Metaheuristics. Many problems in Computer Vision are formalized as an optimization problem, and sometimes no direct solving algorithms or convergent iterative methods are known. In this cases one can try to solve them with the help of metaheuristics. These methods optimize a problem by iteratively trying to improve a candidate solution with regard to a given measure of quality. Some of the classes of the metaheuristics like evolutionary algorithms, swarm intelligence or stochastic optimizations are evaluating different candidates independently in each iteration, making these calculations easily parallelizable [22, 5, 32] Current Technologies. Because of the close relationship the pixel level computer vision algorithms were the first algorithms which were ran on the GPUs. • 1st generation: These early implementations used OpenGL Cg or GLSL and the calcula- tion itself happened in the standard OpenGL pipeline with the help of FBO-s [6, 23, 27] The creators of GPUCV [12] and OpenVidia [24] used this technology to implement parts of OpenCV [39] as the de facto standard computer vision library. • 2nd generation: After its release, the adoption of CUDA was immediate [3, 5, 13, 20, 22, 25, 29, 32, 39, 24]. The adoption of CUDA has not been largely affected by the newer OpenCL standard, except that library writers usually favor OpenCL because of its openness [39, 24, 11, 18, 15].
Metaheuristics. Metaheuristics is described as a heuristic which controls and guides another heuristic to find solutions better than those generated at a local optimum i.e. a heuristic within a heuristic. Another description given is a heuristic that escapes a local optimum by accepting worsening solutions and searching through most of the solution space. Two main classes are known but with many variations and names; Local search based and Population based metaheuristic. Local Search based metaheuristics start with an initial solution and moves to another solution in its neighbourhood utilizing a neighbourhood operator and a criterion for accepting solutions. Among these include Xxxx search, Simulated annealing and Deterministic annealing. The concept underlying xxxx search is briefly described later in section 3.3.4.1 for the purpose of this project. Population based metaheuristics combine a set of initial solutions to generate a subsequent set of new solutions and the population updated. Adaptive memory procedure by Xxxxxx and Xxxxxxxx[12] updates the population by replacing worst individuals with better ones. Genetic algorithms apply the concept of reproduction where alleles represent parent solutions which are combined to produce offspring. A survey on the fleet composition and routing problem in the maritime sector is presented by Xxxxxxxxx et al.[15]. The authors present an overview of industrial cases which have been studied over the years from as early as 1988 to 2007 with the solution approaches adopted in each case.
