基于多分支特征聚合和大核感知单元的钢材缺陷检测算法OA
A Steel Surface Defect Detection Algorithm Based on Multi-Branch Feature Fusion and Large-Kernel Perception Unit
针对因钢材缺陷尺度多样性、纹理结构复杂性导致的检测准确度低和漏检误检等问题,提出了一种基于多分支特征聚合和大核感知单元的钢材缺陷检测算法.首先,提出一种多分支特征聚合方法,利用三阶异构卷积融合策略全面捕获钢材缺陷多尺度特征信息;其次,为了增强算法对目标区域结构信息的建模能力,基于大卷积核结合空间扩张卷积的方法设计大核感知单元,以针对钢材缺陷区域上下文信息进行有效建模,从而实现对原始特征的加权增强;然后,构建结构感知增强模块,通过强化算法对多尺度结构特征与上下文信息的建模能力,提升算法对钢材关键缺陷区域的感知与聚焦效果;最后,利用 Inner_SIoU 损失函数精细化刻画边界框回归过程,提升算法对钢材缺陷目标的检测能力并降低漏检率.实验结果表明,在扩增的 NEU-DET 数据集上,改进算法的平均精度均值、平均精度、召回率分别达到 88.7%、87.8%、81.2%,较 YOLOv11n 算法分别提升 3.8、4.6、2.1 百分点,说明改进算法在检测精度方面实现了显著提升.
To address the issues of low detection accuracy,missed detections and false detections caused by the di-versity of steel defect scales and the complexity of texture structures,we propose a steel defect detection algorithm based on multi-branch feature aggregation and large kernel perception unit.First,a multi-branch feature aggrega-tion method is proposed,which utilizes a third-order heterogeneous convolution fusion strategy to comprehensively capture the multi-scale feature information of steel defects.Second,to enhance the algorithm's ability to model the structural information of target regions,a large kernel perception unit is designed based on a method combining large convolution kernels with spatial dilated convolution,which can effectively model the context information in steel defect regions,thereby achieving weighted enhancement of the original features.Subsequently,we construct a structure-aware enhancement module that strengthens the algorithm's ability to model multi-scale structural features and context information,thereby improving its perception and focus on key defect regions.Finally,the Inner_SIoU loss function is employed to refine the bounding box regression process,which enhances the algorithm's detection capability for steel defect targets and reduces the miss rate.Experimental results demonstrate that on the augmented NEU-DET dataset,the improved algorithm achieved mean average precision,average precision,and recall rates of 88.7%,87.8%,and 81.2%,respectively,representing improvements of 3.8,4.6,and 2.1 percentage points over the YOLOv11n model.This indicates that the improved algorithm has achieved a significant enhancement in detection accuracy.
雷亮;周华勇;徐山雯;刘学涵;赵锦;陈毅;李亚轩
重庆科技大学 计算机科学与工程学院,重庆 401331重庆科技大学 计算机科学与工程学院,重庆 401331重庆科技大学 计算机科学与工程学院,重庆 401331重庆科技大学 计算机科学与工程学院,重庆 401331重庆科技大学 计算机科学与工程学院,重庆 401331重庆城市职业管理学院 大数据与信息产业学院,重庆 401331重庆科技大学 计算机科学与工程学院,重庆 401331
信息技术与安全科学
钢材缺陷检测多分支特征聚合异构卷积融合大核感知单元结构感知增强
steel defect detectionmulti-branch feature aggregationheterogeneous convolution fusionlarge-ker-nel perception unitstructure-aware enhancement
《重庆科技大学学报(自然科学版)》 2026 (2)
81-93,13
重庆市教委科学技术研究项目"基于YOLO-Pose姿态特征检测融合时序的独居老人跌倒识别算法研究"(KJQN202303305)重庆市属本科高校与中国科学院所属院所合作项目"工业互联网内生安全关键技术研究与协同创新"(HZ2021015)重庆科技大学研究生创新计划项目"工业场景下的钢材表面缺陷检测方法研究"(YKJCX2520928)
评论