纱线
纤维
材料科学
人工智能
计算机视觉
计算机科学
工程制图
图像处理
纺纱
复合材料
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-05-01
卷期号:: 2384-2389
标识
DOI:10.1109/yac71005.2026.11615500
摘要
Chemical fiber yarn cakes are key intermediate products in the textile industry, and detecting their tiny surface defects (e.g., hairiness, oil stain and mildew defects) is critical to ensuring fabric quality. These defects feature small scale, weak visual features and low contrast, and are easily obscured by complex texture backgrounds, posing great challenges to traditional machine vision and general deep learning detectors. To solve this problem, a convolutional neural network based on receptive field optimization is proposed for real-time detection of such tiny defects. An improved receptive field enhancement module with a multi-branch dilated convolution structure is designed to expand the receptive field of shallow feature maps and capture fine-grained contextual information, and an efficient multi-scale feature fusion module is constructed to adaptively integrate shallow detail features and deep semantic features for enhancing the representation of defects at different scales. A self-constructed industry-specific dataset for chemical fiber yarn cake surface defects is used for experiments, which makes up for the lack of public datasets in this field of the textile industry and provides a unified benchmark for the performance evaluation of relevant detection algorithms. Experimental results show that compared with similar methods on this self-constructed dataset, the proposed model improves the mean Average Precision (mAP) by 0.5% ∼18.4% while maintaining high detection speed; for the chemical fiber yarn cake defect detection task, the mAP is increased by 1.6% ∼13.2%, with significantly improved recall rates for hairiness and oil stain defects. The proposed method provides an effective solution for accurate detection of tiny industrial defects.
科研通智能强力驱动
Strongly Powered by AbleSci AI