Investigation and development of fusion techniques using night vision images through statistical and multi resolution approach

Abstract

This thesis deals with the research work involved in the design of efficient fusion schemes for surveillance and navigation applications. The motivation is to develop a method that is more efficient than the existing techniques, both in terms of fusion performance and the complexity involved in real-time implementation. There are four new fusion algorithms presented in this thesis. These involve the use of the multi-resolution decomposition technique and the statistical tool of Principal Component Analysis (PCA). The first method introduces a new fusion scheme using the Discrete Wavelet Transform (DWT), where the threshold for fusion is determined from the decomposed wavelet coefficients. The second method is a weighted fusion algorithm that also uses the DWT decomposition technique. Here the approximation sub-images are fused by determining the weights from the energy of the wavelet coefficients. The third method is a (PCA Gaussian) novel scheme which combines the traditional image averaging method along with additive weighted fusion. The research also aims at choosing the most effective fusion evaluation metric for the specific application being considered. The performances of these four algorithms have been determined for various sets of images in this thesis. Based on the study of various fusion metrics reported in the literature such as Mutual Information, Cross Entropy and Standard Deviation the Structural Similarity Index (SSIM) has been found to be the most effective metric that gives results consistent with the human visual system. Finally, a comparative analysis of the newly-developed fusion rules with the existing techniques has been compared to highlight their efficiency. Finally, the algorithms designed are evaluated for their performance under the effect of sensor noise. The fusion efficiency under noise is computed using the root mean square error as a parameter. In addition, the variation of the SSIM index is studied for various levels of source image degradations. newline newline newline