基于注意力机制的水下目标检测算法  

Underwater object detection algorithm based on attention mechanism

作　　者：赵晓飞 于双和 李清波[2] 阎妍 赵颖 ZHAO Xiaofei;YU Shuanghe;LI Qingbo;YAN Yan;ZHAO Ying(College of Marine Electrical Engineering,Dalian Maritime University,Dalian 116026,China;College of Environmental Sciences and Engineering,Dalian Maritime University,Dalian 116026,China)

机构地区：[1]大连海事大学船舶电气工程学院,辽宁大连116026 [2]大连海事大学环境科学与工程学院,辽宁大连116026

出　　处：《扬州大学学报（自然科学版）》2021年第1期62-67,共6页Journal of Yangzhou University：Natural Science Edition

基　　金：国家重点研发计划“海洋环境安全保障”资助项目(2018YFC1406405);国家自然科学基金资助项目(62073054,62003070);中国博士后科学基金资助项目(2020M680037,2020M680930)。

摘　　要：针对传统水下目标检测算法识别精度低的问题,提出一种基于注意力机制的水下目标检测算法(feature refinement and attention mechanism network,FRANet).该算法采用特征融合模块和特征增强模块相结合的方式,使用卷积神经网络提取目标的多尺度特征.同时引入一种由锚框精化模块、空间注意力模块和目标检测模块组成的级联注意力机制方案,通过空间注意力机制解决目标类别的不平衡性,改善水下小目标的分类性能和回归性能.试验表明,利用FRANet算法对水下小目标进行识别的平均精度均值(mean average precision,mAP)达80.5%,验证了算法的有效性,为水下目标识别提供了一种新的研究思路与方法.A FRANet(feature refinement and attention mechanism network)underwater object detection algorithm is proposed to address the problem of low recognition accuracy of traditional underwater object detection algorithms.The algorithm uses a combination of a feature fusion module and a feature enhancement module to extract multi-scale features of the object using a convolutional neural network.A cascaded attention mechanism scheme consisting of an anchor refinement module,a spatial attention module and an object detection module is also introduced to improve the classification performance and regression performance of small underwater object by addressing the imbalance of object classes through the spatial attention mechanism.The experiments show that the mean average precision(mAP)of underwater small object recognition using FRANet reaches 80.5%,which verifies the effectiveness of the underwater object detection algorithm based on the attention mechanism and provides a new research idea and method for underwater object recognition.

关 键 词：深度学习 水下目标检测 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]