TB-MYOLO:融合MLP与YOLOv5的双分支网络船舶目标检测方法

TB-MYOLO:A Two-Branch Network for Remote Sensing Ship Detection Method Fused with MLP and YOLOv5

针对复杂场景下可见光卫星遥感影像的多类别船舶目标检测中易产生误检、漏检的问题,基于YOLOv5算法改进,提出了一种融合MLP的双分支网络船舶目标检测方法TB-MYOLO.该算法通过引入分支网络作为辅助以增强小目标的特征表达能力,分支网络只负责关注小目标物体的学习,携带更多的浅层位置信息.将分支网络学习到的小目标特征向量与主干网络学习到的小目标特征向量相融合,使小目标特征在网络中占有更大的比重,以此增强模型对目标位置的特征学习能力.同时使用MLP模块代替原始YOLOv5的SPPF模块,利用MLP的特征长依赖性的特点对深层网络的特征向量进行筛选加权,突出重点信息,避免了池化层带来的信息损失.实验结果表明,相比原始YOLOv5算法,改进后的TB-MYOLO算法显著提升了小目标类别的召回率和定位精度.对于复杂场景下可见光卫星遥感影像的多类别船舶目标检测,平均准确率mAP50达到了80.8%,相比原始YOLOv5、Retinanet和Faster R-CNN算法、改进后的TB-MYOLO算法,分别提升了2.4%、24.5%和28.1%.

Aiming at the problem of false detection and missed detection in the multi-target detection of visible light satellite remote sensing images in complex scenes,we propose a two-branch ship detection network TBMYOLO fused with MLP based on the improvement of the YOLOv5 algorithm.The algorithm enhances the feature expression ability of small objects by introducing a branch network as an assistant.The branch network is only responsible for the learning of small objects and carries more shallow position information.The small object feature vector learned by the branch network is fused with the small object feature vector learned by the backbone network,then the small object feature occupies a larger proportion in the network,so as to enhance the model’s feature learning ability for the object location.At the same time,the MLP module is used to replace the SPPF module of the original YOLOv5,and the feature vector of the deep network is screened and weighted by using the feature of MLP long feature dependence,which highlights the key information and avoids the information loss caused by the pooling layer.The experimental results show that,compared with the original YOLOv5 algorithm,the improved TB-MYOLO algorithm significantly improves the recall rate and localization accuracy of small object categories.For multi-target ship detection in visible light satellite remote sensing images in complex scenes,the average precision mAP50reached 80.8%.Compared with the original YOLOv5,Retinanet and Faster R-CNN algorithms,and the improved TB-MYOLO algorithm,that value has improved by 2.4%,24.5%and 28.1%respectively.