基于特征增强与损失优化的水下遮挡目标检测算法  

Underwater Occlusion Target Detection Algorithm of Feature Enhancement and Loss Optimization

作　　者：陈亮 杨羽翼 张剑[1] 吴亮红[1] 时慧晶[2] 彭辉[1] CHEN Liang;YANG Yuyi;ZHANG Jian;WU Lianghong;SHI Huijing;PENG Hui(School of Information and Electrical Engineering,Hunan University of Science and Technology,Xiangtan 411201,China;Kunming Shipborne Equipment Research&Test Center,Kunming 650051,China)

机构地区：[1]湖南科技大学信息与电气工程学院,湖南湘潭411201 [2]昆明船舶设备研究试验中心,云南昆明650051

出　　处：《探测与控制学报》2023年第3期109-115,共7页Journal of Detection & Control

基　　金：国家自然科学基金项目(62271199);湖南省教育厅资助科研项目(19C0793,19403D006)。

摘　　要：针对水下探测机器人在海洋作业时由于目标密集、形态重叠等原因容易产生误检、漏检的问题,提出一种基于特征增强与损失优化的水下遮挡目标检测算法。算法以YOLOv4骨干网络为基础,首先在随机通道上引入邻域融合的残差结构模块,通过通道注意力机制,提升通道的信息交互能力;其后,利用α-IoU优化CIoU-Loss损失函数,并采用真值排斥因子与预测排斥因子改进坐标回归损失函数,提高目标位置回归的精度;最后,针对水下图像数据干扰信息多的问题,采用基于密集度引导的自适应非极大值抑制方法完成对输出信息的处理,提升目标检测的召回率。通过对水下海洋生物的检测实验,算法在通用场景与密集遮挡场景下目标探测的mAP值分别提高了1.43%和4.4%,验证了算法的有效性。Aiming at the problem that underwater detection robot is prone to false detection and missing detection due to target density and morphological overlap,an underwater occlusion target detection algorithm based on efficient random residual and enhanced regression loss function was proposed in this paper.Based on YOLOv4,the algorithm firstly introduced the neighborhood fusion residual structure module in the random channel,and improved the information interaction ability of the channel through the channel attention mechanism.Then,CIoU-Loss Loss function was optimized by-IoU,and the coordinate regression loss function was improved by the truth rejection factor and the prediction rejection factor to improve the accuracy of the target position regression.Finally,in view of the problem of much interference information of underwater image data,the algorithm adopted adaptive non-maximum suppression method based on intensity guidance to complete the processing of output information,which could improve the recall rate of target detection.Through experiments on underwater detection in Bohai Sea,the mAP value of the algorithm was improved by 1.43%and 4.4%respectively in general underwater detection and dense target scenarios,which verified the effectiveness of the algorithm.

关 键 词：水下目标探测 YOLOv4 遮挡目标 通道注意力 损失函数优化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]