Tomato Stem Recognition Based on Yolox

In automated tomato harvesting, it is essential to distinguish between the main stem of the plant and the stem with the fruit, which was the aim of this research. We hypothesised that instance segmentation can recognise the main stem and the stem with the fruit in tomato images, even with unbalanced data, with a precision better than 0.7 and an inference speed better than 3 fps. The tomato variety Merlice was used. From the measurement, 399 images were extracted concerning three conditions. The training, testing, and validating sets were randomly divided in an 80 : 10 : 10 ratio, with 10% of the training set allocated as the validation set for cross-validation during model training. We propose to use the YOLO-v8 model for recognition. With image size 800 × 800 pixels and model YOLO-v8m-seg, precision was 0.736, recall 0.738, mAP50 0.712 and mAP50-95 0.301. Inference speed was 6.34 fps. When using the YOLO-v8l-seg model, precision achieved was 0.761, recall 0.673, mAP50 0.689 and mAP50-95 0.308. Interference speed was 3.92 fps. Using integrated GPU, the inference can reach the speeds of up to 10 fps. Experimental results demonstrate usable overall recognition performance for unbalanced samples. The findings of this work provide a technical foundation for developing tomato harvesting robots.