2025年02期 v.46 1-8页
王伊杰1 蔡建扑1 任志刚1,2
（1.广东工业大学自动化学院，广东 广州 510006
2.粤港澳离散制造智能化联合实验室，广东 广州 510006）
摘要：在工业制造领域，对加工后的产品进行表面缺陷检测能够有效提高产品出厂质量。针对工业产品表面缺陷通常具有对比度低、形状不规则、尺寸小且细长等特征，并伴随有明显噪声，导致检测任务充满挑战的问题，提出一种基于高分辨率特征图的工业产品表面缺陷检测算法。首先，提出一种高分辨率特征图实时网络（RHNet）模型，通过将每个阶段的输入输出特征图保持为原始图像分辨率的1/4，有效保留了更多的细节信息；然后，提出短期双分支模块（SDBM），实时处理高分辨率特征图；最后，设计一种快速并行聚合金字塔池化模块（FPAPPM），快速提取深层信息并进行多尺度上下文融合。实验结果表明，RHNet模型在表面缺陷建模能力和检测性能方面均表现较好，能够满足工业场景实时性与应用部署的要求。
关键词：表面缺陷检测；高分辨率特征图实时网络；多尺度融合
中图分类号：TP391.41 文献标志码：A 文章编号：1674-2605(2025)02-0001-08
DOI：10.12475/aie.20250201 开放获取
Surface Defect Detection Algorithm for Industrial Products Based on High-resolution Feature Maps
WANG Yijie1 CAI Jianpu1 REN Zhigang1,2
(1.School of Automation Guangdong University of Technology, Guangzhou 510006, China
2.Guangdong-HongKong-Macao Joint Laboratory for Smart Discrete Manufacturing Guangdong University of Technology, Guangzhou 510006, China)
Abstract: In the field of industrial manufacturing, surface defect detection of processed products can effectively improve the quality of finished products. Industrial product surface defects often exhibit characteristics such as low contrast, irregular shapes, small and slender sizes, and significant noise, making the detection task highly challenging. To address this issue, a surface defect detection algorithm for industrial products based on high-resolution feature maps is proposed. First, a real-time high-resolution network (RHNet) is introduced, which maintains the input and output feature maps of each stage at 1/4 of the original image resolution, effectively preserving more detailed information. Then, a short-term dual-branch module (SDBM) is proposed to process high-resolution feature maps in real time. Finally, a fast parallel aggregation pyramid pooling module (FPAPPM) is designed to rapidly extract deep level information and perform multi-scale context fusion. Experimental results demonstrate that the RHNet model performs well in both surface defect modeling capability and detection performance, meeting the real-time requirements and deployment needs of industrial scenarios.
Keywords: surface defect detection; real-time network based on high-resolution feature maps; multi-scale fusion
