基于改进MSR和YOLOv5的雾天行人检测算法

Pedestrian Detection Algorithm in Hazy Weather Based on Improved MSR and YOLOv5

针对雾气导致图像中行人检测准确率不高的问题,提出了1种基于改进多尺度Retinex(multiple scale Retinex,MSR)和改进YOLOv5(you only look once v5)的雾天行人检测算法。首先,针对MSR算法容易产生光晕和伪影现象的问题,引入均值和均方差对其进行改进,通过伽马校正找到适宜的图像亮度,调整亮度后再对图像进行去雾操作。其次,以传统的YOLOv5检测模型为基础并对其进行改进,引入选择性内核网络(selective kernel networks,SK-Net)模块,与YOLOv5的骨干网络(Backbone)端相融合,输入信息自适应地调整其感受野大小,加强模型对主要信息的提取,提升模型的精度。实验结果表明,改进后的MSR和改进后的YOLOv5相结合得到的算法,各项行人检测指标都有较大提升,其识别精确率、召回率、各类别平均精度均值分别达到了91.2%、87.3%、90.1%,改进后的算法能有效提高雾天行人检测的效率。

To solve the problem of low accuracy of pedestrian detection caused by fog in hazy weather,we propose a pedestrian detection algorithm based on improved multiple scale Retinex(MSR)and improved YOLOv5(you only look once v5)in hazy weather.First,aiming at the halo and artifact phenomenon that MSR is prone to produce,mean value and mean square error are introduced to improve it,and appropriate image brightness is found through gamma correction.After brightness adjustment,the image is dehazed.Secondly,based on the traditional YOLOv5 detection model,it is improved.Selective kernel networks(SK-Net)module is introduced,which is integrated with the backbone terminal of YOLOv5.The input information adaptively adjusts the size of its receptive field,strengthens the extraction of main information from the model,and improves the accuracy of the model.The experimental results show that combining the improved MSR with the improved YOLOv5 has greatly improved the pedestrian detection indicators.The recognition precision,recall and mean average precision value of each category reached 91.2%,87.3%and 90.1%respectively.The improved algorithm can effectively improve the efficiency of pedestrian detection in hazy weather.