【摘要】 遥感船舶图像细粒度分类的难点在于类间差异小和类内差异大，并且该领域公开可用的数据集太少，常规的数据增强方法效率低且效果不够好。为了解决上述问题，提出一种基于注意力机制的遥感船舶图像分类网络。首先，利用CBAM注意力机制生成每张训练图的注意力图以突出目标的显著特征部分；其次，通过注意力引导的区域剪裁和区域删除两种方式进行数据增强；最后，将原图和增强后的图片输入进行训练。在数据集FGSCR-42上对该方法进行验证，实验结果表明，该方法超越了其他现有模型，有效提升了遥感船舶图像细粒度分类精度。更多还原

【Abstract】 The difficulty of fine-grained classification of remote sensing ship images is that the differences between classes are small and the differences within classes are large, and there are too few publicly available datasets in this field, and conventional data enhancement methods are inefficient and not effective enough. In order to solve the above problems, an attention mechanism-based remote sensing ship image classification network is proposed. Specifically, the CBAM attention mechanism is first used to generate the attention map of each training map to highlight the salient feature parts of the target, and then the data augmentation is carried out through attention-guided region clipping and attention-guided region deletion. Finally, the original images and the augmented images are inputted for training. The method is verified on the dataset FGSCR-42. The experimental results show that the method surpasses other existing models and effectively improves the accuracy of the fine-grained classification of remote sensing ship images.更多还原