| China is the country with the largest apple production,and the proportion of apple production has been on the rise.Apple picking has significant labor-intensive characteristics,resulting in a high proportion of labor costs in the picking process.With the loss of a large number of rural labor force in recent years,the shortage of labor force is becoming more and more serious,so it is more and more urgent to realize the automatic picking of apple.As an efficient and brand-new production mode,apple picking robot is of great significance to improve farmers’ living standards,speed up agricultural transformation and build agricultural modernization.The fast and accurate identification of apples is the key to the success of robot fruit picking.Aiming at the problem of fast and accurate recognition of apples in the actual complex orchard environment,this paper studies from the aspects of improving the detection accuracy of the algorithm and lightening the model,and proposes an apple target detection method based on YOLOv4.The main work includes the following points:1.The selection of target detection algorithm.The detection principle and network characteristics of typical algorithms based on candidate region and frame regression in current deep learning are described,and the detection advantages of YOLO algorithm are explained,which lays a theoretical foundation for determining that YOLOv4 algorithm is the basic algorithm of apple detection in this paper.2.Improve the multi-scale structure and frame de-redundancy algorithm.In order to solve the problem that YOLOv4 has many missed detection of apple targets in complex environment,and the regression effect of apple target frame under the condition of fruit overlap and occlusion is poor,an improved YOLOv4 target detection algorithm is proposed.On the one hand,the original YOLOv4 multi-scale prediction structure is improved,by adding a 104 × 104 prediction layer,using 4-scale prediction to improve the accuracy of small target detection,and using K-means clustering algorithm to re-obtain the corresponding anchor frame.On the other hand,the nonmaximum suppression algorithm is improved,and Soft-NMS is proposed as the redundant box removal algorithm to improve the detection effect of overlapping and occluded fruits,and the corresponding loss function is designed for the apple detection task.Finally,the model is tested in the test set.The results show that the mean average precision of the improved YOLOv4 algorithm is 93.46%,which is 3.46% higher than that of the original algorithm,the detection speed reaches38 FPS,and the model generalization ability is stronger,which meets the real-time and accurate detection of apple in complex environment.3.Lightweight design of network model.Aiming at the problem that the YOLOv4 network model has deep layers and high resource occupancy,and cannot adapt to the general performance of hardware equipment,this paper proposes to use Efficientnet B0 lightweight network to replace CSPDarknet53 network as the backbone feature extraction network of YOLOv4.At the same time,using 4-scale prediction structure and Soft-NMS algorithm,a Light-YOLOv4 lightweight network is designed to adapt to low performance embedded devices.Apples in different states are tested under the test set,and the recognition effect of the Light-YOLOv4 detection model is verified.The detection speed reaches 53.1FPS,which increases by 32.7%.The mean average precision is90.47%,which is slightly higher than YOLOv4 by 0.14%.4.Algorithm transplantation and performance evaluation on embedded platform.Through the algorithm transplantation in the Jetson Nano development board to further verify the performance of the improved lightweight detection model in low-performance embedded devices,the results show that the lightweight Light-YOLOv4 can be applied to the Jetson Nano embedded platform,and the detection speed is 8FPS,which is higher than that of YOLOv4.At the same time,a simple version of Apple detection system is developed to realize the application of the algorithm. |
PDF Full Text Request