基于深度学习的印花织物循环图案基元分割

Primitive Segmentation of Repeat Patterns on Printed Fabrics Based on Deep Learning

在实际的生产过程中,织物的印花图案往往由循环图案基元排列而成。然而基于传统的人工织物循环图案基元分割会消耗大量设计成本,因此实现基元的自动分割,有非常重要的研究意义。近年来,随着深度学习技术的快速发展,为解决该问题带来新的希望。针对印花织物循环图案基元分割,该文提出了一种基于深度学习的循环图案基元分割算法。该算法利用预训练AlexNet网络的卷积层进行特征提取,织物图像输入网络后,在网络特征层中会产生规律的激活峰值,每对峰值对应一组位移向量。并且对位移向量进行投票,出现次数最多位移向量的绝对值即为循环图案基元的尺寸。随后在印花织物中找到对应区域,从而实现完整循环图案基元的分割。对比传统算法,该算法不仅可以分割出简单印花织物的循环图案基元,还可以分割复杂印花织物的循环图案基元,达到了更高的准确率,具有更强的鲁棒性。

In the actual production process, the printed patterns of the fabric are often arranged by the repeat pattern primitives. However, the primitive segmentation based on the traditional artificial fabric loop pattern consumes a lot of design costs, so the realization of the automatic segmentation of primitives has very important research significance. In recent years, with the rapid development of deep learning technology, it has brought new hopes for solving this problem. Aiming at the segmentation of repeat pattern primitives on printed fabrics, we propose a repeat pattern primitive segmentation algorithm based on deep learning. The algorithm uses the convolutional layer of the pre-trained AlexNet network for feature extraction. After the fabric image is input to the network, regular activation peaks will be generated in the network feature layer, and each pair of peaks corresponds to a set of displacement vectors. And the displacement vector is voted, and the absolute value of the displacement vector that appears the most times is the size of the repeat pattern primitive. Then the corresponding area in the printed fabric is found, so as to realize the segmentation of the complete repeat pattern primitives. Compared with the traditional algorithm, the proposed algorithm can not only segment the repeat pattern primitives of simple printed fabrics, but also segment the repeat pattern primitives of complex printed fabrics, achieving higher accuracy and stronger robustness.