Foreign Bodies Detector Based on DETR for High-Resolution X-Ray Images of Textiles

Abstract

Foreign bodies (FBs) detection for X-ray images of textiles is a novel and challenging task. To solve the problem of poor performance of anchor-based detectors for FBs detection, we propose a feature-enhanced object detection framework with transformer (FE-DETR). Based on the split-attention of residual split-attention network (ResNeSt), we add convolutional block attention module (CBAM) between residual blocks and replace the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3\times $ </tex-math></inline-formula> 3 convolutional layer of the last residual block with deformable convolution network (DCN) to adapt FBs with different scales. Then, we propose a multiscale feature encoding (MSFE) module to solve the feature dispersion caused by deep convolution. Meanwhile, the transformer module is selected as the prediction head of the detector. During training, several heuristic strategies are used to further optimize the performance of FE-DETR. In addition, we construct a benchmark dataset for the textile FBs detection task. With end-to-end training, FE-DETR achieves higher performance than the baseline and mainstream state-of-the-art methods, with mean average precision (mAP) = 0.74, average precision (AP) = 0.992, average recall (AR) = 0.971, and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$F1$ </tex-math></inline-formula> -score = 0.987. This article has been applied to the production line of medical protective clothing during the Corona Virus Disease 2019 (COVID-19) period and has yielded impressive results in actual production.