SwinBN:一种基于Swin Transformer的针织物疵点检测模型  

Swin BN:A Swin Transformer-based model for knitted fabric defect detection

作　　者：胡越杰 蒋高明[2] HU Yuejie;JIANG Gaoming(College of Textile Science and Engineering,Ministry of Education,Jiangnan University,Wuxi 214122,China;Engineering Research Center for Knitting Technology,Ministry of Education,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学纺织科学与工程学院,江苏无锡214122 [2]江南大学针织技术教育部工程研究中心,江苏无锡214122

出　　处：《丝绸》2023年第1期59-69,共11页Journal of Silk

基　　金：国家自然科学基金项目(61772238);泰山产业领军人才项目(tscy20180224)。

摘　　要：随着针织工业的发展,针织产品疵点的检测与分类成为一个具有广泛应用价值的研究领域。卷积神经网络受限于卷积运算的局部性,无法高效地关注全局特征。基于Transformer模型的研究越来越多,取得了良好的效果,但是仍然存在小目标识别能力差和局部特征提取能力不足等缺陷。为了解决这些问题,文章整合Transformer和CNN的优势对Swin Transformer进行优化,设计了DCSW(Deformable convolution and swin transformer)骨干网络以加强模型的局部感知能力,提高小目标疵点检测的准确率。除此之外,还构造了改进的BiFPN多尺度特征融合网络,有助于增强模型的定位精度。最终结合骨干网络和特征融合框架的多尺度自适应模型SwinBN,在自制的针织物疵点图像数据集上评估,其精确率、召回率和mAP值分别达到72.32%、78.87%和71.07%。实验结果表明,该模型优于现有最佳的目标检测方法,为针织物产品质量控制提供了一种新的解决方案。The knitting industry,which is a traditional advantageous industry in China,is of vital significance to the development of the private economy.Since the 21stcentury,people’s demand for knitted products and clothing has been increasing.In order to improve product quality and maintain market competitiveness,it is particularly critical to realize the detection and classification of knitted fabric defects.In the traditional weaving process,fabric defects are detected manually,which is time-consuming and labor-intensive.Moreover,due to the high intensity work for long time,the inspection personnel are very likely to be fatigued,resulting in low efficiency.Therefore,the automatic fabric defect detection and classification has become a research hotspot,while knitted fabrics are less involved than other fabrics.This is mainly due to the great elasticity of knitted fabrics and their relatively loose loop structure,which easily leads to the inability of the fabric to remain dimensionally stable after being unloaded,thus making the defect information relatively vague.The fabric defect detection method based on machine learning requires artificial design of feature extractors,which have weak generalization performance.Therefore,researchers prefer to adopt deep learning technology to extract fabric image features adaptively and intelligently.As a representative model of deep learning,convolutional neural network(CNN)has shown very competitive performance and achieved excellent results in fabric defect detection and classification tasks.However,because of the locality of convolutional operation,it cannot capture global features well.The self-attention mechanism of transformer can tackle this trouble.In recent years,the transformer-based models have been applied to computer vision and obtained remarkable success,but they still have shortcomings such as poor small target recognition ability and insufficient local feature extraction ability.To make up for these deficiencies,we integrate the advantages of transformer and CNN to

关 键 词：针织物 疵点检测 可变形卷积 图像处理 自注意力 Swin Transformer 计算机视觉 

分 类 号：TS181.9[轻工技术与工程—纺织材料与纺织品设计]