YOLOv8-Rice: a rice leaf disease detection model based on YOLOv8

• Published in: Paddy and water environment Vol. 22; no. 4; pp. 695 - 710
• Main Authors: Lu, Yu, Yu, Jinghu, Zhu, Xingfei, Zhang, Bufan, Sun, Zhaofei
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 01.10.2024; Springer Nature B.V
• Subjects: Adaptability; Agriculture; Biomedical and Life Sciences; Crop yield; Deformation; Disease; Disease detection; Disease recognition; Ecotoxicology; Experimental research; Feature extraction; Food sources; Formability; Geoecology/Natural Processes; Hydrogeology; Hydrology/Water Resources; Laboratory experimentation; Leaves; Life Sciences; Object recognition; Performance evaluation; Plant diseases; Rice; Soil Science & Conservation; Deep learning; Attention mechanism; Loss function; Rice disease; Deformable convolution; Object detection
• ISSN: 1611-2490; 1611-2504
• DOI: 10.1007/s10333-024-00990-w

Rice, being an important global food source, is susceptible to diseases during its growth, resulting in a negative impact on its yield. Existing models for rice disease detection have limitations in recognizing small-sized and irregularly shaped disease types. To address this issue, we propose a new model called YOLOv8_Rice, specifically designed for rice leaf disease detection based on the YOLOv8n object detection model. Firstly, we conducted experimental research to investigate the influence of various common attention mechanisms on the performance of YOLOv8. The aim was to optimize the model’s ability to extract features from different types of targets. Secondly, we enhanced the model’s adaptability to target deformation and spatial changes by incorporating deformable convolutions to improve the C2f module structure in the YOLOv8 model. Furthermore, we replaced the network structure of YOLOv8 with a weighted bidirectional feature pyramid network to achieve weighted feature fusion, aiming to improve model performance and reduce computational complexity. Finally, we replaced the IOU loss function design in the YOLOv8 model with Wise IOU to provide more accurate evaluation results. In comparison to YOLOv8n, our YOLOv8_Rice model achieved an average precision increase of 15.8% and an mAP@0.5 improvement of 18.7% while reducing GFLOPs by 24.7% during testing on the rice disease dataset. These results indicate that YOLOv8_Rice has significant potential for global rice disease detection applications.