Wavelet-Attention CNN for image classification

The feature learning methods based on convolutional neural network (CNN) have successfully produced tremendous achievements in image classification tasks. However, the inherent noise and some other factors may weaken effectiveness of statistics. In this paper, we investigate Discrete Wavelet Transform (DWT) frequency domain design a new Wavelet-Attention (WA) block to only implement attention high-frequency domain. Based this, propose (WA-CNN) for classification. Specifically, WA-CNN decomposes maps into low-frequency components storing structures basic objects, as well detailed information noise, respectively. Then, WA is leveraged capture with different but reserves object Experimental results CIFAR-10 CIFAR-100 datasets show that our proposed achieves significant improvements accuracy compared related networks. MobileNetV2 backbones, 1.26% Top-1 improvement benchmark 1.54% benchmark.