Document Image Classification using SqueezeNet Convolutional Neural Network

摘要

SqueezeNet networks perform well on image classification tasks, achieving accuracies comparable to state of the art convolutional neural networks. In this research we evaluate the performance of SqueezeNet networks in document image classification, showing that an ImageNet pretrained SqueezeNet achieves an accuracy of approximately 75 percent over 10 classes on the Tobacco-3482 dataset, which is comparable to other state of the art convolutional neural networks in terms of accuracy, while containing 50 times less weights compared to them. We then visualize saliency maps of the gradient of our networks output to input, which shows that the network is able to learn meaningful features that are useful for document classification. Features such as the existence of handwritten text, document titles, text alignment and tabular structures, which are proof that the network does not overfit to redundant representations of the dataset itself.