基于注意力机制的水下遮挡目标检测算法  

Underwater Occlusion Target Detection Algorithm Based on Attention Mechanism

作　　者：史建柯 乔美英[1] 李冰锋[1] 赵岩 SHI Jianke;QIAO Meiying;LI Bingfeng;ZHAO Yan(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454000,China)

机构地区：[1]河南理工大学电气工程与自动化学院,河南焦作454000

出　　处：《电子科技》2023年第5期62-70,共9页Electronic Science and Technology

基　　金：国家自然科学基金(41672363)。

摘　　要：针对水下目标检测任务中存在前景遮挡和背景模糊的问题,文中提出一种基于注意力机制的水下目标检测算法。首先采用图像增强算法改善图像质量。然后在非局部神经网络的相似度函数基础上,融合具有逻辑推理能力的级联相似度函数,增强网络对全局上下文特征的表达能力。随后将改进型非局部神经网络与三分支注意力融合,弥补非局部神经网络丢失的通道特征。最后利用空洞卷积模块置换三分支注意力中的池化操作,减少细粒度信息损失。实验表明,该算法在2020年全国水下目标检测大赛提供的数据集上,使基线方法检测精度由65.66%增长至68.55%,证明了所提算法的有效性。In view of the problems of foreground occlusion and background blur in underwater target detection task,an underwater target detection algorithm based on attention mechanism is proposed.Firstly,the image enhancement algorithms are used to improve the image quality.Then,based on the similarity function of the non-local neural network,the concatenation similarity function with logical reasoning capability is fused to enhance the expression ability of the network to the global context features.Additionally,the improved non-local neural network is combined with the triplet attention to make up for the channel features lost by the non-local neural network.Finally,the dilated convolution module is used to replace the pooling operation in triplet attention to reduce the loss of fine-grained information.Experiments show that the proposed algorithm increases the detection accuracy of baseline method from 65.66%to 68.55%on the data set provided by the 2020 National Underwater Target Detection Contest,which proves the effectiveness of the proposed algorithm.

关 键 词：深度学习 卷积神经网络 水下目标检测 遮挡目标检测 注意力机制 相似度函数 空洞卷积 Faster R-CNN 

分 类 号：TN957.52[电子电信—信号与信息处理] TP391[电子电信—信息与通信工程]