摘要
文中通过在深度卷积神经网络的特征提取模块中引入注意力机制,实现不同通道之间的信息交互,缓解船名位置分布多样导致漏检的问题.通过特征金字塔增强模块融合低高级信息来加深不同尺度的特征,基于特征融合模块将不同深度的特征金字塔增强模块产生的特征聚集成最终的特征进行分割,在提高检测精度的同时降低模型复杂度.利用smooth L1损失函数解决船名预设框的位置回归问题.水面船舶船名检测结果表明:改进后的算法在自建船名数据集上检测精确率达到88.1%,相较于现有优势算法DBNet提高了3.4%.
By introducing attention mechanism into the feature extraction module of deep convolutional neural network,the information interaction between different channels was realized,and the problem of missing detection caused by the diverse distribution of ship names was alleviated.The feature pyramid enhancement module fuses low-level and high-level information to deepen the features of different scales.Based on the feature fusion module,the features generated by the feature pyramid enhancement module with different depths were aggregated into the final features for segmentation,which improved the detection accuracy and reduces the model complexity.Smooth L1 loss function was used to solve the position regression problem of ship name preset box.The results of ship name detection on the surface show that the detection accuracy of the improved algorithm on the self-built ship name data set reaches 88.1%,which is 3.4%higher than the existing dominant algorithm DBNet.
作者
甘浪雄
吴金茹
徐海祥
冯辉
张磊
束亚清
张东方
GAN Langxiong;WU Jinru;XU Haixiang;FENG Hui;ZHANG Lei;SHU Yaqing;ZHANG Dongfang(School of Navigation,Wuhan University of Technology,Wuhan 430063,China;Hubei Key Laboratory of Inland Shipping Technology,Wuhan 430063,China;School of Naval Architecture,Ocean and Energy Power Engineering,Wuhan University of Technology,Wuhan 430063,China;Suzhou Port and Shipping Development Center,Suzhou 215000,China)
出处
《武汉理工大学学报（交通科学与工程版）》
2023年第5期850-855,共6页
Journal of Wuhan University of Technology（Transportation Science & Engineering）
关键词
船名检测
注意力机制
特征增强
卷积神经网络
ship name detection
attention mechanism
feature enhance
convolutional neural network
