EMCNet : Graph-Nets for Electron Micrographs Classification

TL;DR

This work proposes an effective end-to-end electron micrograph representation learning-based framework for nanomaterial identification to overcome the challenges and demonstrates that this framework outperforms the popular baselines on the open-source datasets in nanomaterials-based identification tasks.

Abstract

Characterization of materials via electron micrographs is an important and challenging task in several materials processing industries. Classification of electron micrographs is complex due to the high intra-class dissimilarity, high inter-class similarity, and multi-spatial scales of patterns. However, existing methods are ineffective in learning complex image patterns. We propose an effective end-to-end electron micrograph representation learning-based framework for nanomaterial identification to overcome the challenges. We demonstrate that our framework outperforms the popular baselines on the open-source datasets in nanomaterials-based identification tasks. The ablation studies are reported in great detail to support the efficacy of our approach.

Figures (10)

• Figure 1: Overview of EMCNet framework, (a) we represent each image as a grid graph, (b) the $\textbf{GEnc}$ and $\textbf{HGEnc}$ modules of the framework computes the respective grid graph representations, (c) the $\textbf{CTEnc}$ module determines the tree representation, and (d) the output layer predicts the image category.
• Figure 2: For illustration purpose, we split the image into $\textbf{3}\times \textbf{3}$ patches and represent it as an undirected graph.
• Figure 3: For illustration purpose, we split an image into $\textbf{2}\times \textbf{2}$ patches. In the absence of positional information, the GNNs cannot distinguish the images on the left and right.
• Figure 4: The nodes are labeled for illustration. We gradually reduce the graph size in progressive layers, by rejecting the nodes of lower importance and learn the high-level representations through the self-attention based message-passing schemes.
• Figure 5: The nodes are labeled for illustration. We transform the graph into a hierarchical clique tree. The supernodes are known as cliques. The top-left-corner supernode of the clique tree is an induced subgraph(consists of nodes with labels 1, 2, 4, 5, and all of the edges) of the main graph and similarly the other supernodes.