基于通道注意力机制的小样本SAR飞机图像分类方法OA北大核心CSTPCD

Synthetic Aperture Radar(SAR)has become an important device in earth observation because of its all-weather and all-time service,high resolution and wide width,and image classification is an important direction of SAR image interpretation.Compared with the optical image,the imaging mechanism of the SAR image is more complex.There are more noise interference,resulting in poor image clarity and difficulty in sample labeling,which can not guarantee the sample size requirements of the depth learning algorithm.In this context,how to classify few-shot SAR images has become one of the key research issues in the field of SAR image interpretation.To solve this problem,this paper carries out the research of SAR image classification model based on meta-learning,hoping to achieve high-precision recognition of SAR images under the condition of few-shot.A prototypical net classification method based on attention mechanism is constructed,and the importance of automatic acquisition of image features by channel attention module is designed to promote the extraction of features that are more discriminative to image classification.At the same time,a pretraining network is designed for the model to make full use of the information of existing data and learn better priori information,so as to improve the accuracy of classification.Experiments are carried out on the few-shot classification model on the self-built high-resolution SAR image dataset.The ablation experiment shows that both the attention module and the pretraining module improve the performance of the model to a certain extent.Experimental results show that compared with the commonly used few-shot learning methods,the classification method constructed in this paper achieves higher accuracy in SAR image classification,the classification accuracy of the 5-way 1-shot experiment in the first group is improved by 5.9％,and the classification accuracy of the 5-way 5-shot experiment is improved by 1.92％.