CPF-DETR: An End-to-End DETR Model for Detecting Complex Patterned Fabric Defects

Abstract

Fabric defect detection is a prevalent issue in the textile industry. For fabrics with monotone coloration and simple patterns, the existing detection algorithms have been able to meet the detection requirements of industrial production. However, there is still a lack of effective detectors to detect fabrics defects with complex patterns and variable colors. This paper proposed an improved RT-DETR model called CPF-DETR, which improves the detection effect by a noise suppression module (NSM) and a novel encoder using dynamic snake convolution (DSC-Encoder). Firstly, RT-DETR as a complete end-to-end real-time detection model was used as our detection framework to avoid the effect of the lack of a priori knowledge of the anchor in industry detection. Secondly, we designed a noise suppression module to filter out noise from complex backgrounds. Furthermore, we introduced the dynamic snake convolution (DSC) into the encoder and designed a hybrid convolution module (HCM) which helps the encoder to enhancing its ability to acquire elongated structure detail information in complex pattern. Finally, we compared our CPF-DETR with many state-of-the-art models on a Complex Patterned Fabric Dataset (CPF) collected from the Aliyun Tianchi fabric defect dataset. The experimental results demonstrate that the accuracy of our detector is superior to existing models. Our detector achieved 69.1% AP outperforming the RT-DETR by 2.3% and yolv8m by 10.6%. The inference latency of 10.46ms is also able to meet the real-time detection requirements.