Research On Strawberry Image Recognition Algorithm Based On Deep Learnin

In recent years,with the increase of strawberry orchard cultivation area and the loss of rural labor force,the mechanization of strawberry picking research is necessary to put on the agenda.In view of the trend of mechanization picking,the research on efficient strawberry recognition becomes the first task.The overlapping of fruits,the blocking of stems and leaves,and the difference of individual ripeness are the great challenges for efficient and accurate mechanized strawberry picking.At present,most of the domestic picking robots still stay in the traditional machine vision method for the detection and recognition of complex and diversified strawberry picking.There are problems such as slow detection speed,low accuracy,missing detection and false detection,which seriously affect the efficiency of the production link.In recent years,the deep learning object detection algorithm based on GPU computers has greatly improved the detection speed and accuracy,but it has high requirements on hardware resources,large network parameters and slow forward inference speed,which cannot meet the real-time detection requirements of industrial production,and it is difficult to run on computers,embedded devices or mobile terminals with low computing resources.In order to improve the efficiency of automatic picking,this paper studied the fast recognition algorithm of ripe strawberry and the lightweight convolutional neural network with good precision and real-time performance,and transplanted the improved lightweight algorithm to Android terminal,which laid the foundation for the subsequent real-time detection of strawberry and intelligent management of orchard.Specific work and achievements are as follows:(1)Due to the lack of domestic strawberry data sets,this paper chooses homemade strawberry data sets to train and test the improved network.Firstly,the data set was preprocessed,then the strawberry images collected were annotated with Label Img,an open source software,and then the data enhancement method was used to enrich the number of samples.(2)According to the actual needs of target recognition and positioning,the advantages and disadvantages of the current mainstream target detection algorithms are compared,and the YOLOv4 algorithm with both detection speed and accuracy and good open source is selected as the basic algorithm,and then the algorithm model is optimized and compressed.(3)The improved YOLOv4 algorithm model was fused with lightweight neural network Mobilenetv3 to reduce the size of the model and increase the detection rate.After further improvement on Mobilenetv3-YOLOv4,a better model Mobilenet V3-yolov4-ST is obtained.Finally,the improved YOLOv4 algorithm network is trained and tested respectively.Experimental results show that the algorithm proposed in this paper has a good recognition effect on strawberry targets,and has good robustness,with an average recognition accuracy of 92.63%.The model size of the improved algorithm is only187 MB,which is 63 M less than that of the original YOLOv4 network model,and the detection rate reaches 44 frames per second.Compared with YOLOv4 network,the detection rate is improved by nearly 90%.The detection FPS value of the improved lightweight Mobilenet V3-Yolov4-ST model is 5 points higher than that of the Mobile Netv3-YOLOv4 model.