New Developments in Evolutionary Image Registration for Complex 3D Scenarios

Abstract

It can be argued that image registration is a ubiquitous computer vision task, playing a crucial role in a wide variety of problems from industrial to clinical applications. In essence, the registration task involves the search for a correspondence between images acquired under different circumstances. Relating the information of different sets of imaging data is essential for the application of other subsequent image analysis or computer vision tasks involving its integration, comparison, or manipulation. Along with the latest developments of three-dimensional imaging technology, image registration techniques have considerably evolved to a point where efficiency, versatility, robustness, and real-time performance are feasible goals. Nevertheless, these criteria are only fulfilled in ideal conditions or controlled scenarios. In fact, the intensive computational requirements and the unavoidable imprecision present in real world applications prevent a successful accomplishment of the previous goals. Within the framework of this PhD thesis, the focus has been placed on identifying determinant factors leading to complexity from the point of view of soft computing and artificial intelligence. Both fields provide a huge number of techniques that have thrived in solving complex optimization problems. The relevance of such techniques relies on their ability to endure imprecision while reaching reasonable solutions efficiently in evolving environments where noise, incomplete or ambiguous information is present. The main motivation is thus to exploit the latter advantages to address a multidisciplinary IR problem handling the implicit complexity of recent and significant applications. Recently, different metaheuristic approaches, such as evolutionary computation and swarm intelligence methods belonging to a family of nature-inspired algorithms, have awakened a growing interest of the research community due to their promising capabilities to efficiently handle complex environments. Several adverse conditions affect the performance and effectiveness of image registration techniques. In this contribution, different factors have been identified and assessed in order to overcome the difficulties of three distinct applications. In particular, we have addressed a 3D reconstruction problem involving time-of-flight range images. As a result of the low quality and resolution provided by these novel devices, we designed an image preprocessing pipeline to improve their quality along with a novel swarm intelligence registration method to face this challenging problem. Moreover, the proposed nature-inspired method has been extended to face a 3D medical problem involving the alignment of 3D brain magnetic resonance scans. Finally, soft computing techniques have also been applied to address a forensic identification problem. Precisely, we designed a craniofacial superimposition method involving the articulation of the mandible. This problem is solved by applying a skull-face overlay technique, based on the 3D/2D registration of the articulated skull and the photograph of a missing individual.