Welcome | Publications | Research | Industry | Teaching | Biography
UofT Logo (Small)
Mahdi Marsousi
Academic Affiliation:
Multimedia Processing Lab, 
Electrical and Computer Eng. Dept., 
University of Toronto,
Toronto, ON, Canada
Industrial Affiliation:
Research Engineer,
Research and Development Department,
Magna Electronics Inc.,
1 Kenview Blvd, Suite 200,
Brampton, ON, L6T 5E6, Canada
Contact Information:
Academic email:
Work email: 
Email: marsousi@gmail.com
Cell-phone: (647) 967-1585

Adaptive size dictionary learning:

Sparse representation via dictionary learning has found many applications in image processing, such as image compression, image super-resolution, image de-noising, image in-painting, classification, and even image segmentation. Dictionary learning is referred to the process of training a set of basis function, compactly representing patches of a set of images. In other words, each column, also called “atom”, of the learnt-based dictionary is a basis function, learnt to better represent patches of a scene of interest. The number of dictionary atoms has a great impact on the performance of sparse representation. A too small dictionary results in losing details of representing a scene of interest, whereas a too-large dictionary results in overfitting to the training dataset. Overfitting is not desired in de-noising and clustering applications since it lets noise enter the data representation. Although there have been a lot of works reported in literature to address the dictionary learning problem, there has been comparably much less efforts to find an efficient number of atoms for learnt-based dictionaries.

In this research, a novel dictionary learning approach is designed to automatically select an efficient number of dictionary elements to meet representation requirements, including desired representation error and desired average sparsity level, for a given training data set. In the proposed approach, an initial dictionary with a starting number of atoms is progressively spread along the high-dimensional space of image patches until the majority of samples (patches) are sufficiently supported by dictionary atoms. The proposed method is called dictionary learning with efficient number of elements (DLENE).

Adaptive length dictionary learning.

Additional resources:

Download source code here.

2016 (c) Copy-right preserved, Welcome | Publications | Research | Industry | Biography