Design of citrus peel defect and fruit morphology detection method based on machine vision.

• The purpose of this paper is to achieve the appearance quality inspection of citrus, addressing the challenges of defect target and fruit morphology detection. • A detection algorithm for defect targets on citrus peel and a method for fruit morphology detection have been proposed. • A citrus quality detection data set is constructed and evaluated. Experimental results show that the method in this paper achieves satisfactory results in performance. Identifying defects in citrus peels and analyzing fruit morphology are two core challenges in citrus quality inspection. In order to more accurately identify minor defects on citrus peels, we proposed a detection model Yolo-FD (Yolo for defects). The model was based on the Yolov5 network framework, and the backbone network embedded the Three-dimensional Coordinate Attention (TDCA) mechanism innovatively designed in this study. It accurately captured the subtle changes and feature associations of the target in spatial location, significantly enhancing the model's ability to perceive defects in fruit peels. Moreover, we employed a simplified Bidirectional Weighted Feature Pyramid Network (BiFPN) in the model to achieve cross-scale connections and improve the feature fusion ability of the model. At the same time, Contextual Transformer block (COT) was introduced into Neck network and the CoT3 module was built to fully capture the static and dynamic contextual information in the citrus defects images and enhance the expression of the feature map. Through this series of improvement methods, missed detections and false detections caused by small targets were effectively reduced. Fruit morphology detection was combined with the Partice Swarm Optimized Extreme Learning Machine (PSO-ELM) model to determine whether the citrus fruit morphology was well-formed, using the symmetry index, roundness and tilt angle of the citrus as input parameters. The experimental results indicated that the mean average precision of the Yolo-FD model is 98.7 % (mAP-0.5). Compared with Yolov5s, Yolov7-tiny, and Yolov8n, the mAP was improved by 1.4 %, 1.5 %, and 0.5 % respectively. Its average detection time for a single frame image on the server was 19.5 ms. And the PSO-ELM model achieved a fruit morphology detection accuracy of 91.42 %, a coefficient of determination of 0.9044, and a mean squared error of 0.8497. The research results met the accuracy and real-time requirements for citrus sorting on the production line, and could provide an effective solution for citrus grading and quality assessment. [ABSTRACT FROM AUTHOR]