结合注意力的纺织品瑕疵检测方法研究

PDF下载HTML

Title:

Research on textile defect detection method combining attention mechanism

作者:

周在雍¹, 狄岚²

1. 江南大学 人工智能与计算机学院, 江苏 无锡 214000;
2. 江南大学 数字媒体学院, 江苏 无锡 214122

Author(s):

ZHOU Zaiyong¹, DI Lan²

1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214000, China;
2. School of Digital Media, Jiangnan University, Wuxi 214122, China

关键词:

注意力机制; 尺度聚合; 双线性插值; 离散余弦变换; 多尺度特征; 特征融合; 纺织品瑕疵检测; 计算机视觉

Keywords:

attention mechanism; scale aggregation; bilinear interpolation; discrete cosine transform; multi-scale feature; feature fusion; textile defect detection; computer vision

分类号:

TP391

DOI:

10.11992/tis.202304045

摘要:

本文阐述了一种名为SAAM-YOLOX的基于改进YOLOX的纺织品瑕疵检测模型，旨在解决纺织品瑕疵检测中针对犬牙花纹与格纹背景中出现的误检和漏检问题，以及整体检测精度不高的问题。在特征提取阶段，该模型引入了离散余弦变换所构建的多分支离散余弦注意力机制（multi-branch discrete cosine attention，MDCA），能够解决模型在犬牙花纹与格纹背景中出现的误检和漏检问题，并且在检测精度上有一定的提高；在特征融合阶段，为了聚集和加强不同尺度的语义特征，SAAM-YOLOX模型采用了尺度聚合技术和注意力机制来构建尺度聚合注意力模块（scale aggregation attention module，SAAM）。在SAAM的上采样过程中，使用双线性插值结合自注意力机制来增强特征信息的有效性，从而进一步提高检测的精度。在完成尺度聚合后，加入注意力模块来增强混合尺度的特征表示，最终实现提高检测效果的目的。实验结果表明，本文检测模型解决了犬牙花纹与格纹背景中出现的误检和漏检问题，并且提高了瑕疵检测的精度。

Abstract:

This paper presents a textile defect detection model called SAAM-YOLOX, which is based on the improved YOLOX. The model aims to address the issues of false positives and false negatives in textile defect detection, particularly in detecting Houndstooth and Gingham backgrounds, as well as the problem of overall low detection accuracy. In the feature extraction stage, the model introduces a multi-branch discrete cosine attention mechanism (MDCA) based on the discrete cosine transform to address false positives and false negatives in the Houndstooth and Gingham backgrounds, and thereby improve the detection accuracy. In the feature fusion stage, the SAAM-YOLOX model adopts scale aggregation technology and attention mechanism to construct a scale aggregation attention module (SAAM) to aggregate and enhance semantic features of different scales. SAAM uses bilinear interpolation combined with self-attention mechanism to enhance validity of feature information in the upsampling process, further improving the detection accuracy. After completing the scale aggregation, an attention module is added to enhance the mixed-scale feature representation, ultimately achieving the goal of improving detection performance. Experimental results demonstrate that the proposed detection model can effectively solve the problem of false positives and false negatives in Houndstooth and Gingham backgrounds, and improve the accuracy of defect detection.