Study On The Method Of Millet Ear Detection In Field Based On YOLO Algorithm

The traditional method of ear detection is manual observation.This method is inefficient,labor-intensive,and subjective.It is not suitable for field ear statistics;while ear detection based on image processing technology,although it can identify ears by extracting the characteristics of ears,it is susceptible to the impact of field problems such as light and soil,and cannot get a generalized model.Feature extraction requires a lot of experience.In response to the above problems,this paper proposes to use the YOLO series model to quickly and efficiently detect the ears of rice in a specific frame.While improving the efficiency of counting ears,it also reduces the workload of researchers in the field.It satisfies the requirements for counting ears and is estimated to have a certain scientific value for the yield of millet per mu.The main work of this research is as follows:(1)In this study,the male sterile line GBS grown in the millet experimental site of Shanxi Agricultural University was used as the experimental object,and 784 millet ear images were collected to construct a millet ear detection data set.(2)This article uses three models of YOLOv2,YOLOv3 and YOLOv4 to detect millet ears.I have verified the advantages of YOLOv4;for these three models,I fixed confidence scores and select different IOU thresholds to study the detection performance of models.Experiments show that the performance of the model shows a downward trend with the increase of IOU.(3)In this paper,the two models of YOLOv3 and YOLOv4 study the detection performance of millet ears from the two aspects of batch data samples and anchors.For different batch data samples,experiments show that the batch data size will affect the performance of the model,and the algorithm performance will increase with the decrease of the batch samples;In this paper,k-means clustering is used to modify the anchor boxes of these two networks according to the grain data set,and the performance of the modified model is better than that of the original model.(4)At the end of this article,YOLOv5s,YOLOv5m,YOLOv5l and YOLOv5x are used to detect millet ears respectively.In the case of not fitting,as the depth of the model increases,the detection performance of the model gets better and better;Secondly,the impact of different IOUs on the detection performance of these four models is analyzed.As the IOU threshold increases from 0.2 to 0.5,the detection performance of the model is improving.Experimental results show that the YOLO-based neural network can effectively detect the ears of millet in a specific frame,which is an efficient and accurate method.