Maize Seedling and Weed Detection based on MobileNetv3-YOLOv4
Cheng Liu, Shi-Quan Shao, Gao Wei
Abstract
Aiming at the problems of low identification accuracy, poor real-time performance and robustness caused by the cross growth of crops and weeds, an improved YOLOv4 detector model based on lightweight convolutional neural network(CNN) was proposed with maize and its associated weeds in seedling stage as the research object. First, in order to reduce the number of parameters and improve the speed of feature extraction, MobileNetv3 was used to build a lightweight feature extraction backbone network to replace CSPDarkNet53 network in YOLOv4. Second, use transfer learning strategy to accelerate the training speed. The experimental results show that the mean average precision(mAP) is 89.98%, the detection speed is 69.76 f/s and the number of parameters is 8.17x106 of our model proposed for corn and its associated weeds in natural environment. The mAP is decreased by 1.85%, but the detection speed is increased by 180%, and the number of parameters is decreased by 9 times compared with YOLOv4. And mAP is increased by 2.70%, detection speed is decreased by 8%, and the number of parameters is increased by 30% compared with YOLOv4_Tiny. The overall results indicate that out model can give consideration to both speed and accuracy in the maize seeding and its associated weeds detection, which can provide technical support for precision weeding in agricultural automation.