Recognition of maize seedling under weed disturbance using improved YOLOv5 algorithm

Abstract

Using UAV-based RGB images to recognize maize seedlings is of great significant for precise weed control, efficient water and fertilizer management. However, the presence of weeds with morphological resemblances at the maize seedling stage affects the recognition of maize seedlings. This research employs UAV RGB images and deep learning algorithms to achieve accurate recognition of maize seedlings under weed disturbance. Firstly, the adaptive anchor frame algorithm is employed to intelligently select optimal anchor frame sizes suited for the maize seedling from UAV images. This strategic selection minimizes time and computational demands associated with multiple anchor frame sampling. Subsequently, the Global Attention Mechanism (GAM) is introduced, bolstering feature extraction capabilities. A range of deep learning models, including YOLOv3 and YOLOv5, are applied for maize seedling recognition, culminating in the identification of an optimal model. To account for real-world scenarios, we investigate the influences of UAV flight height and weed disturbance on maize seedling recognition. The results indicate a multi-class Average Precision (mAP) of 94.5 % and 88.2 % for detecting maize seedlings at flight heights of 15 m and 30 m, respectively, with an average detection speed of 0.025 s per single image. This emphasizes the efficacy of the improved YOLOv5 deep learning model in recognizing maize seedlings under weed disturbance using UAV RGB images.