基于卷积神经网络的图像超分辨率方法研究  

RESEARCH ON IMAGE SUPER-RESOLUTION METHOD BASED ON CONVOLUTIONAL NEURAL NETWORK

作　　者：杨志政 王春兴[1] Yang Zhizheng;Wang Chunxing(School of Physics and Electronics,Shandong Normal University,250358,Jinan,China)

机构地区：[1]山东师范大学物理与电子科学学院,济南250358

出　　处：《山东师范大学学报（自然科学版）》2018年第4期427-433,共7页Journal of Shandong Normal University(Natural Science)

摘　　要：结合认知构架ACT-R模型(Adaptive Control of Thought-Rational),基于邻域嵌入算法和深度学习的图像超分辨率重建方法,构建一个应用于超分辨率重建研究的ACT-R研究模型.在匹配阶段,根据低分辨率(Low Resolution,LR)测试图像的结构和内容特征,运用图像的多尺度相似性和非局部相似性,对图像进行特征提取;在选择阶段,把邻域嵌入算法分为两层,进行邻域图像块的寻找,同时构建一个端到端的深层门限卷积神经网络,把从匹配阶段得到的高分辨率无细节小图像块输入到卷积神经网络中并得到输出图像,将输出图像与高分辨率无细节小图像块相加得到该低分辨率小图像块对应的高分辨率小图像块,最后把高分辨率图像块组合成高分辨率(High Resolution,HR)图像.决策阶段,我们进行实验并与其他方法对比.结果表明,该模型对单帧图像具有良好的重建能力,在视觉效果上和客观评价标准上都取得了不错的效果,能够较好的重建低分辨率图像.Combined with the cognitive architecture ACT-R model and image super-resolution reconstruction method based on neighborhood embedding algorithm and deep learning,a ACT-R research model applied to superresolution reconstruction research was built.According to the structure and content features of the low resolution test images,we can make use of the multi-scale similarity and non-local similarity of the image to extract the features of the images.At the selection stage,the neighborhood embedding algorithm is divided into two layers to search for neighborhood image blocks.At the same time,an end-to-end deep threshold convolution neural network is constructed.The high-resolution small image blocks obtained from the matching stage are input into the convolution neural network.The output images and the high-resolution small detail-free images are obtained.The output image blocks are added to the high-resolution and detail-free small image blocks to obtain the high-resolution small image blocks.The high resolution image blocks are combined into a high resolution(HR)image.At the decision stage,we conduct experiments and compare with other methods.The method proposed in this paper has achieved good results in both visual effects and objective evaluation criteria.The results show that the model has a good ability of reconstructing a single frame image.

关 键 词：卷积神经网络 邻域嵌入算法 ACT-R认知模型 超分辨率重建 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]