Abstract

This thesis deals with methods for fusion of images as well as videos using multi-scale transforms. First, a novel image fusion algorithm based primarily on an improved multi-scale coefficient decomposition framework is proposed. The proposed framework uses a combination of non-subsampled contourlet and wavelet transforms for the initial multi-scale decompositions. The decomposed multi-scale
coefficients are then fused twice using various local activity measures. Experimental results show that the proposed approach performs better or on par with the existing state-of-the art image fusion algorithms in terms of quantitative and qualitative performance. In addition, the proposed image fusion algorithm can produce high
quality fused images even with a computationally inexpensive two-scale decomposition. Finally, we extend the proposed framework to formulate a novel video fusion algorithm for camouflaged target detection from infrared and visible sensor inputs. The proposed framework consists of a novel target identification method based on conventional thresholding techniques proposed by Otsu and Kapur et al. These thresholding techniques are further extended to formulate novel region-based fusion rules using local statistical measures. The proposed video fusion algorithm, when used in target highlighting mode, can further enhance the hidden target, making it much easier to localize the hidden camouflaged target. Experimental results show
that the proposed video fusion algorithm performs much better than its counterparts in terms of quantitative and qualitative results as well as in terms of time complexity.
The relative low complexity of the proposed video fusion algorithm makes it an ideal candidate for real-time video surveillance applications.