基于金字塔边缘增强的自矫正低光照目标检测  

A self correcting low-light object detection method based on pyramid edge enhancement

作　　者：蒋占军[1] 吴佰靖 马龙 廉敬[1] JIANG Zhanjun;WU Baijing;MA Long;LIAN Jing(School of Electronics&Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070

出　　处：《光学精密工程》2024年第20期3099-3111,共13页Optics and Precision Engineering

基　　金：国家自然科学基金(No.62061023);甘肃省水利厅省级项目(No.LZJT523029)。

摘　　要：针对低光照图像整体亮度和对比度低,且目标边缘特征有限,导致目标检测算法识别定位精度不高的问题,提出一种低光照目标检测方法。首先,提出低光照图像增强网络,利用图像高斯金字塔、Retinex和暗通道去雾在低光照图像增强的优点,并在暗通道去雾算法中加入边缘轮廓特征,在增强整体亮度对比度的同时,突出目标的边缘特征;其次,为提高特征的提取准确性,在RTDETR的特征提取部分,设计了轻量化自矫正特征提取网络,以更小的计算量生成并矫正主干特征提取网络生成的特征图,提升目标检测精度。在ExDark数据集上的实验结果表明:较于基准RTDETR,精度提高了2.34%,召回率提升了2.09%,参数量减少了4.95 M,模型大小减少了13.31 MB,本文方法能够有效提升低光照场景下的目标检测性能。A low-light target detection method was proposed to overcome the problem of low overall brightness,contrast and limited edge features in low-light images,which lead to poor recognition and localization of target detection algorithms.Firstly,a low-light enhancement network was designed to utilize the advantages of image Gaussian pyramid,Retinex and dark-channel defogging in low-light image enhancement,and edge contour features were added to the dark-channel defogging algorithm to enhance the overall luminance contrast while highlighting the edge features of the target.Secondly,to improve the accuracy of feature extraction in the feature extraction section of RTDETR,a lightweight self correcting feature extraction network was designed to generate and correct the feature maps generated by the backbone feature extraction network with smaller computational complexity,thereby improving the accuracy of object detection.The experimental results on the ExDark dataset shows that compared with the benchmark RTDETR,the mAP improves by 2.34%,the recall improves by 2.09%,the parameter amount reduces by 4.95 M,the model size reduces by 13.31 MB,and the proposed method is able to effectively improve the performance of the target detection in the low-light scene.

关 键 词：目标检测 暗目标 低光照增强 高斯金字塔 自矫正网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]