基于Dark-YOLO的低照度目标检测方法  

Low-Illumination Object Detection Method Based on Dark-YOLO

作　　者：江泽涛[1] 肖芸 张少钦[2] 朱玲红 何玉婷 翟丰硕 Jiang Zetao;Xiao Yun;Zhang Shaoqin;Zhu Linghong;He Yuting;Zhai Fengshuo(Guangxi Key Laboratory of Image and Graphics Intelligent Processing,Guilin University of Electronic Technology,Guilin 541004;School of Civil and Architecture,Nanchang Hangkong University,Nanchang 330063;School of Information Engineering,Nanchang Hangkong University,Nanchang 330063)

机构地区：[1]桂林电子科技大学广西图像图形与智能处理重点实验室,桂林541004 [2]南昌航空大学土木建筑学院,南昌330063 [3]南昌航空大学信息工程学院,南昌330063

出　　处：《计算机辅助设计与图形学学报》2023年第3期441-451,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(62172118,61876049);广西自然学科基金重点项目(2021GXNSFDA196002);广西图像图形智能处理重点实验项目(GIIP2008);广西研究生教育创新计划资助项目(YCBZ2021070,YCBZ2018052,2020YCXS050).

摘　　要：在复杂的低照度环境中获取的图像存在亮度低、噪声多和细节信息丢失等问题,直接使用通用的目标检测方法无法达到较为理想的效果.为此,提出低照度目标检测方法——Dark-YOLO.首先,使用CSPDarkNet-53骨干网络提取低照度图像特征,并提出路径聚合增强模块以进一步增强特征表征能力;然后,设计金字塔平衡注意力模块捕获多尺度特征并加以有效利用,生成包含不同尺度且更具判别力的特征;最后,使用预测交并比(intersection over union,IoU)改进检测头,IoU预测分支为每个预测框预测IoU值,使得目标定位更加准确.在ExDark数据集上的实验结果表明,相较于YOLOv4,均值平均精度(mAP)提升了4.10%,Dark-YOLO方法能够有效地提高在低照度场景下目标检测的性能.The images acquired in a complex low-illumination environment have problems such as low brightness,high noise,and loss of detailed information.The general object detection method cannot be used directly to achieve relatively ideal results.In this situation,a low-illumination object detection method-Dark-YOLO is proposed.Firstly,the CSPDarkNet-53 backbone network is used to extract the features of low-illumination image,and the path aggregation enhanced module is proposed to further enhance the ability of feature representation.Then,the pyramid balanced attention module is designed to capture multi-scale features and make use of them effectively to generate more discriminant features with different scales.Finally,the prediction intersection over union(IoU)is used to improve the performance of detection head.The IoU prediction branch predicts the IoU value for each prediction box,which makes the object positioning more accurate.The experimental results on the ExDark dataset show that compared with YOLOv4,the mean average precision(mAP)is improved by 4.10%.Dark-YOLO method can effectively improve the performance of object detection in low-illumination scenes.

关 键 词：目标检测 低照度图像 注意力机制 多尺度特征 预测交并比 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]