基于密集连接任务对齐的小目标检测算法
Small target detection algorithm based on dense connection task alignment
田春欣 1陈绪君 1郑有凯1
作者信息
- 1. 华中师范大学 物理科学与技术学院,湖北 武汉 430079
- 折叠
摘要
针对当前基于卷积神经网络的单阶段目标检测算法(YOLO系列、TOOD等)对高空拍摄场景下的小目标检测存在精度低、漏检、误检等问题,在TOOD算法基础上,提出一种基于密集连接任务对齐的小目标检测算法DATNet.为提升网络的检测能力,采用CSPDarkNet网络提取输入图像特征,通过密集连接的方式融入空洞卷积,添加注意力模块捕捉感兴趣的目标区域,引入DIoU回归损失函数通过任务对齐的检测头来训练模型.实验结果表明,DATNet在VisDrone-DET数据集上相对于TOOD算法平均准确率提升1.7％,为小目标检测场景提供了一种技术方案.
Abstract
To solve the problems of low accuracy,missed detection and false detection of current single-stage target detection algorithms based on convolutional neural networks(YOLO series,TOOD,etc.)for small target detection in high-altitude shoo-ting scenes,a small target detection algorithm DATNet based on dense connection task alignment was proposed on the basis of TOOD algorithm.To improve the detection ability of the network,CSPDarkNet network was used to extract the input image features,the hole convolution was integrated through dense connection,the attention module was added to capture the target area of interest,and DIoU regression loss function was introduced to train the model through task aligned detection heads.Experimental results show that the average accuracy of DATNet on Visdrone dataset is improved by 1.7％compared with that of TOOD algorithm,which provides a technical solution for small target detection scenarios.
关键词
密集连接/空洞卷积/任务对齐/目标检测/小目标/上下文信息/特征提取Key words
dense connection/atrous convolution/task alignment/target detection/small targets/context information/feature extraction引用本文复制引用
基金项目
国家自然科学基金项目(60101204)
湖北省自然科学基金项目(2020CFB474)
出版年
2024
NETL
NSTL
万方数据