船舶采集图像智能分类研究  

Research on intelligent classification of ship images based on classification algorithm

作　　者：苑靖国 YUAN Jing-guo(Tianjin Maritime Collage,Tianjin 300350,China)

机构地区：[1]天津海运职业学院,天津300350

出　　处：《舰船科学技术》2022年第22期158-161,共4页Ship Science and Technology

摘　　要：为实现不同角度和不同距离下,船舶采集图像的智能分类,提出基于多尺度注意力深度卷积神经网络分类算法的船舶采集图像智能分类。将采集的船舶图像输入该网络中,网络的多尺度深度卷积层采用3个多尺度特征注意力模块结合深度残差模块,提取船舶采集图像不同层次的局部不变性特征;池化层对该特征转换处理后形成特征向量;全连接层池化层引入尺寸匹配函数融合特征向量,形成多尺度纹理特征向量并输入分类层,实现船舶采集图像智能分类。测试结果显示:该方法可实现不同船舶类别图像特征提取,gini指数结果均在0.963以上,可依据分类需求,实现不同角度以及距离条件下、不同的船舶图像类别的准确分类。In order to realize the intelligent classification of ship image acquisition under different angles and different distances,an intelligent classification of ship image acquisition based on multi-scale attention deep convolutional neural network classification algorithm is proposed.The collected ship images were input into the network,and the multi-scale deep convolution layer of the network used three multi-scale feature attention modules combined with the depth residual module to extract the local invariance features at different levels of the ship image acquisition.The pooling layer transforms the feature and forms the feature vector.In the pooling layer of the fully connected layer,the size matching function is introduced to fuse the feature vectors,and the multi-scale texture feature vectors are formed and input into the classification layer to realize the intelligent classification of ship image acquisition.The test results show that the proposed method can achieve image feature extraction of different ship categories,and the gini index results are all above 0.963.According to the classification requirements,the method can achieve accurate classification of different ship image categories under different angles and distances.

关 键 词：分类算法 船舶采集图像 智能分类 局部不变性特征 多尺度 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]