FocusDet: towards high-quality digital printing fabric defect detection

Abstract

Deep-learning models have been effectively applied to the fabric defect detection field, in which dilemmas still exist for further improving product quality. For the self-built digital printing fabric defect detection dataset, the dilemmas can be expressed in aspects. First, the existing detection models are more inclined to learn many shot categories (head classes) and directly ignore low shot categories (tail classes); Second, the sampled positive and negative anchors in each mini-batch are not equally important, therefore they should be unequally attended to according to their importance. To solve these problems, in this article, a high-quality model for digital printing fabric defect detection was proposed, termed FocusDet. Specially, we construct the model based on the Faster-RCNN framework with two well-designed modules: the balanced group softmax module and the importance-based sample reweighting module, which improve the detection accuracy. Experimental results demonstrate that our proposed model reaches state-of-the-art accuracy on COCO metrics compared with other advanced detection models in the digital printing fabric defect detection dataset.