基于广度残差与像素点注意力的图像去模糊模型  

Image Deblurring Model Based on Width Residual and Pixel Attention

作　　者：况发 熊邦书[1] 欧巧凤[1] 余磊[1] KUANG Fa;XIONG Bangshu;OU Qiaofeng;YU Lei(Key Laboratory of Image Processing and Pattern Recognition of Jiangxi Province,Nanchang Hangkong University,Nanchang 330063,Jiangxi,China)

机构地区：[1]南昌航空大学图像处理与模式识别江西省重点实验室,江西南昌330063

出　　处：《应用科学学报》2022年第6期996-1005,共10页Journal of Applied Sciences

基　　金：国家自然科学基金(No.61866027,No.62162044);江西省自然科学基金(No.20202BAB202016);江西省重点研发计划(No.20212BBE53017);南昌航空大学研究生创新专项基金(No.YC2020031)资助。

摘　　要：针对现有方法难以快速从模糊图像中恢复高质量清晰图像的问题,提出了基于广度残差与像素点注意力的图像去模糊模型。该模型以编解码网络为基础,采用广度卷积与多阶残差方法,构建广度残差模块,提高了模型处理速度;同时,采用局部平均与矩阵叉乘,构建像素点注意力模块,增强了模型去模糊质量。在GOPRO数据集上进行的实验结果表明,在模型大小仅为22.24 MB情况下,结构相似度为0.9223,峰值信噪比为31.74 dB,平均运行时间为0.37 s。所提出方法与尺度循环网络方法相比,其峰值信噪比提高了4%,并且性能优于现有其他去模糊方法。In order to solve the problem that existing methods suffer difficulty in quickly recovering high-quality sharp images from blurred images,an image deblurring model based on width residual and pixel attention is proposed.Based on encoder-decoder networks,the model uses wide convolution and multi-order residual method to construct width residual modules,improving the processing speed of the model.At the same time,local average and matrix cross multiplication are used to construct pixel attention modules,which enhance the model deblurring quality.The experimental results on GOPRO datasets show that the structural similarity of the proposed method is 0.9223,the peak signal-to-noise ratio is 31.74 dB,and the average running time is 0.37 seconds when the model size is 22.24 MB.Compared with the scale-recurrent network method,the peak signal-to-noise ratio of the proposed method improves by 4%,and its performance is better than the other existing deblurring methods.

关 键 词：图像去模糊 编解码网络 像素点注意力 广度残差 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]