Using improved YOLO V5s to recognize tomatoes in a continuous working environment

In the continuous working environment of the picking robots, factors such as illumination change, camera hardware, the movement of the picking robots, and image background interference have a great impact on the tomato detection accuracy of the picking robots. At present, the research on continuous working environment has not effectively solved the problem that multiple environmental factors affect tomato recognition. To solve this problem, based on YOLO V5s, this paper proposes an improved YOLO V5s for tomato identification in the continuous working environment of picking robots. The improved YOLO V5s are improved from the following aspects. Firstly, image sample data in continuous work environment is enriched by a data enhancement method. In addition, CBAM attention mechanism module is added to YOLO V5s feature extraction network to improve the capability of tomato feature extraction in the continuous work environment. On this basis, in the prediction part of YOLO V5s network model, an improved Soft Non-Maximum Suppression (Soft-NMS) method was proposed to improve the recognition ability of the network model to cluster tomatoes in the continuous working environment. The test results show that the tomato detection precision and recall rate of the improved YOLO V5s model are 92.08% and 82.42% respectively, the mAP is 92.75%. Compared with the original model, the improved YOLO V5s recognition precision increased by 2.72% and mAP increased by 1.29%. It has strong robustness under different lighting conditions and different tomatoes blocking each other, and can be better used in the continuous working environment of the picking robots.