Identification and Location Method of Cherry Tomato Picking Point Based on Si-YOLO

Abstract

In order to address the challenge of simultaneously recognizing and calculating the picking points of tomatoes in an unstructured environment, a method of identifying and locating cherry tomato picking points in a greenhouse environment was proposed. Thus, to improve the recognition of small targets, the attention mechanism was combined with the target detection algorithm, Si-YOLO, which has been proven effective in exploring the mechanisms and methods for enhancing small targets recognition. Therefore, Si-YOLO was used to solve the complex problem of tomato picking recognition. Firstly, the SimAM attention module was added to the backbone network of YOLOv5 to give more attention to the target fruit and to solve the problem of fruit stem recognition in similar color backgrounds. Second, a dataset was augmented by GAN combined with traditional image data augmentation methods such as mosaic, rotating 90 degrees, and Hue, which helps to locate the picking point of tomato string more accurately and to improve the generalization ability of the model. Si-YOLO achieves higher detection accuracy than classic models like Mask R-CNN, YOLOv5, Deeplabv3+, etc. Finally, the Si-YOLO model was deployed to Android system of mobile phones, and the stability of the model running on mobile terminal devices was verified by testing different cell phones. Experimental results showed that the Si-YOLO model's average accuracy, recall, and precision were 95.3%, 80%, and 44.6%. This study provides technical support for mobile edge computing-based robot target detection and visual location for harvesting operations in a facility environment.