Activity Detection and Retrieval for Image and Video Data with Limited Training
Abstract
The main objective of this project is to exploit image analysis tools for solving image segmentation and classification applications. Here we propose two techniques for image segmentation. The first involves an automata based multiple threshold selection scheme, where a mixture of Gaussian is fitted to the image intensity histogram, and the parameters for each of the Gaussian distribution are estimated by learning automata. For our second approach to segmentation, we employ a region based segmentation technique that is capable of handling intensity inhomogeneity, and use a sparse representation based dictionary learning technique to learn the basis function from a given set of training data. The image intensities of the test image to be segmented are then represented as a linear combination of these learned bases. We also implement a kernel based dictionary learning scheme for image classification. Here we extract multiple feature types from the images and learn a dictionary based on the combination of their kernel representation, and implement a mutual information based technique for determining the kernel combination weights. The final classification is based on the minimum reconstruction error for these features. We provide detailed description and supporting experimental results for each of these methods.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 10, 2015
- Accession Number
- AD1001424
Entities
People
- Rituparna Sarkar
- Scott T. Acton
- Zongli Lin
Organizations
- University of Virginia