Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (12): 155-162.DOI: 10.3778/j.issn.1002-8331.2012-0089

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Improved Ship Target Detection Method for RFBnet Network

FANG Jian, LIU Kun   

  1. College of Information Engineering, Shanghai Maritime University, Shanghai 201306, China
  • Online:2022-06-15 Published:2022-06-15

改进RFBnet网络的船只目标检测方法

方健,刘坤   

  1. 上海海事大学 信息工程学院,上海 201306

Abstract: In order to solve the problem of ship target missing and classification error, a new method of natural image target detection based on improved RFBnet(I-RFBNET) is proposed. Deconvolution feature fusion module and pooling feature fusion module are used for feature fusion to form six new effective feature layers. A step long convolutions method is proposed to extract the concerned area information of the feature unit in the original image. The dilate convolutions block, which integrates the attention mechanism, and the new first three effective feature layers are designed for feature fusion. Focusing classification loss function is introduced to solve the problem of unbalanced distribution of positive and negative samples during training. The model of SeaShips is saved after training SeaShips, a data set of sizing ships. The experimental results show that the improved algorithm has a good detection effect, especially for small targets under multi-target occlusion. The average accuracy is 96.26%, 4.74 percentage points higher than the previous algorithm, and the speed reaches 26 FPS (frame per second), meeting the need for real-time detection.

Key words: ship detection, attention mechanism, expansion convolution module, feature fusion, small target

摘要: 针对目前舰船目标检测中,多目标情况下的舰船目标很容易被多目标遮挡,造成舰船目标漏检、分类错误等问题,提出了一种基于改进RFBnet(I-RFBnet)的自然图像目标检测方法。使用池化特征融合模块(PFF)和反卷积特征融合模块(DFF)进行特征融合,形成新的六个有效特征层。提出一种跨步长卷积方式来提取特征单元在原图中的关心区域信息,设计了融入注意力机制的膨胀卷积模块(dilate convolutions block,DB)和新的前三个有效特征层再次进行特征融合。引入聚焦分类损失函数解决训练过程中正负样本分布不均衡的问题;最后通过对规模船只检测数据集SeaShips训练后,保存其模型。实验结果表明:改进后的算法检测效果良好,尤其在多目标遮挡下的小目标效果显著。平均精度均值为96.26%,比改进前的算法提高了4.74个百分点,帧率达到26?FPS(frame per second),满足实时检测的需求。

关键词: 舰船检测, 注意力机制, 膨胀卷积模块, 特征融合, 小目标