Structural Image Classification with Graph Neural Networks

1. Project Aims

• 通过将图像表示为无向图来利用区域之间的模式和关系。
• 使用最近提出的Graph Neural Networks [1]模型来处理结构化数据的分类：

2. Image Structure as Graphs

图像的结构上下文可以表示为图G = {N，E}，其中N（节点）对应于感兴趣区域，E（边缘）对应于两个不同区域之间的连接。图1说明了所考虑的结构表示。

1. Region Adjacency Graph (RAG):根据颜色和纹理划分区域，拟合椭圆区域并连接相邻区域
2. Grid structure:
3. Minimum Spanning Tree (MST):最小生成树（MST）：检测具有尺度不变的斑点检测器的区域，Hessian-Laplace [3]。识别每个图像大约200-3000个区域。MST在形成到所有节点的连接路径时产生最低的总成本。
4. Delaunay Triangulation:

3. Node and Edge Labels

4. Filtering Nodes detected with Hessian-Laplace

1. Form a representative node set using k-means clustering in feature space. This is performed on a subset of images from the entire dataset.
2. Apply further heuristics to remove outliers and non-meaningful clusters.应用进一步的启发式方法来删除异常值和无意义的簇。
3. For each image, associate all nodes to an appropriate cluster centroid.
4. Keep the node set in which each node is closet to its associated centroid. This yields a filtered node set containing at most k-nodes.

5. Classification with Graph Neural Networks (GNN)