Biformer Attention and ASF-YOLO for Cordyceps Sinensis Target Recognition

Abstract

Cordyceps sinensis, a highly valued traditional Chinese medicine, faces challenges in collection due to inefficiencies in manual searching, strenuous labor, and the impact of subjective expertise. The integration of deep learning into Cordyceps sinensis identification is an unexplored area. To alleviate the manual labor and enhance the precision and speed of identifying Cordyceps sinensis, a novel detection approach that combines attention mechanisms with the ASF-YOLO model has been developed. This approach replaces the Spatial Pyramid Pooling Fast (SPPF) with a Context Augmentation Module (CAM) and swaps the original C3 model with a lighter model, C3-Faster, which is based on FasterNet. Additionally, it incorporates the Bi-level Routing Attention (BiFormer) mechanism and a Context Integration module to better detect smaller targets and increase accuracy. For the detection of tiny Cordyceps sinensis targets against intricate backgrounds, a novel fusion framework, ASF-YOLO, which leverages attention scale sequence fusion, has been introduced to boost detection accuracy further. Through experimental verification, the average accuracy rate (MAP) for Cordyceps sinensis can reach 99.2, 0.6% higher than that of traditional YOLOv5. The enhanced YOLOv5 boasts an average detection accuracy of up to 95% and it could identify some cordyceps sinensis that could not be identified by traditional YOLOv5.