Authors

Abstract

Egg appearance defect detection in conveyor-belt environments is challenged by irregular defect morphology, illumination variations, and the limited computational resources of intelligent sorting equipment, making it difficult to simultaneously achieve high detection accuracy and lightweight deployment. In this study, a lightweight detection model, termed YOLOv11n-DAL, is proposed to improve egg defect detection performance while reducing computational complexity. The proposed model integrates C3k2_DCNv4, ADown, and Detect_LSDECD modules into the YOLOv11n framework to enhance defect feature representation, improve effective information retention during downsampling, and reduce computational redundancy in the prediction stage. A conveyor-belt egg image dataset containing three categories, namely damaged, dirty, and good, was constructed. The dataset consisted of 1,730 original images, and the training set was expanded to 4,152 images through data augmentation. Comparative experiments and ablation studies were conducted to validate the effectiveness of the proposed model. The experimental results demonstrated that YOLOv11n-DAL achieved 95.2% Precision, 94.5% Recall, 94.8% F1-score, 96.7% mAP@0.5, and 95.8% mAP@0.5:0.95. Compared with the original YOLOv11n model, YOLOv11n-DAL reduced the number of parameters from 2.6 M to 1.7 M, GFLOPs from 6.3 to 4.8, and model size from 5.2 MB to 3.8 MB, while maintaining superior detection performance. The results indicate that YOLOv11n-DAL achieves an effective balance between detection accuracy and computational efficiency, providing a lightweight and practical solution for intelligent egg sorting systems in conveyor-belt environments.

Abstract in Chinese

传送带环境下鸡蛋外观缺陷检测受到缺陷形态不规则、光照变化以及智能分拣设备计算资源有限等因素影响，难以同时满足检测精度和轻量化部署需求。本文提出一种轻量化检测模型YOLOv11n-DAL，在降低计算复杂度的同时提升鸡蛋缺陷检测性能。该模型将C3k2_DCNv4、ADown和Detect_LSDECD融合至YOLOv11n框架中，以增强缺陷特征表达能力、提高下采样过程中的有效信息保留能力，并降低预测阶段的计算冗余。本文构建了包含damaged、dirty和good三类的传送带鸡蛋图像数据集，包括1730张原始图像，训练集经数据增强后扩展至4152张。通过对比实验和消融实验对模型性能进行验证。实验结果表明，YOLOv11n-DAL模型的Precision、Recall、F1-score、mAP@0.5和mAP@0.5:0.95分别达到95.2%、94.5%、94.8%、96.7%和95.8%。相比YOLOv11n，所提出模型将参数量由2.6 M降低至1.7 M，GFLOPs由6.3降低至4.8，模型大小由5.2 MB降低至3.8 MB，同时保持更高检测性能。研究结果表明，YOLOv11n-DAL能够在检测精度和计算效率之间取得良好平衡，为传送带鸡蛋智能分拣系统提供了一种轻量化解决方案