Abstract:

The coated fabrics introduce crease damage in the process of manufacturing, transportation, inflation and deflation. The manual crease detection is low efficient. The crease detection accuracy of conventional image processing algorithms cannot meet the practical requirements. An end-to-end deep convolutional neural network for crease detection of coated fabrics is proposed. Datasets are established by performing standard flex durability tests. The encoder and decoder adopt a multi-scale feature fusion structure and an optimized up-sampling model, respectively. Real-time statistics of the crease geometric information are computed by image morphology processing. The crease detection accuracy of the proposed method is as high as 95.78%. In comparison with the conventional semantic segmentation method and other deep learning networks, the performance of the crease detection has been greatly improved.

Key words: coated fabrics, automated crease detection, Deep Convolutional Neural Network（DCNN）, morphological processing

摘要：

涂层织物在生产制造和使用中易产生折皱损伤，人工折皱检测效率较低，传统图像处理方法的检测精度无法满足要求。提出一种基于深度卷积神经网络的涂层织物折皱识别和检测方法。通过标准揉搓试验建立数据集，网络编码和解码器分别采用多尺度特征融合结构和优化上采样模块，使用形态学方法进行折皱几何信息的实时统计。当前检测方法准确率达到95.78%，比传统语义分割技术及其他深度学习模型有很大的提升。

关键词: 涂层织物, 折皱智能识别, 深度卷积神经网络, 形态学处理