Article Preview
TopIntroduction
The image fusion can be explained as the process by which a set of images is combined to construct a new image with the actual facts from the sources. The resultant image acquired from image fusion technique is more clear and suitable for the purposes of human visual perception and further computer processing tasks (Goshtasby & Nikolo, 2007). However, the limitations in the depth-of-focus of optical lenses in charge-coupled devices leads to acquire an image that may not contain all relevant objects in focus. To obtain an image with all-in-focus, an image fusion process is needed to combine the images taken from same view point under different focal settings. In this paper, a novel method for multi-focus image fusion is proposed.
During the past two decades, many image fusion methods are developed (Li, Yang & Hu, 2011). The image fusion algorithms have been categorized into pixel, feature, and decision levels (Hu & Li, 2012) based on the representation format at which image information is handled. Pixel-level image fusion is the process of extracting information according to the strength of pixels from more than one images of the same scene to get factual inputs which helps in further image processing. Such methods blend more than one input images into a single composite image in a raw image delineation and conserve more real information compared with feature- and decision-level methods.
The multi focus image fusion has been useful in the areas of medical imaging by combining the images obtained by Computed Tomography (CT) scan and Magnetic Resonance Imaging (MRI) to get more information and additional data from fused image (Agarwal & Bedi, 2015). In Remote sensing, image fusion can combine the spatial information of panchromatic (PAN) image and the spectral detail of a low-resolution multispectral (MS) image to construct a fused MS image with high spatial resolution (Li, Yin & Fang, 2013). In traffic control, different image processing and data fusion techniques were tested and evaluated to provide real-time output as well as to obtain useful statistical information such as traffic loads, lane changes, average velocity, etc. (Koutsia, Semertzidis & Dimitropoulos,2008). Image fusion combined with perimeter surveillance and automatic target detection is helpful in detecting the intruder accurately (Riley & Smith, 2006). In all such cases, analyzing a series of input source images one by one is impractical and inefficient. Hence, there is a need to perform fusion of the source images and to enhance the salient features for better interpretation of the images. The present research study particularly focuses on the fusion methods to be applied on multi-focus image datasets namely Lab, Jug, Doll, Desk, Clock and Book.
Several basic transform-domain schemes were proposed such as (i) fusion by averaging: fuse by averaging the corresponding coefficients in each image (“mean” rule), (ii) fusion by selecting the greatest in absolute value of the corresponding coefficients in each image (iii) fusion by de-noising: perform simultaneous fusion and de-noising by thresholding the transforms coefficients (iv) high/low fusion, and, many more (Mitianoudis & Stathaki, 2007). These methods do not suit for multi-focus image data sets.
Since multiscale transforms can effectively extract the important information of the images such as lines and details, they are the most commonly used methods to fuse images. To name a few, the Laplacian pyramid transform (LAP) (Burt & Adelson, 1983), the discrete wavelet transform (DWT), the Principal Component Analysis (PCA) and the dual-tree complex wavelet transform (DT-CWT) (Lewis et al. 2007). To better capture the intrinsic geometrical structure of images, various multiscale geometrical analysis methods including curvelet, contourlet and so on have been developed (Hu & Li, 2012). Recently, a window empirical mode decomposition (WEMD) based method has been developed (Qin et al., 2017) for multi-focus image fusion. The rules were framed rules for bidimensional intrinsic mode function (BIMF) components and the residue component. The sum-modified Laplacian and local regional visibility concept were used for fusion of these two components respectively. Few of them are utilized for the design and implementation of present study.