邓超,赵永昆,孙俊岭,等.基于YOLO-RCES的轻量级钢材表面瑕疵检测算法[J].河南理工大学学报（自然科学版）,2026,45(6):126-135.

Deng C, Zhao Y K, Sun J L,et al.A lightweight steel surface defect detection algorithm based on YOLO-RCES[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(6):126-135.

基于YOLO-RCES的轻量级钢材表面瑕疵检测算法

邓超, 赵永昆, 孙俊岭, 王伟东

河南理工大学 物理与电子信息学院，河南 焦作 454003

摘要: 目的 为提高深度学习钢材表面瑕疵检测算法在边缘设备上的部署能力和检测效果，针对现有检测算法参数量大、计算复杂度高、检测精度较低等问题，提出一种基于YOLO-RCES的轻量级钢材表面瑕疵检测算法。  方法 针对模型复杂度高且不能有效利用底层纹理信息的问题，使用RevCol网络改造YOLOv8n主干，提高模型检测精度，降低模型复杂度。针对YOLOv8n的C2f模块中卷积堆叠致使模型复杂度较高的问题，向特征融合网络C2f模块引入高效多尺度卷积，降低网络的参数量和浮点计算量，丰富网络多尺度信息。针对模型提取特征能力不强的问题，向改进后的主干引入注意力机制，提高网络提取长距离依赖和上下文信息的能力。最后，针对YOLOv8n损失函数不能很好平衡难易样本的问题，引入SlideLoss强化模型，提高难检测瑕疵的检测效果。  结果 在GC10-DET数据集上测试，与YOLOv8n模型相比，改进模型的精确度提高2.3%，mAP@0.5提高1.6%，mAP @0.5∶0.95提高1.9%，参数量降低26.7%，浮点计算量降低24.4%。在NEU-DET数据集上进行测试，改进模型的mAP @0.5提高1.4%，进一步验证了模型的泛化能力。  结论 改进模型提高了轻量化程度和检测精度，与目前主流检测模型的轻量级版本和先进钢材检测轻量级模型对比，具有更好的综合优势。

关键词:钢材表面瑕疵检测;轻量化;可逆柱网络;通道先验卷积注意力;高效多尺度卷积

doi:10.16186/j.cnki.1673-9787.2024030069

基金项目:国家自然科学基金资助项目（62101176）；河南省科技攻关项目（232102210100，242102210082）；河南理工大学基本科研业务费基础研究项目（B类）（NSFRF230601，NSFRF240621）

收稿日期:2024/03/26

修回日期:2024/10/10

出版日期:2026/10/09

A lightweight steel surface defect detection algorithm based on YOLO-RCES

Deng Chao, Zhao Yongkun, Sun Junling, Wang Weidong

School of Physics & Electronic Information Engineering， Henan Polytechnic University， Jiaozuo 454003， Henan， China

Abstract: Objectives To improve the deployability and detection performance of deep learning-based steel surface defect detection algorithms on edge devices， a lightweight detection algorithm， YOLO-RCES， is proposed to address the issues of large parameter size， high computational complexity， and low detection accuracy in existing methods.  Methods To reduce model complexity and enhance the utilization of low-level texture information， the RevCol network is employed to reconstruct the backbone of YOLOv8n， achieving improved detection accuracy with reduced complexity. To address the high computational cost caused by convolution stacking in the C2f module， efficient multi-scale convolution is introduced into the feature fusion network， reducing the number of parameters and floating-point operations while enriching multi-scale feature representation. Furthermore， an attention mechanism is incorporated into the improved backbone to enhance the model’s ability to capture long-range dependencies and contextual information. To alleviate the imbalance between hard and easy samples， SlideLoss is introduced to improve the detection performance for hard defects. Results Experiments on the GC10-DET dataset show that， compared with YOLOv8n， the proposed model improves precision by 2.3%， mAP@0.5 by 1.6%， and mAP @0.5∶0.95 by 1.9%， while reducing parameters by 26.7% and computational cost by 24.4%. On the NEU-DET dataset， mAP @0.5 is further improved by 1.4%， demonstrating good generalization ability.  Conclusions The proposed model achieves higher detection accuracy while maintaining a lightweight structure. Compared with existing lightweight versions of mainstream detection models and advanced lightweight steel defect detection methods， it demonstrates superior overall performance.

Key words: steel surface defect detection; lightweight; RevCol network; channel prior convolutional attention; efficient multi-scale convolution