Research On Maturity Discrimination Of Waxbarry Under Complex Orchard Environment

Waxberry shows the growth characteristics of short fruit maturity and significant differences in individual maturity time.During the harvest period,waxberry are mainly picked manually in batches.This repeated picking mode leads to high harvest cost and labor intensity.Therefore,automatic picking robot has become a research hotspot in Intelligent Agriculture in recent years.It uses machine vision technology to identify and locate mature fruits,and simulates manual positioning and fruit picking operation through end effector.However,the fruit recognition and positioning accuracy of picking robot is easily affected by unstructured environment and the growth morphological characteristics of waxberry under complex orchard environment.Therefore,this paper studies the location and maturity discrimination method of waxberry under the conditions of variable illumination and cluster growth based on monocular machine vision technology,combined with traditional image analysis and deep learning strategies.The main research contents and conclusions of this paper are as follows:(1)A method for judging the maturity of waxberry by combining the local sliding window technique and the morphological characteristics of fruit appearance.In order to maintain the actual color information of the waxberry area in the optical image,in the HSV(Hue/Saturation/Value)color space,contrast limited adaptive histogram equalization(CLAHE)is used to deal with the hue component,which alleviates the problem of unstable image of the fruit area due to the change of orchard ambient light.In YCb Cr color space,Cb Cr color difference method and Otsu threshold segmentation algorithm are used to filter most of the background regions.In order to solve the problem of fruit region over segmentation in Cb Cr color difference map,the candidate waxberry region is extracted combined with regional growth strategy;Further,the local sliding window method is used to traverse the candidate red waxberry region,and the classification model of red waxberry target is established by combining local binary pattern(LBP)and support vector machine(SVM).The classification results of adjacent regions are fused by soft non maximum suppression(soft NMS)to output more accurate fruit detection results;Finally,based on the prior knowledge of the appearance color of mature waxberry,a fruit maturity recognition algorithm combining color evolution model and red ratio is proposed to realize the maturity discrimination of waxberry.The experimental results show that compared with the three algorithms of global sliding window method,marker controlled watershed convex hull method and super-pixel clustering,this method has achieved the best performance in the detection and maturity discrimination of waxberry in different illumination.Its accuracy,recall and F1-score are 92.51%,90.82% and 91.65% respectively,and the maturity discrimination accuracy of red waxberry fruit is 90.85%.In addition,the local sliding window strategy is conducive to suppress a large number of candidate image regions containing only background,and significantly reduce the computational complexity of fruit detection.The average detection rate of this method is 173 ms/frame.From the perspective of traditional target detection methods,this method shows good near real-time detection performance.(2)Waxberry maturity discrimination based on deep convolution neural network(DCNN).According to the demand of DCNN for the quantity and diversity of training data,the methods of defocus blur,motion blur,rotation transform,random clipping,random flip,brightness and contrast adjustment are used to expand the training data.Further,the performance of waxberry detection and maturity discrimination of DCNN with different architectures of YOLOv5 s,Center Net,SSD and Faster RCNN were evaluated on the expanded training data.The results of experiment show that Faster RCNN obtains the best performance of waxberry detection and maturity discrimination in terms of overall average performance,its F1-socre is 95.72% and 92.62% respectively.But the average detection rate is the lowest among the four network models,which is 117ms/frame;YOLOv5s obtained the fastest detection and maturity discrimination rate of waxberry,reaching 30 ms/frame,its F1-socre was 95.52% and 90.61% respectively.The overall accuracy was slightly lower than that of Faster RCNN.At the same time,the total parameters and GPU operation space complexity of the network model were lower than those of the other three network models.In terms of model generalization performance,Faster RCNN and YOLOv5 s are least affected by the changes of orchard lighting conditions and the random growth morphological characteristics of waxberry.They can more accurately detect and distinguish red waxberry in blocked and clustered forms.Compared with Center Net and SSD models,the highest detection accuracy is improved by about 15%.