基于卷积通道注意力及改进ResNet的遥感场景分类

Remote Sensing Scene Classification Based on Convolutional Channel Attention and Improved ResNet

在遥感场景分类中,针对场景图像具有高类内差异性和类间相似性的问题,提出一种改进ResNet50的分类方法。以ResNet50为骨干网络,通过构建并行下采样结构来替代ResNet50的第1个层块,以充分提取浅层信息,同时降低参数数量,提升模型的训练和推理速度;设计轻量级的卷积通道注意力(CCA)模块,将其嵌入到后续的每一个特征提取阶段中,使得模型可以在每个阶段中更好地捕捉重要的空间和通道信息并增强模型对输入的关注能力。在RSSCN7遥感场景分类公开数据集上,分类准确率相比于ResNet50提升了4.43%,参数量减少了45.44%。试验结果表明:并行下采样结构和CCA能显著降低模型复杂度和提升分类性能。

In remote sensing scene classification,an improved ResNet50 classification method is proposed to address the problem of high intra-class variability and inter-class similarity for scene images.Taking ResNet50 as the backbone network,a parallel downsampling structure is constructed to replace the first layer block of ResNet50,which is used to fully extract the shallow information,and reduce the number of parameters to speed up the training and inference speed of the model;Then,a lightweight convolutional channel attention(CCA)module is designed and embedded into each subsequent feature extraction stage,so that the model can better capture the important spatial and channel information,and improve the model's ability to pay attention to the inputs.On the RSSCN7 Remote Sensing Scene Classification public dataset,the classification accuracy is improved by 4.43%and the number of parameters is reduced by 45.44%compared to ResNet50.The experimental results show that the parallel downsampling structure and CCA can significantly reduce the model complexity and improve the classification performance.