Abstract:

[Significance] Efficient and accurate quality detection is a key link in the post-harvest fruit industry. It is critical for ensuring food safety and improving industrial efficiency. With the growth of global population and the increasing demand for food safety, traditional manual inspection has become incompetent. Manual inspection relies heavily on the experience of operators. It is time-consuming, labor-intensive and subjective. The results are often unstable and inaccurate. It also may cause secondary damage to fruits. In addition, the rising cost of human resources makes the transformation of the fruit industry towards automation and intelligence an inevitable trend. Computer vision and machine learning technologies provide a solution. They can improve detection accuracy and efficiency, while reducing labor costs. Among various object detection algorithms, the YOLO (You Only Look Once) series stands out for its balance of real-time performance, high accuracy and low computational cost. It has achieved excellent results in fruit defect detection. However, there is still a lack of comprehensive reviews on the application of YOLO series algorithms in this field. The purpose of this paper is to fill the blank of YOLO series algorithms in the field of fruit defect detection and provide strong support for the further development of agricultural fruit defect detection technology. [Progress] Focusing on fruit defect detection, this paper reviews the development history of YOLO models from YOLOv1 to YOLO26. It also analyzes the innovative improvements of each generation. The development of YOLO models can be divided into three stages. The first stage (YOLOv1-YOLOv3) established the single-stage detection framework. YOLOv1 proposed the idea of single-shot detection, laying a preliminary foundation for real-time fruit defect detection. YOLOv2 introduced the anchor box mechanism and a more efficient backbone network, improving the positioning robustness for fruits and their defects of different sizes and shapes. YOLOv3 proposed a multi-scale feature pyramid, enhancing the detection ability for small defects such as early lesions and minor bruises. The second stage (YOLOv4-YOLOv8) focused on modular design, engineering construction and efficiency optimization. YOLOv4 clarified the modular design of backbone, neck and head, providing a clear framework for algorithm improvement. YOLOv5, YOLOv6 and YOLOv8 made continuous breakthroughs in lightweight design, anchor-free detection and decoupled head structure. YOLOv7 introduced the E-ELAN (Extended Efficient Layer Aggregation Network) architecture and auxiliary head, balancing detection accuracy and speed. The third stage (YOLOv9-YOLO26) focused on information extraction and transmission, post-processing simplification and high-level semantic modeling. YOLOv9 alleviated information attenuation in deep networks. YOLOv10 simplified post-processing to meet the high throughput needs of online sorting systems. YOLOv13 and YOLO26 further improved the model's performance in complex scenarios. This review also summarizes the application of YOLO series models in various fruit defect detection tasks, including apples, citrus, tomatoes and pears. It focuses on three key challenges: detection of internal and latent defects, defect detection under complex surface morphologies, and model lightweighting and edge deployment. For internal and latent defects, researchers combine YOLO models with multi-modal imaging technologies such as near-infrared and hyperspectral imaging. For complex surface morphologies, they improve the model's feature extraction ability by introducing attention mechanisms and multi-scale feature fusion. For lightweight and edge deployment, they adopt strategies such as lightweight architecture design, model compression and training optimization. [Conclusions and Prospects] The YOLO series models have become one of the most widely used technical frameworks in fruit defect detection. They have achieved remarkable phased results. However, they still face several challenges. These include algorithm limitations, such as the coupling optimization conflict between positioning and classification, and the limitation of local feature extraction. Dataset specificity leads to poor model generalization. Hardware constraints restrict the deployment of models on edge devices. Complex detection environments reduce the model's robustness. Future research directions include four aspects. First, build standardized and large-scale public fruit defect datasets, covering different fruit varieties and defect types. Second, improve the interpretability of YOLO models, building intuitive explanation interfaces for non-professional users. Third, develop human-machine collaboration strategies, combining the advantages of deep learning models and farmers' practical experience. Fourth, integrate YOLO with large models, optimizing data generation strategies and model lightweighting to achieve efficient deployment. This review aims to provide theoretical support and technical reference for the further development of agricultural fruit defect detection technology.

Key words: deep learning, object detection, YOLO, automation, fruit defect detection