An Improved Real-Time Lightweight YOLOv8 Algorithm for Bloom Maturity Recognition in Torreya Grandis

To address the challenge of rapidly and accurately detecting male cones of Chinese Torreya at various stages of maturity in natural environments, the research team proposes a target detection algorithm, GFM-YOLOV8s, based on an improved YOLOv8s. By utilizing the YOLOv8s network model as the foundation, we replace the backbone feature extraction network with c2f-faster-ema to lighten the model and simultaneously enhance its ability to capture and express important image features. Additionally, the PAN-FPN feature extraction structure in the neck is substituted with a BiFPN structure. By removing less contributive nodes and adding cross-layer connections, the algorithm achieves better fusion and utilization of features at different scales. The WIoU loss function is introduced to mitigate the mismatch in orientation between the predicted and ground truth bounding boxes. Furthermore, a structured pruning strategy was applied to the optimized network, significantly reducing redundant parameters while preserving accuracy. Results: The improved GFM-YOLOV8 has a detection accuracy of 88.2% for Torreya male cones, the detection time of a single image is 8.3 ms, and the model size is 4.44 M, fps is 120 frames, parameters is 2.20´106. Compared with the original YOLOv8s algorithm, map50 and recall are increased by 2.0% and 2.0% respectively, and the model size and model parameters are reduced by 79.2% and 80.1% respectively. The refined lightweight model can swiftly and accurately detect male cones of Torreya at different stages of maturity in natural settings, providing technical support for the visual recognition system used in growth monitoring at Torreya bases.