Performance Analysis of Dictionary Learning Techniques for Sparse Representation based Image Fusion

Abstract

Images are real description of objects. Due to the limited depth of field (DOF) of the conventional cameras, a complete description of the scene with all the relevant objects in focus cannot be acquired in a single image. For accurately analysing and interpreting the images, image fusion technology came into existence which combines the relevant and complementary information of multiple images into a single comprehensive image. There are many algorithms proposed in the literature for image fusion. Recently, Sparse Representation (SR) theory has gained significant importance in the field of image fusion. Proper construction of an over-complete dictionary plays a very crucial role in sparse representation based fusion algorithms. There are two key issues involved in this. First one is the ability of the dictionary to fit the given input data and the second one is the computational effort taken during the learning process. The existing techniques either concentrate on the former part or on the latter part but fails to satisfy both the criteria simultaneously in a single fusion algorithm. Motivated by these concerns, supervised classification based dictionary learning approaches are proposed for fusion of different types of source image pairs. These approaches are denoted as Global Dictionary Based on Focus Feature Classification (GDFFC) and Global Dictionary using Morphology Based Classification Scheme (GDMC). In both the approaches, the global dictionaries are constructed based on pre-classification of initial training data set. The computational effort of the dictionary learning process is greatly reduced by the parameter settings of the dictionary which include the training data size, number of iterations and the size of resultant over-complete dictionary. The subjective and objective evaluations prove that both the approaches produces best visual effect and competes the existing state-of-the art methods. newline