Texture modeling and labeling using divide and conquer

Texture detection plays an important role in many computer vision tasks. Application ranges from finding an object in satellite images to anomaly detection in medical imaging. Recently sparse representation based texture modeling and detection scheme is proposed. One major drawback using such a generative data modeling is the time taken for texture detection. In this paper we propose and investigate divide and conquer strategy in sparse framework. Our study shows that divide and conquer strategy can be efficiently used with sparsity based data modeling without compromising the accuracy. Advantage of such an approach is reduced computation with minimal degradation in detection accuracy. Detection is performed by comparing a test texture with each dictionary, in conventional approach; whereas the proposed approach speeds up detection using divide and conquer strategy. Detection time complexity of conventional and proposed approaches for n classes are O(n) and O log(4n) respectively. Proposed approach learns dictionaries at different levels where a group of classes are merged to form a reduced set of classes. This leads to learning overhead. We also propose an algorithm to remove the learning overhead.