Abstract:
In order to improve the accuracy of intelligent sugarcane harvesting, reduce the high computing power requirement for the algorithm deployment, the recognition algorithm of YOLOv4-tiny based on MobileNet and the recognition algorithm of YOLOv4-tiny based on network slimming are proposed, which make use of the advantages of the lightweight target detection algorithm YOLOv4-tiny compared with YOLOv4 algorithm, such as more simplified network structure and higher reasoning speed. The accuracy and complexity of the two models are compared. When the average precision of YOLOv4-tiny based on network slimming algorithm was only 0.6% lower than that before slimming(94.7%), its model complexity was reduced to 1/3 of the original, that was, the FLOPs and Params after slimming were 1.1 G and 1 789 658. However, when the average precision of YOLOv4-tiny with MobileNet as the backbone decreased by 1.92%, its FLOPs and Params were 1.29 G and 2 600 068. The performance of the latter in average precision and model complexity was not as good as that of the YOLOv4-tiny model after slimming. The results showed that the YOLOv4-tiny sugarcane stem node recognition model based on network slimming algorithm could effectively reduce the complexity of the model, and its computation was friendly to embed devices and mobile devices. The research can provide a technical method for the development of the intelligent sugarcane harvesting machine.