基于双路增强残差块连接的图像去块效应网络  

Image Deblocking Network Based on Dual Path Enhanced Residual Block Connection

作　　者：冯洁丽 何小海[1] 任超[1] 陈洪刚[1] 王新欢 FENG Jieli;HE Xiaohai;REN Chao;CHEN Honggang;WANG Xinhuan(College of Electronic Information,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,四川成都610065

出　　处：《无线电工程》2022年第12期2271-2279,共9页Radio Engineering

基　　金：国家自然科学基金(61871279,62081330105);中央高校基本科研业务费专项资金(2021SCU12061)。

摘　　要：在实际工作中,由于存储空间和传输带宽的限制,往往需要对采集的图像和视频进行一定的压缩。在基于块的图像压缩中,一种很明显影响视觉体验的压缩伪影被称为图像的块效应。针对此问题,构建了一种基于双路增强残差块连接的图像去块效应网络(Dual Path Enhanced Residual Block Network,DPERBN)。具体地,提出的算法主要由双路增强块(Dual Path Enhanced Block,DPEB)构建而成。提出的DPEB不仅能够捕获图像丰富的多尺度特征,而且其中的双路结构能重复利用残差连接和密集连接2种网络结构探索到的新特征。双路单元(Dual Path Unit,DPU)能激活非线性特征,促进信息在网络中的流动;过渡单元(Transition Unit,TU)能将信息有效融合并传递。消融实验验证了提出DPERBN的有效性。验证结果表明,该方法获得的去压缩效应图像具有更高的客观评价指标,图像细节更丰富,有助于后续的处理及分析。In practice,compression is necessary for the collected images and videos because of the limitation of storage and transmission bandwidth.One of the visually noticeable compression artifacts in block-based image/video compression is called image blocking effect.To solve the problem,an image deblocking network based on Dual Path Enhanced Residual Block Connection(DPEBRN)is established.Specifically,the proposed algorithm is mainly constructed of Dual Path Enhanced Block(DPEB).Not only can the proposed DPEB capture rich multi-scale features of images,but also its dual path structure can reuse new features explored by the two network structures of residual connection and dense connection.Dual Path Unit(DPU)can activate nonlinear features and promote information flow in the network,while Transition Unit(TU)can effectively fuse and transmit information.The effectiveness of the proposed DPEBRN is verified through ablation experiment.Experimental results show that the decompression effect image captured by this algorithm has higher objective evaluation index and more details,which is beneficial to further processing and analysis.

关 键 词：双路增强块 去压缩效应 卷积神经网络 空洞卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]