Tomato yellow leaf curl virus detection based on cross-domain shared attention and enhanced BiFPN

Accurate identification of Tomato Yellow Leaf Curl Virus (TYLCV) is essential for ensuring the sustainability of tomato cultivation, as tomatoes are a globally important commercial crop. TYLCV manifests as stunted growth, yellowing, and curling of leaves, severely affecting tomato yield and quality. Traditional manual identification methods are not only costly but also inefficient. Hence, computer-assisted detection may serve as an effective means to ensure the long-term sustainability of tomato production. Existing models face challenges such as diverse lesion characteristics, detection of small targets, uneven illumination, and occlusions in complex orchard backgrounds, which lead to low recognition accuracy and poor adaptability. Based on the YOLOv8n model, we propose the YOLOv8n-CDSA-BiFPN model for detecting TYLCV. We introduce an Adaptive Random Mix-Cut Splicing (ARMS) image augmentation technique that merges affected leaf images with healthy leaves, enhancing the training and validation datasets through noise addition, geometric transformations, color changes, and affine transformations, thus increasing the diversity of background targets. A Cross-Domain Shared Attention mechanism (CDSA) deconstructs the backbone network’s input feature map into two parts, each assigned different attention weights for spatial and channel domains, facilitating shared computation and redistribution, and reconstructing key information to capture subtle semantic features of leaf yellowing and curling. The network’s neck incorporates an enhanced Bidirectional Feature Pyramid Network (BiFPN), integrating feature maps of various scales to boost early symptom detection capabilities. Moreover, the introduction of Inner-CIOU with a scale factor-controlled auxiliary bounding box optimizes the precision and regression efficiency in detecting small target features. Experimental results demonstrate that the model achieves an average accuracy of 88.25%, a recall rate of 86.59%, and a Mean Average Precision (mAP) of 85.34%, significantly outperforming the original YOLOv8n and mainstream detection models such as YOLOv5s, SSD, and Faster-RCNN. This model exhibits exceptional performance in detecting TYLCV in complex environments, providing robust technical support for tomato growth monitoring and offering insights for the detection of other crop diseases. Our work can be accessed at https://github.com/mohenghui/yolov8_CDSA .