Few-Shot Image Classification Based on Asymmetric Convolution and Attention Mechanism

Abstract

Deep learning has achieved great success in the field of image processing, but it is often difficult to achieve good results when the amount of annotation data is small, and it is prone to overfitting. Few-shot learning can effectively solve this problem. The existing few-shot image classification models usually lead to a large computational loss, and the design of the network structure is complicated and costly. In addition, they do not pay attention to the key features of the image. To solve the above problems, this paper proposes a few-shot image classification method based on asymmetric convolution and attention mechanism. Through replacing the standard $3\times3$ convolution kernel with asymmetric convolution blocks, the feature extraction ability of the model is improved without increasing additional computational consumption. Secondly, the SENet (Squeeze-and-Excitation Networks) attention mechanism is introduced to explicitly construct the correlation between feature channels, so that the salient features of the image are highlighted and emphasized. Finally, the K-Nearest Neighbor (KNN) is used to calculate the similarity between the query image and the category. A large number of experimental results show that the proposed algorithm performs well in the task of few-shot image classification, which verifies the effectiveness and superiority of the algorithm.