An Adaptive Multi-granularity Graph-based Representation for Image

Current image processing methods typically operate at the finest granularity level-pixel level, which poses challenges in terms of efficiency and interpretability in deep neural models. Inspired by the idea of multi-granularity cognitive computing, we explore the construction of a multi-granularity representation for images, aiming to build efficient and interpretable deep learning models. In this work, we propose a method to represent images as graph, where each node represents a structural block in the image and each edge represents the association between two nodes. Specifically: (1) We have designed a gradient-based strategy to adaptively segment the image into several rectangular regions, where each region can be considered as a node. (2) Each node is connected to the nodes sharing the same region, and we have extracted image features corresponding to the nodes as attributes for each node. (3) Finally, all nodes and their corresponding edges form the graph representation of the digital image. In experiments, our method, combined with graph neural network (GNN), is applied to two benchmark datasets for image classification tasks. The results show that our proposed method significantly improves the learning and generalization of GNN models in computer vision tasks compared to previous studies.