Study on fusion clustering and improved YOLOv5 algorithm based on multiple occlusion of Camellia oleifera fruit

The detection method of Camellia oleifera fruits is not only an essential part of the Camellia oleifera fruit detection and locating process, but also a key foundation of the vision robot harvesting. Due to the occlusions of branches and leaves, as well as the overlap of fruits, it is a daunting task to improve the precision of Camellia oleifera fruit detection in a natural environment. In this paper, we propose a YOLO-COF lightweight multiclass occlusion target detection method for Camellia oleifera fruit, introducing K-means++ clustering algorithm under the framework of YOLOv5s to automatically filter and select the target dataset; incorporating Coordinate Attention to enhance the feature extraction ability of occluded target and improve the generalizability and robustness of the model. The study assesses four categories of datasets: complete fruit occlusion NO, fruit occluded by leaf OL, fruit occluded by branches OB, and fruit occluded by other fruits OF, combining the new detection model and comparing with five Faster R-CNN, YOLOv4-MobileNetV3, YOLOv4-CSPDarknet, YOLOv5s, and YOLOv5x models. The experimental results indicate that the AP values of YOLO-COF for the detection of NO, OL, OB, and OF are 0.97, 0.94, 0.93, and 0.92, respectively. In comparison to other neural network models, YOLO-COF is well-balanced in terms of precision, speed, and size. Its mAP value is 94.10%, its frame rate is 74.8FPS, and its model size is 27.1 MB. The proposed model has the potential to facilitate the operational decision-making of vision robot during the process of Camellia oleifera fruit harvesting.

Automated summary

This study proposes a YOLO-COF lightweight multiclass occlusion target detection method based on the YOLOv5s framework, which automatically filters and selects target datasets using the K-means++ clustering algorithm. It incorporates Coordinate Attention to enhance the feature extraction ability of occluded targets and improve the generalizability and robustness of the model. Experimental results show that YOLO-COF achieves AP values of 0.97, 0.94, 0.93, and 0.92 for the detection of NO, OL, OB, and OF, respectively. Compared to other neural network models, YOLO-COF achieves a good balance between precision, speed, and size. Its mAP value is 94.10%, frame rate is 74.8FPS, and model size is 27.1 MB. The proposed model has the potential to facilitate the operational decision-making of vision robot during the process of Camellia oleifera fruit harvesting.