QSVG: quaternion vision graph neural network with structure inference for texture classification

In the realm of texture classification, while traditional approaches like Bag of Words (BoW) and Convolutional Neural Networks (CNNs) are commonly used, recent research highlights the remarkable flexibility offered by Graph Neural Networks (GNNs). Specifically, GNNs can effectively represent image blocks as one-dimensional multi-channel node features from three-dimensional data, interconnecting node feature channels. However, existing methods tend to process channels separately during graph convolution, overlooking critical inter-channel connections that are integral to capture intricate relationships between texture features, potentially limiting the effectiveness of GNNs in texture classification. Moreover, textures contain vital global information essential for classification performance, but GNNs tend to focus on information from similar neighboring nodes, which can restrict their capability to capture the global feature of the entire image during data processing. To address the aforementioned challenges, we propose a novel GNN architecture called Quaternion Structure inference Vision Graph neural network (QSVG) for texture classification, which combines quaternions and structure inference. Unlike previous approaches, QSVG comprises two key components: first, Quaternion Max-related convolution (Qmax), which uses the Hamilton transform to enhance node characterization by exploiting inter-channel connections; second, Structure Inference Block (SIF Block), which enriches node information through Bidirectional Gated Recurrent Unit (Bi-GRU)-based structure inference, fusing graph neural network node features with global features. Experimental results on three standard texture datasets demonstrate the competitive classification performance of our method.