Research On Detection And Localization Method Of Potato Seedling Images

Accurate recognition of potato leaf bud is the key technology to enhance the cutting seedling production efficiency in the potato automatic breeding system.Since potato leaf bud detection has the characteristics of small targets,severe occlusion of stems and leaf buds,and unstable environmental light,strict requirements are put forward for the accurate detection and positioning algorithm of leaf buds.To this end,the deep learning network based potato leaf bud identification and detection methods are studied in this paper.Firstly,in order to improve the accuracy of recognition,basic network design research is carried out.In order to reduce the rate of missed detection,the number of the anchor boxes in the basic network is set to 12.For improving the accuracy of small target recognition,four scaling scales are added according to the input size of the basic network.For the purpose of reducing the false recognition rate and improve the distinction between leaf bud and stem,color loss information item is added to the loss function of the network,and pre-training weight of Darknet is loaded with migration learning helps the network to converge.The identification experiment is carried out,and the recall rate of the newly designed basic network is 6.61%higher than that of the YOLO v3 network,the accuracy rate is 3.67%higher than that of YOLO v3 network,the F₁score is 5.19%higher than that of YOLO v3 network.At the same time,the newly designed basic network is also compared with SSD and Faster R-CNN,both the time and accuracy of identification are better than these two networks.Secondly,combining with the latest strategies of YOLO v4,based on the above-mentioned basic network,a deep learning network for potato leaf bud recognition is constructed.On the basis of 12 anchor boxes,for the input size,according to the multi-scale scaling of the basic network,the scale prediction layer of19×19,38×38,76×76,152×152 is generated,and the color is fused into the loss function of YOLO v4 and a new loss function is constructed,retaining the characteristics of YOLO v4,the recognition experiments is conducted.The recall rate of the new convolutional neural network is 0.51%higher than that of YOLO v4 and3.12%higher than that of the basic network.The accuracy rate is 0.82%higher than that of YOLO v4 and 0.81%higher than that of the basic network.The F₁score is0.65%higher than the YOLO v4 network and 1.98%higher than the basic network.And the average recognition speed of each image can reach 0.38s.Combining with the YOLO v4 strategy,the use of anchor boxes,multi-scale prediction,and loss function method of the basic network can further improve the accuracy of network recognition.Finally,through newly designed recognition network detection,the detection results are preprocessed,and the maximum entropy threshold is used to obtain a clear position of the stem.Combining with the test results,after matching,the cutting point of the stem is determined.By transforming coordinate,the position of the knife is finally obtained.Combining with the actual situation,the shear prototype can meet the current application requirements of the laboratory.