Research On Young Apple Image Recognition Method Based On YOLO Algorithm

With the agricultural market reform and the rapid development of China’s apple industry,apple bagging as a cultivation measure to produce pollution-free fruit can prevent pests and diseases,reduce pesticide residues and improve apple quality.Automatic bagging technology can reduce labor intensity,reduce danger and improve economic efficiency,which will become a major trend of apple bagging development.Research on visual detection system is a precondition for the development of automatic bagging technology.Accurate access to the location information of young apples is a prerequisite for automatic bagging.The image recognition technology of young apples is recognized by extracting the characteristic information of young apples,but the recognition is challenging due to the complexity of background environment,the overlapping of shading between branches and leaves and fruits,the variability of light conditions and the interference of other factors.Therefore,we conducted a study on the accurate and rapid recognition of young apples in natural environments containing multiple complex interference factors,as follows:(1)Using young apples as the research object,images of Morris and Red Fuji young apples were collected in the field,image annotation was performed,and a dataset was constructed.The four algorithms,YOLOv3,YOLOv4,YOLOv5 and YOLOx,were compared and analyzed in terms of structure,function and innovation mechanism,and the model training and recognition experiments of the four algorithm networks were completed for the Morris apple,Red Fuji apple and mixed apple image datasets,respectively,and the recognition results were analyzed and evaluated.(2)To address the problem that the recognition results are affected by the existence of multiple complex interference factors in the natural environment,we analyzed and categorized the interference factors affecting recognition,designed a test set of young apple images containing multiple complex interference factors,completed the recognition tests of different algorithms of YOLO series on the test set containing multiple interference factors,and conducted a comparative analysis of the recognition results,and the recognition accuracy and recall rate of all models were reduced.The recognition accuracy and recall rate of all models were reduced.(3)To improve the accuracy of apple young fruit recognition in images containing multiple complex interference factors,the YOLOv3 Darknet-53 feature extraction network was improved and validated based on the idea of YOLOv3 algorithm using residual network and multiscale feature fusion to detect small targets.A feature map of size(104,104,128)was used as the output instead of the original feature map of size(13,13,1024),and the improved YOLOv3 young apple target detection model was proposed.By increasing the size of the output feature map of the feature extraction network and reducing the size of the perceptual field,the ability of the algorithm to capture young apples in images and the recognition accuracy were improved.(4)The improved YOLOv3 algorithm was used for training and recognition of young apple images with multiple complex interference factors,and the recognition results were compared with the original YOLOv3 and other five target detection models.Structurally,the improved YOLOv3 had a simple internal structure and faster recognition speed.In terms of performance,the recognition accuracy of the improved YOLOv3 on the test set of images containing multiple complex interference factors was 92.44%,which was 3.85%higher than the original YOLOv3.The recall rate was 88.54%,which was 11.99%higher than the original YOLOv3,and the recognition speed per image was 1.76 seconds.The performance improvement of improved YOLOv3 was mainly reflected in the ability to recognize the correct number of targets,and the ability to detect the correct number of targets and the image recognition speed on the test set of young apple images containing multiple complex interference factors was better than the current YOLOv7 algorithm.This paper provided theoretical and technical support for the YOLO series of target detection algorithms in young apple recognition and evaluation,and provided a research direction for the development of vision systems for automatic fruit bagging technology.