LFA-YOLO: Detection of Lychee Fruit Anthracnose Based on UAV Images and Deep Learning

Abstract

HighlightsMulti-path coordinated attention mechanism incorporating salient information is proposed.LeanNet, a backbone network based on depth-separable convolution, is proposed.Channel grouping multiscale fusion convolution is proposed and applied in C2f and detection head.Deployed LFA-YOLO in edge devices for real-time applications.Abstract. Anthracnose is a communicable agricultural disease with a high incidence and significant damage, posing as a major threat to litchi production. In this article, we propose a lightweight target detection method called LFA-YOLO to address the problem of leakage and misdetection in the detection of lychee fruit anthracnose due to the complexity of the agricultural environment. Our method incorporates a Multipath Coordinate Attention (MCA) mechanism into the YOLOv8 algorithm to mitigate detection errors caused by mutual occlusion between fruits. Additionally, in order to reduce the number of model parameters, we utilize the more efficient LeanNet to replace the backbone network of the baseline model. Moreover, we propose a more efficient channel grouping multi-scale fusion convolution (CGMF convolution) and apply it to the C2f module and decoupled detection head. Experimental results on the lychee fruit anthracnose (LFA) dataset demonstrate that the proposed LFA-YOLO model achieves a mAP50 of 88.61, with precision and recall rates of 90.18 and 82.77, respectively. Furthermore, the number of model parameters and floating-point operations (FLOPs) of the new method decreases to 61% and 48% of that of the baseline network, while the frames per second (FPS) improves to 90.3. Finally, we deploy the proposed LFA-YOLO on the NVIDIA Jetson Nano edge computing device to provide vision support for agricultural unmanned aerial vehicles (UAVs). Keywords: Attention mechanism, Lychee fruit anthracnose, Object detection, YOLOv8.