Lightweight Multi-Path Wavelet Transform Attention and Adaptive Downsampling for Enhanced Convolutional Neural Networks

Feature extraction is crucial in machine learning and computer vision, enabling tasks such as pattern recognition and classification. Convolutional Neural Networks have demonstrated excellent performance in image classification and object detection but face challenges like limited receptive fields, spatial invariance, and insufficient cross-channel interactions. To address these limitations, we propose a novel, lightweight multi-path wavelet transform attention mechanism. This mechanism integrates semantic information and texture details of images, enriching spatial attention and achieving comprehensive feature aggregation. In terms of texture details, by utilizing Haar wavelet transforms, our method effectively captures local details in approximation, horizontal, vertical, and diagonal directions, enhancing the model's sensitivity to complex spatial relationships. A multi-scale feature selection module dynamically selects information-rich features from various decomposition levels, improving the model's ability to understand fine details and global structures. In addition, we introduce a light adaptive weight downsampling module (LAWDS) based on discrete wavelet transform and channel attention mechanism to preserve edge and texture details during downsampling. These designs are lightweight and can be easily integrated into existing networks with minimal computational overhead. Extensive experiments on CIFAR and ImageNet demonstrate clear improvements in classification performance, while downstream tasks such as semantic segmentation also benefit from our approach. Comprehensive ablation studies validate the effectiveness of the proposed attention mechanism, highlighting its potential to enhance CNN-based models in various computer vision applications.