Dynamic graph neural network for super-pixel image classification

Abstract

Convolutional Neural Networks (CNN) haveachieved a huge success in computer vision tasks. In spite of thefact that some CNN models can out-perform human in manyspecific tasks, the traditional CNN models can only handleimages with fixed dimensions, i.e. width and height. Currentstate-of-the-art CNN models overcome the issue by resizing inputimages to maintain the consistency of dimensions among inputs.This approach has the main drawback of over up-sampling orover down-sampling with images whose dimensions are far toodifferent to the standard of the models. Moreover, most of CNNmodels are built over-parameterized, which means most of themare heavier than necessary. In this paper, we aim to make useof graph neural networks to broaden the very new researchfield of applying the networks on visual tasks. We propose aposition-aware dynamic graph propagation scheme to handlesuper-pixel images created by popular super-pixel segmentationalgorithms. Although our model did not out-perform state-of-the-art traditional CNN models on all datasets due to theinevitable information loss of segmentation step, we achieveda huge improvement in accuracy compared to existing graphneural network methods and achieved state-of-the-art accuracyon Superpixel MNIST-75 dataset by 23% error rate drop; wealso reduced the number of parameters by 99% compared toVGG16 model.