摘要
针对微小目标检测存在准确率低、速度慢等问题,提出了一种基于改进YOLOX的微小目标快速检测方法。改进YOLOX算法中嵌入了一种轻量级的卷积块注意力模块(Convolutional Block Attention Module,CBAM),以帮助算法在大背景中找寻局部密集而微小的潜在目标区域,增强微小目标特征提取能力;使用DIoU损失函数优化收敛速度,提升目标检测的定位能力;在解耦头中引入Focal Loss作为置信度损失函数,缓解难易样本数量的不均衡。结果表明,改进算法在TinyPerson数据集上平均正确率(Average Precision,AP)最高可达52.69%,同时,改进YOLOX-s算法在Tesla P100上分辨率为1 440×928像素的图片检测速度可达35 FPS(Frames Per Second)。
To resolve the problems in tiny object detection,such as low accuracy and slow speed,a fast tiny object detection method is proposed based on improved YOLOX.A lightweight Convolutional Block Attention Module(CBAM) is embedded in the YOLOX algorithm to improve the feature extraction ability of tiny objects by helping the algorithm recognize locally dense and tiny areas as potential targets in big data background.Meanwhile,the DIoU loss function is used to optimize the convergence speed and improve the positioning ability of object detection.Moreover,the Focal Loss is used as the loss function of the confidence degree in the decoupling head to alleviate the imbalance between difficult and simple samples sizes.The results show that the Average Precision(AP) of the improved algorithm can reach up to 52.69% for TinyPerson,and the detection speed for the improved YOLOX-s algorithm on a Tesla P100 can reach 35 frames per second for an image with a resolution of 1440×928.
作者
李家俊
杨杰
LI Jiajun;YANG Jie(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi,China)
出处
《江西冶金》
2023年第2期164-172,共9页
Jiangxi Metallurgy
基金
国家自然科学基金资助项目(62063009)。
