School of Science, Engineering and Information Technology

Large scale image classification based on object proposal

Project Title:

Large scale image classification based on object proposal

Supervisor(s):

Dr Dengsheng Zhang

Contact person and email address:

Dengsheng Zhang dengsheng.zhang@federation.edu.au

A brief description of the project:

There is a massive amount of image data on the Internet and in the storage, however, find a desirable image for a user is still a challenge task. This is due to image data are not structured as textual data and there is a deep lack of understanding of image data. Traditional approaches to image data classification are based on semantic labels collected from web pages or from manual annotations. However, these approaches are impractical because vast majority of images are either unlabelled or mislabelled.

In this research, we set image classification as an object proposal problem, and investigate/develop automatic techniques to classify images based on dominant object presented in each type of images. The project will investigate and make use of ImageNet which is the largest image collection for research purpose to learn different types of objects for image classification. The project will investigate most effective features such as colour, texture and shape for each types of objects and use convolutional neural network (CNN) to learn such types of features for best classification outcomes.

The project will aim at the frontier of image data mining and machine learning by using the sate-of-the-art of image analysis methodology and benchmark datasets. New method of object proposal will be resulted from the research and a new toolbox/system of image classification will be developed as an outcome.

Time frame: three years. In the first year, the candidate will conduct a literature review on state-of-the-art methods of image classification, investigate ImageNet database for data collection and prototype an image classification toolbox for experiments. In the second year, the candidate will investigate various features for object description, develop a new method of object proposal and conduct CNN learning. In the third year, the candidate will focus on compiling results and writing the thesis.