基于对抗注意力机制的水下遮挡目标检测算法  

Underwater Occlusion Target Detection Algorithm Based on Adversarial Attention Mechanism

作　　者：罗偲 李凯扬 吴吉花 任鹏[1] LUO Cai;LI Kaiyang;WU Jihua;REN Peng(College of Oceanography and Space Informatics,China University of Petroleum(East China),Qingdao 266580,Shandong,China)

机构地区：[1]中国石油大学(华东)海洋与空间信息学院,山东青岛266580

出　　处：《计算机工程》2024年第10期313-321,共9页Computer Engineering

基　　金：国家重点研发计划(2021YFE0111600)。

摘　　要：水下环境复杂,遮挡目标信息缺失严重而难以提取到足够的特征信息,导致水下遮挡目标易被漏检。为解决该问题,提出一种基于对抗注意力机制的水下目标检测算法。以Faster R-CNN算法为框架,提出基于空间注意力机制的对抗生成遮挡样本网络(AOGN)。AOGN与Faster R-CNN网络相互竞争,通过三阶段训练过程,在不增加推理负担的情况下学习生成检测网络难以正确区分的样本,提高Faster R-CNN网络对水下遮挡目标的检测精度。使用Focal loss增加困难样本的损失比重,解决水下数据集难易样本不平衡的问题。在此基础上,为获得更丰富的水下目标特征信息,使用SE-ResNet50代替VGG16作为骨干网络,通过残差网络和SE模块的结合获得更有效、更丰富的水下目标信息,提高对检测目标的特征提取能力,同时加入多条ROIpooling支路实现多尺度特征融合,增加特征的丰富性。实验结果表明,该算法在URPC数据集和水下垃圾数据集上分别取得了73.76%和86.85%的平均精度均值(mAP),遮挡目标漏检率分别达到2%和7%,相较于其他检测算法能够有效提升检测性能。The complexity of underwater environments and the severe lack of occluded target information,make the extraction of sufficient information difficult,resulting in a high omission factor for underwater occlusion targets.To solve this problem,the present study proposes an occluded underwater target detection algorithm based on an improved adversarial attention mechanism.Using Faster R-CNN as an adversary network,the Adversarial Occlusion sample Generation Network(AOGN),which has a competitive relationship with the Faster R-CNN is designed to improve the detection accuracy for occlusion targets.Through a three-stage learning process,AOGN learns how to generate samples that are difficult for the detection network to classify correctly,thereby improving the detection accuracy of the Faster R-CNN for underwater occlusion targets.Subsequently,the Focal loss function is used to increase the proportion of difficult samples in the loss.Finally,to solve the problem of low resolution of underwater images,SE-ResNet50 is used as the backbone in place of VGG16,thereby enhancing the feature extraction ability.Furthermore,multi-scale feature fusion is adopted based on multi-ROIpooling branches to increase the richness of features.The proposed algorithm achieves mean Average Precision(mAP)values of 73.76%and 86.85%and omission factor values of 2%and 7%,on the URPC and underwater common trash datasets,respectively.These results demonstrate that the algorithm effectively outperforms existing detection methods.

关 键 词：机器视觉 水下目标检测 对抗样本 损失函数 SE-ResNet50网络 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]