图像多尺度密集网络去模糊模型  

Deblurring Model of Image Multi-Scale Dense Network

作　　者：宋昊泽 吴小俊[1] Song Haoze;Wu Xiaojun(School of Internet of Things Engineering,Jiangnan University,Wuxi,Jiangsu 214122,China)

机构地区：[1]江南大学物联网工程学院

出　　处：《激光与光电子学进展》2019年第21期24-32,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61672265,61373055)

摘　　要：使用基于深度学习的端到端去模糊方法,将模糊图像编码后再解码成高清图像。针对编码过程中网络模型存在提取特征信息不足,导致重建的去模糊图像质量下降的问题,提出两种网络结构改进方法:在自编码网络中添加密集网络结构以提高网络提取特征信息的能力;引入多尺度感受野结构,该结构由4个尺度的平均池化层和上采样层组成,从而提取更多输入图像的上下文特征信息。在GOPRO数据集和Kohler数据集,两种网络改进方法均取得了较好的图像重建效果。This study uses an end-to-end method for image deblurring based on deep learning to encode the blurred image and to subsequently decode it into a high-definition image.However,the lack of extracted feature information during encoding decreases the quality of the reconstructed deblurred image.To solve this problem,we propose two methods for improving the network structure.First,a dense network structure is added to the autoencoder network for extracting considerable feature information.Second,a multiscale perceptual field structure is introduced to extract considerable contextual feature information,comprising 4 scales of average pooling layers and up-sampling layers.The two improved methods achieve good image deblurring effects using the GOPRO and Kohler datasets.

关 键 词：图像处理 图像去模糊 多尺度结构 平均池化层 上采样层 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]