Torreya Sinensis Mature Fruit Detection and Morphological Center Point Labeling Based on YOLOV5

The small size and dense growth pattern of mature Torreya grandis fruits, compounded by different ripeness clustered on the same plant, it is a great challenge for the morphological center identification accurately. A detection method that integrates Otsu threshold segmentation with the YOLOv5 network was proposed in this paper. Firstly, a Yolov5 network was trained with different pictures, and the U channel of the YUV model was used for noise reduction followed the detected images. Then, the morphological center was calculated with the method of Otsu threshold segmentation from the images. Above all, the canny method is employed to detect the edges of mature fruits, allowing for the absolute coordinates extracted in the image. The absolute coordinates of the morphological center can be obtained through the normalization. It is demonstrated that the detection accuracy can reach at 97.5% based on the YOLOv5 network. The issue, caused by non-coincidence between the morphological center and the detection box center, can be solved thanks to the optimized detections followed by Otsu threshold segmentation algorithm. The proposed method can be a promising application for Torreya grandis fruit picking automatically.