Identification of wormholes in soybean leaves based on multi-feature structure and attention mechanism

Soybean leaf wormholes resulting from crop pest infestations can adversely affect crop quality. Traditional manual and machine learning recognition algorithms designed to detect wormholes cannot meet testing accuracy and speed requirements, due to complex factors related to environmental conditions, dense planting practices, and leaf wormhole pattern diversity. To address this problem, here an improved method for identifying soybean pests is proposed that incorporates a YOLO-v5s (You Only Live Once) network model utilizing an improved wormhole-recognition algorithm for enhanced detection of soybean leaf wormholes as a measure of pest infestation severity. This algorithm could effectively recognize leaf wormholes regardless of leaf multi-blade morphological diversity, due to the incorporation of a sample transformation method that reduced rates of false positives and misses through elimination of redundant bounding boxes. Results obtained using the improved model to analyse a soybean sample data set assembled here from test data revealed that the improved YOLO-v5s model-based method achieved an average accuracy rate of 95.24% that was 2.50%, 12.13%, and 2.81% higher than respective results obtained using algorithms based on faster R-CNN, YOLO-v3, and YOLO-v5s models. In addition, the improved model required a storage space size of only 15.1 MB and achieved a data transmission rate of 91 frames per second (f/s). Therefore, the method proposed here achieved significantly improved wormhole recognition accuracy and speed and required only minimal resources for model deployment as a superior method than currently used methods for use in soybean wormhole identification.