In this paper, a novel feature‐based attention network with sequential information is proposed for fabric defect detection. As an important part of the textiles industry, fabric quality inspection needs to be automated in the latest wave of intelligent transformation. To this end, fabric defect detection algorithms have been widely studied. In this paper, an efficient defect detection model for fabrics was built. As the appearances of defects change with the type of fabric, manual features which denote the overall situation of the fabric are used as prior knowledge. The feature‐based attention module discussed in this paper can generate attention maps from these manual features to rectify the responses of the feature maps according to the whole situation of the input image. Multidirectional long short‐term memory networks are implemented to extract context information from continuous defects. When making a judgement, taking sequential information into consideration may reduce the number of unexpected misjudged decisions compared with the number of those made independently pixel by pixel. Both of those two modules can be integrated into any existing convolutional neural network model and trained in an end‐to‐end manner. A fabric defect dataset is built to train and test the models. In this paper, several models with different architectures are implemented to verify our ideas, and are supported by results confirming the efficiency of the proposed methods.

中文翻译：

---

具有顺序信息的新型基于特征的网络，用于纺织品缺陷检测

在本文中，提出了一种具有顺序信息的新颖的基于特征的注意力网络，用于织物缺陷检测。作为纺织品行业的重要组成部分，需要在最新的智能化转型浪潮中使织物质量检查自动化。为此，已经广泛研究了织物缺陷检测算法。本文建立了一种有效的织物缺陷检测模型。由于缺陷的外观随织物类型的变化而变化，因此将表示织物总体情况的手动功能用作现有知识。本文讨论的基于特征的注意力模块可以从这些手动特征中生成注意力图，以根据输入图像的整体情况来纠正特征图的响应。多方向长短期存储网络的实现是从连续缺陷中提取上下文信息。进行判断时，与逐个像素独立做出的决策相比，考虑到顺序信息可以减少意外决策错误的数量。这两个模块都可以集成到任何现有的卷积神经网络模型中，并以端到端的方式进行训练。建立织物缺陷数据集以训练和测试模型。在本文中，实现了几种具有不同体系结构的模型以验证我们的想法，并得到证实所提出方法效率的结果的支持。与逐个像素独立做出的决策相比，考虑到顺序信息可以减少意外决策错误的数量。这两个模块都可以集成到任何现有的卷积神经网络模型中，并以端到端的方式进行训练。建立织物缺陷数据集以训练和测试模型。在本文中，实现了几种具有不同体系结构的模型以验证我们的想法，并得到证实所提出方法效率的结果的支持。与逐个像素独立做出的决策相比，考虑到顺序信息可以减少意外决策错误的数量。这两个模块都可以集成到任何现有的卷积神经网络模型中，并以端到端的方式进行训练。建立织物缺陷数据集以训练和测试模型。在本文中，实现了几种具有不同体系结构的模型以验证我们的想法，并得到证实所提出方法效率的结果的支持。