Few-Shot Incremental Learning for Aerial Image Scene Classification Based on Feature Adaptation and Prototype Continuous Optimization

Abstract

Aerial image scene classification plays a significant role in many applications, however, it generally faces the problem of lacking labeled samples and incremental classification tasks. Existing deep networks directly applied for incremental learning may lead to knowledge forgetting of old categories and overfitting to identify novel category. In this paper, a few-shot incremental learning (FSIL) model based on feature adaptation and prototype continuous optimization is presented for aerial image scene classification. In the proposed framework, a feature extractor and a cosine distance classifier based on residual networks are constructed firstly. Then, a feature adaptation module is developed to calculate relation coefficient between training samples and test samples of incremental stage in order to adjust feature weights. In addition, a prototype continuous optimizer is proposed to correct prototype features to enhance the discrimination during incremental stage and reduce knowledge forgetting of old categories. Experimental results on two datasets verify that the presented method is effective for few-shot incremental aerial scene image classification(FS-IASIC).