Image clustering for mobile applications

Abstract

This project looks into how images can be grouped into clusters using a variety of clustering methods and parameters. The clustering results are aimed at satisfying a few different use case scenarios concerning mobile users.

The first use case is concerned with how mobile users can be presented with clusters of images that are relevant in context. By using clustering algorithms like fuzzy c-means, k-means and subtractive clustering, images are clustered first by the color and texture inputs, and then by their location and time information inputs. A database of 3 NTU-newsworthy events consisting 201 images taken under various considerations serve as the first set of images experimented with. Initial clustering results were adjusted with suitable modifications to the input parameters and clustering methods, such as experimenting with different number of clusters to form and different estimates for cluster centers. The best visually-clustered images context-wise are then presented.

The next use case is concerned with how a mobile user can organize his travel blog effectively. The travel blog should be concise, with images fronting the blog to be the most representative to catch browsers’ attention. Scale Invariant Feature Transform (SIFT) keypoints are used as the inputs for the forming of clusters. The clustering methodology is designed based on a few guidelines that considers users’ expectations, and tested against a database of 100 images containing scenes of various NTU landmarks. The strengths of the SIFT algorithm, such as invariance to image rotation and scaling, are evident in the clustering results. The clustering results are reasonable, and improvements can be made on the method of clustering.