基于改进CycleGAN的夜间道路环境下非机动车特征增强方法  

Pedestrian feature enhancement method in night road environment based on improved CycleGAN

作　　者：黄佳恒 叶青[1] 邱实卓 Huang Jiaheng;Ye Qing;Qiu Shizhuo(School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114,China)

机构地区：[1]长沙理工大学电气与信息工程学院,长沙410114

出　　处：《现代计算机》2023年第20期1-8,共8页Modern Computer

摘　　要：夜间光照度低或者光照分布不均匀,非机动车识别率受到很大影响。为有效增强目标图像,研究结合一种热图自注意力Grad-CAM网络和一种自适应归一化层AdaLIN改进CycleGAN的生成器,使用自适应归一化层AdaLIN改进CycleGAN判别器,在改进CycleGAN网络中加入Merge网络,生成类夜间非机动车特征增强图像。经过实验验证,相较于对比算法RetinexNet、MSRCP和CycleGAN+Merge算法,主要图像评价指标明显提高,比改进前CycleGAN+Merge算法在PNSR、SSIM指标上分别提高了4.16%、7.22%,在人类视觉主观分析上也更优。证明对改进算法进行研究后能够有效改善夜间道路情况下非机动车特征的视觉效果。Low illumination at night or uneven light distribution greatly affects the recognition rate of non-motor vehicles.In order to effectively enhance the target image,the generator of improved CycleGAN is combined with a heatmap self-attention Grad-CAM network and an adaptive normalization layer AdaLIN to improve the generator of CycleGAN.Improve the CycleGAN discriminator using the adaptive normalization layer AdaLIN.The Merge network is added to the improved CycleGAN network to generate a nocturnal non-motor vehicle feature enhancement image.After experimental verification,compared with the comparison algorithm RetinexNet,MSRCP and CycleGAN+Merge algorithm,the main image evaluation index is significantly improved,which is 4.16%and 7.22% higher than the PNSR and SSIM indicators of the improved CycleGAN+Merge algorithm,respectively,and it is also better in the subjective analysis of human vision.It is proved that the proposed algorithm can effectively improve the visual effect of non-motor vehicle features under night road conditions.

关 键 词：Grad-CAM AdaLIN U-net 无监督学习 夜间低照度图像 CycleGAN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U463.6[自动化与计算机技术—计算机科学与技术] U495[机械工程—车辆工程]