Improved YOLOv5-Based Method for Date Palm Leaf Disease Recognition

The Significance of Deep Learning-Based Date Palm Leaf Disease Recognition for the Local Economy of Xinjiang, China” Cultivation of date palms is intricately tied to the daily needs of people, and it plays a vital role in the local economy of Xinjiang, China. Establishing a deep learning model capable of recognizing diseases in date palm leaves is of paramount importance. Such a model can enhance the quality of dates and reduce economic losses for farmers. Addressing challenges like disease spot occlusion, small disease targets that often lead to missed detections, and low recognition accuracy in complex background images, this study introduces an enhanced YOLOv5s-based model referred to as YOLOv5s-CBS. By incorporating the CBAM attention mechanism into the network's prediction head, it strengthens the feature extraction capability for small disease targets. This enhancement allows the model to better extract the features of the disease targets on date palm leaves, addressing the challenges of occluded and small targets and improving detection accuracy. The model employs the SIoU function to accelerate training and enhance inference accuracy. It also improves the feature fusion portion of the backbone network by using a bidirectional weighted feature pyramid network (BiFPN) to strengthen multi-scale feature fusion, ultimately improving recognition across different scales. To validate the effectiveness of YOLOv5s-CBS, it is compared with six other models: Faster R-CNN, YOLOv5m, YOLOv5l, YOLOv5x, YOLOv5s, and VGG16. In terms of various performance metrics, YOLOv5s-CBS performs the best in recognizing date palm leaf diseases, achieving an accuracy rate of 89.23%, a recall rate of 91.46%, and a mean average precision (mAP) of 78.86% at an overlap ratio of 0.5. Particularly high recognition accuracy is observed for date white spot disease and date palm mite disease, reaching 97.9% and 94.7%, respectively. Recognition performance for date rust dis-ease and date palm leaf spot disease is relatively lower, at 67% and 55.5%, respectively. Experimental results demonstrate that this model effectively addresses challenges such as oc-clusion and small disease targets in complex backgrounds, ultimately improving the accuracy of date palm leaf disease recognition.