融合形变卷积和自注意力的素色布匹瑕疵检测

Plain Fabric Defect Detection Using Deformation Convolution and Self-Attention

在布匹生产过程中,会产生不同种类的瑕疵,准确的瑕疵检测对于提高纺织工业的生产效率具有重要意义。然而,布匹瑕疵仍存在小目标、极端纵横比、数量不均衡等问题,导致布匹瑕疵检测识别率低、错检漏检频发。因此提出了融合形变卷积和自注意力的素色布匹检测方法。首先,提出了融合形变卷积的多尺度特征提取,缓解模型对不规则瑕疵特征提取能力不足的问题。其次,构建多通道注意力聚合网络,生成新的具有强语义和精确位置信息的瑕疵特征图,提高小目标瑕疵检测的准确率。最后,设计自适应边框生成器,指导生成更精确的瑕疵边界框,解决部分纵横比悬殊的瑕疵无法生成紧密包围框的问题。在标准数据集上的实验结果表明,提出的方法有效提升了布匹瑕疵检测的准确率和效率。

In the process of fabric production,different kinds of defects will occur,and accurate defect detection is of great significance to improve the production efficiency of the textile industry.However,there are still problems such as small targets,extreme aspect ratios,and unbalanced numbers of cloth defects,which lead to a low recognition rate of cloth defect detection and frequent false detections and missed detections.Therefore,a plain cloth detection method integrating deformed convolution and self-attention is proposed.First,a multi-scale feature extraction fused with deformation convolution is proposed to alleviate the problem of the model's insufficient ability to extract irregular flaw features.Secondly,a multi-channel attention aggregation network is constructed to generate a new defect feature map with strong semantics and precise location information to improve the accuracy of small target defect detection.Finally,an adaptive bounding box generator is designed to guide the generation of more accurate defect bounding boxes to solve the problem that some defects with large aspect ratios cannot generate tight bounding boxes.The experimental results on standard datasets show that the proposed method effectively improves the accuracy and efficiency of fabric defect detection.