Research On Identification And Detection Method Of Orchard Peach Based On YOLOv3

There are many varieties of agricultural products in China,among which peaches are the fruit products that are in large demand in the daily life of residents and are widely planted in China.The current peach picking still relies mainly on labor,and the labor demand is huge.If there is a shortage of labor,it will delay the picking of mature peaches and cause huge economic losses.This article provides a solution for the society to deal with labor shortages.It uses automatic picking machines to pick ripe fruits,and uses object detection methods to detect peaches.Because the peach plants are densely distributed and the branches and leaves are luxuriant,the branches and fruits are interlaced to produce occlusion,which increases the difficulty of peach identification and detection.Therefore,this paper launched a research on the target detection of orchard peaches based on Yolov3.In order to further improve the accuracy of Yolov3 in peach recognition and detection,and enhance the detection effect.Based on the self-built "autumn peach" data set,this paper proposes an optimization method for orchard peach detection based on Yolov3,and completes the realization of the peach recognition and detection system.Optimization method: Use the ResNet-101 network to perform pre-training on the public data set COCO data set,optimize the initial network weights to improve the convergence speed and accuracy of the method,to adapt to the small amount of training data in engineering applications,and save data collection requirements Time cost.Using the ResNet-101 network as the backbone network for feature extraction is conducive to better extracting the deep information in the image.For the shallow feature information of small targets,multi-scale feature fusion is performed on convolution features of different multiples.The subsequent processing method is softer than the NMS in YOLOv3.The soft-NMS algorithm filters out the pre-selected areas of the image to improve the detection effect of adjacent objects.Through a series of up-sampling and down-sampling calculations,three different areas are obtained.The feature map of the size realizes the multi-scale detection of peaches in the orchard.The method proposed in this paper has an average correct detection rate of 96.3%on the test set of orchard autumn peaches,an average correct recognition rate of 95.1%,a mAP value of 0.82,and a single image sample running time of 69 ms.Experimental results show that for the data set of orchard peaches,the improved peach recognition and detection method based on YOLOv3 is better than SSD,Faster R-CNN,YOLO,YOLOv3 target detection methods,and can effectively identify the labels and positions of mature and immature peaches,single sample The processing time can meet the requirements of real-time control.