An Image Classification Tool of Wikimedia Commons

Sisi Huang, Humboldt-Universität Zu Berlin
Labelling massive datasets consisting of images from webpages manually is quite time-consuming and also exhausting. If there was a tool which can help us to classify those unlabeled images automatically, it would not overwhelm us nearly as much. In this thesis we aim to extract significant features from images and to automate the annotation of unlabeled images. Due to the variety of images, we focus our attention on solving the problem of chart image classification. Chart images are frequently
more » ... ges are frequently presented in documents and used as a common tool for visualizing relationships within the data. Especially, they are able to distinguish themselves by their patterns or shapes. To deal with this problem we propose machine learning models that can extract the images' features automatically, and predict their labels. Convolutional neural networks are the popular models for solving such problem of image classification. Thus, it is our goal to bridge the relationship between chart images and neural networks. In this thesis we attempt two directions to implement convolutional neural networks: transfer learning and self-training models. On a set of testing data a model using transfer learning based on the VGG-16 pre-trained model, achieves a test accuracy of up to 0.65. Self-training models are LeNet-5, Alex blocks and VGG blocks, which are grounded by AlexNet and VGG. However, performances of self-training models are sightly worse than transfer learning, the highest prediction accuracy of the self-training models is only 0.47.
doi:10.18452/21576 fatcat:thfmsqppsbaifbopvawolkfizq