用于水下声呐目标检测的弱特征共焦通道调控方法  

Weak feature confocal channel regulation for underwater sonar target detection

作　　者：何梦云 何自芬[1] 张印辉[1] 陈光晨 张枫 HE Meng-yun;HE Zi-fen;ZHANG Yin-hui;CHEN Guang-chen;ZHANG Feng(Faculty of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学机电工程学院,云南昆明650500

出　　处：《中国光学（中英文）》2024年第6期1281-1296,共16页Chinese Optics

基　　金：国家自然科学基金资助项目(No.62171206,No.62061022)。

摘　　要：声呐图像视觉检测是复杂水域资源勘探和水下异物目标探测领域的重要技术之一。针对声呐图像中小目标存在的特征微弱和背景信息干扰问题,本文提出弱特征共焦通道调控水下声呐目标检测算法。为了提高模型对弱小目标的信息捕获和表征能力,设计弱小目标特征激活策略,并引入先验框尺度校准机制匹配底层语义特征检测分支,以提高小目标检测精度。应用全局信息聚合模块深入挖掘弱小目标的全局特征,避免冗余信息覆盖小目标微弱关键特征。为解决传统空间金字塔池化易忽视通道信息的问题,提出共焦通道调控池化模块,保留有效通道域小目标信息并克服复杂背景信息干扰。实验结果表明,本文所提模型在水下声呐数据集的9类弱小目标识别的平均检测精度达83.3%,相较基准提高了5.5%,其中铁桶、人体模型和立方体检测精度得到显著提高,分别提高24%、8.6%和7.3%,有效改善水下复杂环境中弱小目标漏检和误检问题。Visual detection of sonar images is a critical technology in complex water resource exploration and underwater foreign object target detection.Aiming at the problems of weak features and background information interference of small targets in sonar images,we propose a weak feature confocal channel modulation algorithm for underwater sonar target detection.First,in order to improve the model's ability to capture and characterize the information of weak targets,we design a weak target-specific activation strategy and introduce an a priori frame scale calibration mechanism to match the underlying semantic feature detection branch to improve the accuracy of small target detection;second,we apply the global information aggregation module to deeply excavate the global features of weak targets to avoid the redundant information from covering the small target's weak key features;finally,in order to solve the problem of traditional spatial pyramid pooling which is easy to ignore the channel information,the confocal channel regulation pooling module is proposed to retain effective channel domain small target information and overcome interference from complex background information.Experiment results show that the model in this paper achieves an average detection accuracy of 83.3%on nine types of weak targets in the underwater sonar dataset,which is 5.5%higher than the benchmark.Among these,the detection accuracy of iron buckets,human body models and cubes is significantly improved by 24%,8.6%,and 7.3%,respectively,effectively solving the problem of leakage and misdetection of weak targets in complex underwater environment.

关 键 词：弱小目标检测 水下声呐图像 全局信息聚合 共焦通道调控池化 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]