UAV remote sensing detection of tea leaf blight based on DDMA-YOLO

Abstract

Tea leaf blight (TLB) is a common disease that affects the yield and quality of tea. Timely and accurate detection and monitoring of TLB can help support the precise control of the disease. This study proposed an unmanned aerial vehicle (UAV) remote sensing method based on DDMA-YOLO for effectively detecting and monitoring TLB while reducing the workload and time consumption of this process. This method used the RCAN to reconstruct high-resolution tea images to solve the problem of insufficient resolution of UAV remote sensing images. In this method, Retinex was selected to enhance the image contrast and to reduce the influence of uneven illumination. The amount of training sample data was expanded to improve the model’s generalization performance. The DDMA-YOLO model was constructed to improve the accuracy of monitoring TLB. The DDMA-YOLO model was developed using the YOLOv5 network as the baseline and by adding a multiscale RFB module to the backbone to improve the extraction ability of the detailed features of diseased leaves and to reduce the problem of missed detection caused by small leaves. A dual-dimensional mixed attention (DDMA) was added to the Neck, which parallels coordinate attention with channel attention and spatial attention, integrates nonlocal attention information and local attention information, and reduces missed detection and false detection caused by dense blade distribution. The experimental results show that the proposed method was superior to the classic target detection methods Fast R-CNN, SSD, RetinaNet, YOLOv3, YOLOv4 and YOLOv5. Compared with the baseline network, the AP@0.5 of the proposed method increased by 3.8%, and the recall increased by 6.5%.