Multifocus image fusion framework using discrete cosine transform and spatial frequency

Abstract

Image fusion aims to merge all prominent and essential information present in each and every image for effective visualization or any subsequent processing. It serves as pre-processing step in many applications where there is a need for intricate details present in the images. The major application areas of image fusion are medical, defence, remote sensing and astronomy. In recent years, there is an exponential demand for image fusion techniques with the technological improvement in multi sensor technology. The categories of image fusion are multifocus, multiexposure, multimodal, multitemporal or multiview image fusion and this research work emphasizes particularly on multifocus image fusion. Multifocus image fusion aims to combine the in-focus information from individual source images and produce the resultant fused image as all-in-one focus image. In this research, multifocus image fusion is achieved through two different approaches. The first work combines pixel and feature level fusion which comprises of three major steps. As the initial step, the source images are pre-processed by trimmed mean process in blocks of 3x3 to remove noise. The second step is edge super imposition which enhances the edge features and improves image resolution. Since edge features of the source images vary from each other, edge detail from the first image is superimposed on the next image and vice versa. Finally, Dense Scale Invariant Feature Transform (Dense SIFT) is adopted for feature level processing. Dense SIFT descriptor is well suited for extracting orientation information of each and every pixel and it generates a feature vector, which serves as an activity level measurement to decide the resultant pixels in the fused image. The second work combines spatial and frequency domain and gains the advantages of both domains. Frequency domain method is a three stage process - decomposition, fusion and reconstruction. The frequency domain techniques as a general rule modify the coefficients of a transform without considering the spatial consistency of the coefficients. This lack of consideration can result in visual artifacts. The resulting artifact can be prevented by including the spatial domain technique. The spatial domain processing is rich enough to extract local neighbourhood information. Since, Discrete Cosine Transform (DCT) has excellent energy computational property, it is chosen for frequency domain, whereas spatial frequency is adapted for spatial domain. newline