A Texture Classification Method Based on Improved Graph Scattering Transform

Texture classification is an important topic in image processing and has a significant impact on picture analysis. The graph scattering transform network is an untrained, interpretive neural network that can extract graph-level characteristics from textures for categorization. We have worked to develop the network in order to overcome its flaws, which include the inability to extract multi-level information from the graph and insufficiently rich output feature information. In order to perform multi-level information acquisition of the graph, we first implement graph coarsening modules in the graph scattering transform network to coarsen the graph to various degrees before feeding it into the network to extract features. Second, we combine the output features obtained by the overall maximum and average operation to enhance the expressiveness of the output coefficients. Classification experiments on texture datasets KTH-TIPS, UMD and UIUC show that the improved network proposed in this paper achieves superior outcomes.