Abstract
Cordyceps sinensis (C. sinensis) is a valuable herbal medicine with wide-ranging applications. However, automating C. sinensis recognition is challenging due to the high morphological similarity and limited phenotypic variation among its subspecies. In this paper, we propose a novel approach called Progressive Feature Learning Network (PFL-Net) that mines multiple biological features to recognize different subspecies. Firstly, to comprehensively capture multi-scale discriminative features of C. sinensis, we propose the Spatial-aware Semantic Refinement Module (SSRM), which constructs discriminative feature groups by utilizing relative positions to model the intrinsic feature relations. Secondly, the Multi-scale Collaborative Perception Module (MCPM) avoids isolated biological features during modeling by establishing relations between different feature groups to enhance the recognition integrity of C. sinensis. Furthermore, to prevent the model from focusing on the same discriminative regions of C. sinensis, we propose a Channel Decouple (CD) loss that decouples features along the channel dimension, enhancing the diversity of C. sinensis discriminative features. In addition, we construct a C. sinensis dataset (CSD) to facilitate the application of biometric recognition, representing the first study focused on fine-grained C. sinensis recognition. Extensive experiments conducted on the CSD and three benchmark datasets validate the effectiveness of our proposed method, achieving a top-1 accuracy of 94.43% on the CSD dataset, which surpasses all existing approaches.