Signal2seq: Generation Model of Textile Component Analysis Sequence for Textile Fiber Composition Analysis

The analysis of textile fiber composition using deep learning technology represents an intersection between the textile industry and artificial intelligence, holding significant theoretical and practical importance. The goal of textile fiber composition analysis is to identify the constituent materials and their respective proportions, which can be categorized into qualitative and quantitative analysis. Previous approaches have primarily relied on traditional machine learning methods. In this paper, we propose a unified modeling approach for composition analysis tasks using sequence generation techniques in deep learning, simply treating it as a sequence generation task based on input signals from collected samples. Component descriptions (category labels and content labels) are represented as discrete label sequences. We train neural networks to perceive signals and generate the desired sequences. During model training, we observed that different label orders significantly impact model performance. Compared to traditional machine learning algorithms, our method requires minimal assumptions and prior knowledge, achieving state-of-the-art experimental results on the FENIR dataset.