Article Preview
Top1. Introduction
Image registration is a process to spatially align two geometrically displaced images, one of them called the source image and the other reference image, so that the corresponding points assume the same coordinates. Registration of images plays a vital role in various fields ranging from medical diagnostics, weather forecasting, crop monitoring, military surveillance to computer vision and artificial intelligence. There have been a number of established theories and techniques described in previous references for image registration (Moigne, Netanyahu, & Eastman, 2011; Brown, 1992; Goshtasby, 2012; Zitova & Flusser, 2003). However, with increase in variety, complexity and heterogeneity of images and in the view of the vast amount of information available, the need for newer of image processing algorithms has become necessary and this field is continuously evolving (Brown, 1992). In addition to the classical analytical techniques like correlation based or frequency plane-based methods (Goshtasby, 2012), optimization based iterative methods also takes a very important part in image registration.
When two images only have translational, rotational or affine differences, an intensity-based rigid transformation algorithm can be used for registration, for example principle axes registration or multiresolution registration methods. However, in presence of additional changes other than the geometric mismatch, for example, topological, non-rigid methods, such as adaptive transformation methods are used. In such cases, a registration may be considered optimal if a criterion of similarity or dissimilarity measured is defined, and the optimization algorithm either maximizes (similarity) or minimizes (dissimilarity) the measurement criterion. It is not possible to prescribe a universal method applicable to all types of image registration tasks. This is due to variation in image acquisition process, as the images can be acquired at different times (multi-temporal), from different viewpoint (multi-view) or from different sensing devices (multi-modal) (Argyriou et al., 2015).
For non-rigid image registration, in addition to Gradient-descent algorithms, stochastic optimization methods have been proposed which worked well on multimodal and multiresolution images (Klein, Staring, & Pluim, 2007). In recent years, research in metaheuristic approaches are gradually growing due to its flexibility and ease of applicability. Swarm Based meta-heuristic algorithms, imitating the behavior of the swarming of a species occurring in nature are frequently being used in the optimization problems in diverse fields of research due to their simplicity and easier implementation than strictly analytical algorithms. Some of the examples in this category are Particle Swarm Optimization (PSO) algorithm (Eberhart & Kennedy, 1995), Ant Colony optimization algorithm (Dorigo & Di Caro, 1999), Cuckoo Search optimization algorithm (Gandomi, Yang, & Alavi, 2013), Firefly algorithm (Yang, 2009) etc. Among all the swarm-based algorithms PSO based image registration algorithms have been reported extensively in literature (Wachowiak et al., 2004; Pramanik, Dalai & Rana, 2015; Wang & Bian, 2012 ; Maddaiah & Pournami, 2019) though there have been very few attempts of using other swarm-based algorithms for image registration task (Daniel & Anitha, 2016; Zhang et al., 2013).