改进AOD-Net的轻量级图像去雾算法  

Lightweight Image Defogging Algorithm Based on Improved AOD-Net

作　　者：张骞 陈紫强[1] 姜弘岳 赵玖龙 ZHANG Qian;CHEN Ziqiang;JIANG Hongyue;ZHAO Jiulong(School of Information and Communication,Guilin University of Electronic Science and Technology,Guilin 541004,Guangxi,China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《实验室研究与探索》2022年第7期18-22,共5页Research and Exploration In Laboratory

基　　金：国家自然科学基金项目(61861011)。

摘　　要：针对AOD-Net图像去雾算法质量偏低,存在色差的问题,提出一种改进的图像快速去雾算法。该算法对AOD-Net的前两个卷积层进行位置归一化,并将提取的矩信息输入后续的网络层中进行仿射变换,在改善原有网络中的数据分布及提高网络的收敛速度后,引入PSA注意力模块,使用多尺度的卷积核提取特征信息,加权融合特征并调整网络通道的权重,抑制冗余信息,提高模型的去雾质量。利用公开数据集RESIDE与现有的轻量级去雾算法进行实验对比,改进算法的单幅图像去雾耗时仅为4.3 ms,去雾质量优于DCP、CAP和Dehaze-Net等去雾算法。与AOD-Net相比,该方法的峰值信噪比提高了2.71 dB,结构相似度达到0.95,有效提升网络的图像去雾能力。Aiming at the low quality of AOD-Net image defogging algorithm and the problem of chromatic aberration,an improved fast image defogging algorithm is proposed.The algorithm normalizes the position of the first two convolutional layers of AOD-Net,and the extracted moment information is input into the subsequent network layers for affine transformation to improve the data distribution in the original network and increase the convergence speed of the network.The PSA attention module is introduced to extract feature information by using multi-scale convolution kernel,and the weighted fusion feature is used to adjust the weight of network channel to suppress redundant information and improve the dehazing quality of the module.Using the public data set RESIDE to compare with the existing lightweight dehazing algorithm,the improved algorithm takes only 4.3 ms to dehaze a single image,and the defogging quality is better than DCP,CAP and Dehaze-Net,and other defogging algorithms.Compared with AOD-Net,the peak signal-to-noise ratio PSNR of this method is increased by 2.71 dB,and the structural similarity SSIM reaches 0.95,which effectively improves the image defogging ability of the network.

关 键 词：图像去雾 深度学习 位置归一化 注意力模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]