细粒度遥感舰船开集识别

Fine-grained remote sensing ship open set recognition

为了解决传统深度卷积神经网络在舰船图像细粒度分类中的局限性,本文设计了细粒度遥感舰船开集识别模型。首先,引入了基于注意力机制的STN模块,加在特征提取网络前用来过滤背景信息;然后在STN模块后接一个多尺度的并行的卷积结构,强化网络对不同尺度的局部区域的特征提取能力;接着将提取到的特征分别输入基分支和元嵌入分支,用来增大类间方差和减小类内方差,同时强化模型对尾类小样本的学习;最后对两个分支的分类结果进行决策融合,根据设定的阈值判别已知类和未知类进一步对已知类进行细分。在平衡与不平衡分布的FGSCR-42数据集上进行了4种开放度实验,结果表明:在平衡分布的数据集上4种开放度的平均准确率为90.5%,86.3%,85.7%,85.1%,不平衡分布数据集的平均准确率为90.0%,85.1%,84.3%,84.1%。与当前主流的舰船识别方法相比,本文方法分类具有更高的识别准确率和更好的泛化能力。

In this study,a fine-grained remote sensing ship open-set recognition model is designed to ad⁃dress the limitations of traditional deep convolutional neural networks in fine-grained classification of ship images.First,a STN module based on attention mechanism is introduced before the feature extraction net⁃work to filter background information.In addition,a multi-scale parallel convolution structure is added af⁃ter the STN module to enhance the feature extraction ability of the network for local regions of different scales.The extracted features are input into the base and meta-embedded branches,to increase inter-class variance and reduce intra-class variance,strengthening the model's learning of the tail class small samples concomitantly.Finally,the classification results of the two branches are fused;known and unknown classes are distinguished according to the set threshold;and known classes are subdivided.Four types of openness experiments were conducted on the FGSCR-42 datasets with balanced and unbalanced distribu⁃tions.The results show that the average accuracies of the four types of openness in the balanced distribu⁃tion dataset are 90.5%,86.3%,85.7%,and 85.1%;the corresponding average accuracies of the un⁃balanced distribution dataset are 90.0%,85.1%,84.3%,and 84.1%.Compared with the current mainstream ship recognition methods,the proposed method has higher recognition accuracy and better generalization ability.