Evaluating YOLOv5 Deep Learning Models for the Detection of Tomato Berries for Crop Yield Estimation

摘要

A major advantage of harvesting robots is automatic fruit detection. Fruit recognition is difficult due to complex environmental variables such as lighting change, branch and leaf occlusion, and tomato overlap. Based on YOLOv5, an enhanced tomato detection model dubbed Tomato-YOLO is provided in this study to address these issues. YOLOv5 has a dense architecture, which makes it easier to reuse features and develop a more concise and accurate model. Furthermore, for tomato localisation, the model uses a rectangle bounding box. The bounding boxes can then more accurately match the tomatoes, improving the Non-Maximum Suppression Intersection-over-Union (IoU) calculation (NMS). They also reduce the size of the prediction coordinates. This will afford for more advancements in edge deep learning models for in situ and real-time visual tomato detection, which is necessary for harvesting robot development. The effectiveness of these alterations was demonstrated in an excision research. The research demonstrated that the system can distinguish green and reddish tomatoes, even when they are shrouded by leaves. With the NVIDIA GEFORCE GTX Architecture platform, Tomato-YOLO had the best performance, with an F1-score of 66.15 percent, a mAP of 52.26 percent, and an inference time of 16.14 ms.