Multi-sensor image fusion using optimized and adjustable non-subsampled shearlet transform and measurement of fusion performance
No Thumbnail Available
Image fusion is a technique to assimilate information acquired from similar or dissimilar sources of images. Compared to the source images, fused images convey more information, and they have more clarity. However, distorted fused images are obtained in many cases, which only misinterpret the actual scene information. This happens due to the shortcomings of fusion rules, fusion framework, image decomposition techniques, and inaccurate source image registration. To address these issues, we proposed few efficient fusion techniques to generate better quality fused images for different computer vision applications. In the proposed method presented in Chapter 2, non-subsampled shearlet transform (NSST) is used to decompose pre-registered source images into low- and high-frequency components. These low- and high-frequency coefficients are fused by using our proposed modified weighted salience and local difference fusion rules, respectively. To enrich edge information in the fused image, Canny edge detector with scale multiplication is employed. Moreover, a metric QTS is proposed to jointly measure both texture and structural information present in the fused image. The proposed metric is formulated on the basis of local standard deviation filtering, local information entropy, and local difference filtering. Both subjective and objective results validate the proposed fusion framework and the metric QTS. The proposed method can deal with noisy source images due to the deployment of anisotropic diffusion filtering. However, proposed method is unable to produce state-of-the-art results in case of multifocus images due to non-optimal frequency response of NSST filters and lack of additivity in fusion rules. Moreover, there is still a scope of formulating a new metric to measure other attributes of fused images.
Supervisor: M.K. Bhuyan
ELECTRONICS AND ELECTRICAL ENGINEERING