Modeling the Relationship Between Texture Semantics and Textile Images

Texture semantics, which is the kind of feelings that the texture feature of an image would arouse in people, is important in texture analysis. In this study, we study the relationship between texture semantics and textile images, and propose a novel parametric mapping model to predict texture semantics from textile images. To represent rich texture semantics and enable it to participate in computation, 2D continuous semantic space, where the axes correspond to hard-soft and warm-cool, is first adopted to quantitatively describe texture semantics. Then texture features of textile images are extracted using Gabor decomposition. Finally, the mapping model between texture features and texture semantics in the semantic space is built using three different methods: linear regression, k-nearest neighbor (KNN) and Multi-layered Perceptron (MLP). The performance of the proposed mapping model is evaluated with a dataset of 1352 textile images. The results confirm that the mapping model is effective and especially KNN and MLP reach the good performance. We further apply the mapping model to two applications: automatic textile image annotation with texture semantics and textile image search based on texture semantics. The subjective experimental results are consistent with human perception, which verifies the effectiveness of the proposed mapping model. The proposed model and its applications can be applied to various automation systems in commercial textile industry.