Research On Fast Apple Detection Method Based On Deep Learning

Apple detection is one of the important technologies of apple picking robot system.Because of the complex scenes and variety in orchard,it is difficult to detect apples.Therefore,it is of great significance to achieve fast and accurate apple detection to promote the development of apple industry.Aiming at the complex scenes of the orchard,this thesis starts with four common apple detection research objects,namely red apple,green apple,bagged apple and defective apple,and studies apple detection based on YOLOv4 series models.The main research contents of this thesis are as follows:(1)Aiming at the problem that the apple data collection is not comprehensive enough,and combining with the complex scenes of the orchard,this thesis proposes an illustration data augmentation method,which randomly inserts the collected apple leaf illustrations into the original apple image to simulate the apple leaf occlusion scene with the biggest interference factors of the orchard.At the same time,some traditional data augmentation methods are used to further enhance the data,so that the apple data set contains richer scenes,and the model can learn more comprehensive apple features,thus improving the detection effect of apples.Through experimental comparison,the illustration data augmentation method proposed in this thesis is beneficial to the improvement of model performance and can accurately detect apples in complex scenes.(2)Due to the huge YOLOv4 model and low processing efficiency,this thesis proposes to replace CSPDarknet53,the backbone network with the largest proportion in YOLOv4,with a lightweight Efficient Net network,and superimpose a custom convolution block Conv2D on the output three features to further extract and match the features.Through experimental comparison,the overall performance of the improved Efficient Net-B0-YOLOv4 model in this thesis is better than the current mainstream apple detection model.In addition,in view of the diversity of apple species and the singleness of current apple detection research,this thesis compares four common apple detection studies,namely,red apple,green apple,bagged apple and defective apple,based on the improved model.The experimental results show that the improved Efficient Net-B0-YOLOv4 model can accurately detect apples in different scenes,which indicates that the model has good generalization performance.(3)Combining the advantages of Efficient Net network,the residual module in YOLOv4-Tiny model is optimized,and the attention mechanism SE module and the adaptive Conv module are added to further make full use of the extracted feature information.The improved YOLOv4-Tiny-SE~*model is compared with the improved Efficient Net-B0-YOLOv4 model.The experimental results show that it is superior to the improved Efficient Net-B0-YOLOv4 model in terms of model size and processing efficiency,and can meet the requirements of apple detection in complex scenes.(4)The online apple detection system based on C/S mode is designed and implemented,and a multifunctional man-machine interface is designed for the client users by using tools such as PyQt5,Qt Designer and Python.By inputting data and configuring it to the server,the model deployed by the server is used for calculation and processing,and finally the detection results are returned to the client,thus simulating and realizing the rapid apple detection operation in complex scenes.