Rapid and accurate identification of crop diseases in the early stage is an important guarantee to reduce crop economic losses. In view of the actual production environment, crop yellow leaf curl virus (YLCV) cannot be accurately and quickly identified by color or texture features by traditional image processing algorithms in the early stage of disease, and the YOLO v5s general model has poor recognition effect and low efficiency in complex environments. The dataset was made from two sources: the images of single diseased leaves in the public dataset of Plant Village and the canopy images of diseased crop taken by mobile phones in the actual production, and manually labeled the diseased leaves in the images to achieve the correct identification of targets in complex terrain background and leaf occlusion, that was, to accurately identify all diseased leaves in healthy leaves, diseased leaves, withered leaves, weeds and soil. In addition, a smartphone was used to shoot images at the production site, there would be a variety of factors such as mobile phone resolution, light, shooting angle, etc., which would lead to problems such as reduced recognition accuracy, and it was necessary to preprocess data and enhance the collected images to improve the model recognition rate, and enhance the extraction ability of YOLO algorithm to key information by repeatedly increasing the CA attention mechanism module (coordinate attention) for many times on the YOLO v5s original model backbone network. The weighted bidirectional feature pyramid network (BiFPN) was used to enhance the fusion ability of different feature layers of the model, thereby improving the generalization ability of the model, replacing the loss function EIoU (Efficient IoU loss), further optimizing the algorithm model, and realizing the comprehensive improvement of the target recognition performance of the system after multi-method superposition optimization. Under the same experimental conditions, compared with the original YOLO v5, YOLO v8, Faster R-CNN, SSD and other models, the precision rate P, recall rate R, average recognition accuracy mAP0.5, mAP0.5:0.95 reached 97.40%, 94.20%, 97.20% and 79.10%, respectively, and the proposed algorithm maintained a high operation speed while improving the accuracy and average accuracy. It met the requirements of accuracy and timeliness of the detection of crop yellowing leaf curvature virus disease, and provided a theoretical basis for the intelligent identification of crop leaf diseases on mobile terminals.