Wavelet convolution and multi-scale attention for Chinese Herbal Medicine recognition

The intelligent and precise recognition of Chinese Herbal Medicine is a critical challenge for advancing smart healthcare, particularly as artificial intelligence integrates with the modernization of traditional Chinese medicine. Existing methods struggle to model the multi-scale morphological characteristics of Chinese Herbal Medicine and achieve high classification accuracy under complex backgrounds. To address these limitations, we constructed the Chinese Herbal Medicine Multi-Morphology Image Dataset-50 (CHM-Morph-50), comprising images of 50 species with annotations for leaf texture, three-dimensional contour, and graded background complexity. Leveraging this dataset, we propose the Wavelet Convolution and Multi-Scale Attention Network (WCMSA-Net), which uses wavelet convolution to decompose images into high- and low-frequency components for parallel extraction of complementary details, combined with a multi-scale spatial pyramid attention mechanism to dynamically capture discriminative features. An improved label-smoothing focal loss with class-weight adjustment is introduced to mitigate class imbalance. Experimental results show that WCMSA-Net achieves a Top-1 accuracy of 88.6% on CHM-Morph-50, exceeding ResNet50 and MobileNet by 3.4% and 10.4%, respectively, and maintaining robustness under occlusion and blur. This study offers a high-accuracy, scalable framework for the digital recognition of Chinese Herbal Medicine, with potential applications in the identification of other medicinal plant species.