WRAN: A Plug and Play Wide Residual Attention Network Based on Few-shot Classification for Image Reconstruction

The classification of distorted images is crucial in the field of distorted inpainting. However, most of the existing methods for distorted image classification rely on fully supervised learning, which demands a substantial amount of manually labeled image data sets. The practical application of image classification has been hindered by certain obstacles. While small sample learning-based image classification methods can help solve this problem, existing methods fail to achieve high classification accuracy due to the significant amount of occlusion in distorted images. This article proposes a new method for small sample image classification that can accurately classify distorted images. The method employs two modules: the Self-Attention Wide Residual Network (WARN) and the TL module. Firstly, we introduced channel attention and spatial self-attention mechanisms in the WRAN module to enhance the ability to extract critical features from distorted images. Next, the TL module further improves the accuracy of cosine similarity by using anti-sine functions and low-pass filters. Finally, we conducted experiments on the proposed distorted image dataset, and the results showed that our proposed method achieved a classification accuracy of 70.04% at 1 shot and an accuracy of 81.95% at 5 shot, representing optimal performance with improvements of 1.46% and 4.04%, respectively, compared to the suboptimal method. Our experimental code address is: https://github.com/shliar/WRAN_net.git.