Fabric defect detection algorithm based on improved RT-DETR

Textiles are important raw materials in industry and life, and China's textile industry plays a key role, but there are more than 80 kinds of surface defects in fabric production, which affect the quality and development of the industry. Current detection algorithms have problems such as insufficient accuracy and limited application scenarios. Manual detection and traditional machine vision methods also have obvious defects. Although algorithms based on deep learning have applications, they have their own shortcomings. Therefore, an improved RT-DETR fabric detection algorithm RT-FDTR is proposed in this study: optimizing the backbone network, introducing C2f_AdditiveBlock module to enhance feature extraction ability; designing DHSA-AIFI module to enhance small target detection and anti-interference ability; developing SCOK-CCFF feature pyramid to optimize feature fusion. Experiments on the fabric defect dataset of Aliyun Tianchi show that the P, R and AP50 of the improved model are 82.2%, 77.1% and 76.5% respectively, which are 6.9%, 2.5% and 3% higher than those of the original RT-DETR-r18, and the parameters are reduced by 20.6%. The detection speed is increased by 9.7FPS, which meets the accuracy and real-time requirements of fabric defect detection in industry.