2]中国科学院苏州生物医学工程技术研究所,江苏苏州215163

出　　处：《中国光学（中英文）》2023年第5期1022-1033,共12页Chinese Optics

基　　金：国家重点研发计划(No.2021YFB3602200);苏州市科技计划项目(No.SZS201903)。

摘　　要：针对结直肠镜图像分辨率偏低、纹理信息偏少和细节模糊等缺点,提出了一种基于残差注意力网络的图像超分辨率重建算法SMRAN,选取结直肠息肉内窥镜图像数据集PolypsSet中的部分图像作为原始数据进行实验。首先,使用卷积网络提取低分辨率图像的浅层特征;其次,设计Res-Sobel结构对图像边缘特征进行增强;然后,通过引入不同大小的卷积核,设计多尺度特征融合模块(Multi-Scale feature Extraction Block,MEB),自适应地提取不同尺度的特征,从而得到有效的图像信息,并通过残差注意力网络将Res-Sobel模块和多尺度特征融合模块MEB进行连接;最后,通过亚像素卷积层对图像进行重建,得到最终的高分辨率图像。在尺度因子为×4时,网络在测试集上的测试结果如下:峰值信噪比PSNR为34.25 dB,结构相似性SSIM为0.8675。实验结果表明,与传统的双三次插值算法及常用的SRCNN、RCAN等深度学习算法相比,本文提出的SMRAN对结直肠内窥镜图像具有更好的超分辨率重建效果。In this paper,an image super-resolution reconstruction multi-scale algorithm based on a residual attention network(SMRAN)is proposed to solve the problems caused by low resolutions,less texture information and blurred details in colorectal endoscopic images.Images from the colorectal polyp endoscope image dataset PolypsSet are selected as the raw data for these experiments.A convolutional network is built to extract the shallow features of the low-resolution image and a Res-Sobel block is designed to enhance its edge features.A multi-scale feature fusion block MEB is designed by introducing convolution kernels of dif-ferent sizes to adaptively extract image features of different scales and obtain effective image information.The Res-Sobel block and multi-scale feature fusion module block MEB are connected through the residual at-tention network.Finally,a high-resolution image is reconstructed at the sub-pixel convolution layer.When the amplification factor is×4,the performance of the proposed algorithm on the test set are as follows:the peak signal-to-noise ratio(PSNR)is 34.25 dB and the structural similarity(SSIM)is 0.8675.Compared with the traditional bicubic interpolation algorithm and commonly used deep learning algorithms such as SRCNN and RCAN,the proposed SMRAN algorithm shows better super-resolution reconstruction results on colorectal endoscopic images.

关 键 词：内窥镜图像 超分辨率重建 残差结构 注意力机制 多尺度特征融合 索贝尔算子 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]