An Improved YOLOv8-Based Method for Detecting Citrus Leaf Disease

Considering the intricate morphology and diverse disease types of citrus leaves, along with the slow detection speed associated with traditional manual observation methods, this study proposes a novel approach for detecting citrus leaf diseases utilizing an enhanced YOLOv8 model. Initially, the Multi-Head Self-Attention (MHSA) module is integrated following the SPPF layer within the network architecture. This multi-head mechanism allows the model to attend to input features from multiple distinct perspectives, thereby improving performance in scenarios involving small objects, complex backgrounds, and occlusions. Subsequently, certain convolutional networks within some C2f modules are substituted with DCNv2 deformable convolutional networks. The core principle of deformable convolution involves introducing offsets to standard convolutional networks, enabling them to adapt more effectively to the geometric shapes of input features and further enhancing the model's adaptability to disease-affected leaf structures. Finally, experimental results indicate that the improved YOLOv8 model exhibits varying degrees of enhancement in metrics such as P value, recall rate, [email protected], and [email protected] compared to baseline models; additionally, both training and validation loss values have shown a reduction.