Hybrid Metaheuristics for Image Analysis

Metaheuristic algorithms are an upper level type of heuristic algorithm. They are known for their efficiency in solving many difficult nondeterministic polynomial (NP) problems such as timetable scheduling, the traveling salesmen, telecommunications, geosciences, and many other scientific, economic, and social problems. There are many metaheuristic algorithms, but the most important one is the Genetic Algorithm (GA). What makes GA an exceptional algorithm is the ability to adapt to the problem to find the most suitable solution—that is, the global optimal solution. Adaptability of GA is the result of the population consisting of “chromosomes” which are replaced with a new one using genetics stimulated operators of crossover (reproduction), and mutation. The performance of the algorithm can be enhanced if hybridized with heuristic algorithms. These heuristics are sometimes needed to slow the convergence of GA toward the local optimal solution that can occur with some problems, and to help in obtaining the global optimal solution. GA is known to be very slow compared to other known optimization algorithms such as Simulated Annealing (SA). This speed will further decrease when GA is hybridized (HyGA). To overcome this issue, it is important to change the structure of the chromosomes and the population. In general, this is done by creating variable length chromosomes. This type of structure is called a Hybrid Dynamic Genetic Algorithm (HyDyGA). In this chapter, GA is covered in detail, including hybridization using the Hill-Climbing Algorithm. The improvements to GA are used to solve a very complex NP problem, which is image segmentation. Using multicomponent images increases the complexity of the segmentation task and puts more burden on GA performance. The efficiency of HyGA and HyDyGA in the segmentation process of multicomponent images is proved using collected field samples; it can reach more than 97%. In addition, the reliability and the robustness of the new algorithms are proved using different analysis methods.

This chapter presents a new algorithm for edge detection based on the hybridization of quantum computing and metaheuristics. The main idea is the use of cellular automata (CA) as a complex system for image modeling, and quantum algorithms as a search strategy. CA is a grid of cells which cooperate in parallel and have local interaction with their neighbors using simple transition rules. The aim is to produce a global function and exhibit new structures. CA is used to find a subset of a large set of transition rules, which leads to the final result, in our case: edge detection. To tackle this difficult problem, the authors propose the use of a Quantum Genetic Algorithm (QGA) for training CA to carry out edge detection tasks. The efficiency and the enforceability of QGA are demonstrated by visual and quantitative results. A comparison is made with the Conventional Genetic Algorithm. The obtained results are encouraging.

Hybridization represents a promising approach for solving any recognition problem. This chapter presents two face recognition frameworks involving the hybridization of both the feature extraction and classification stages. Feature extraction is performed through the two proposed hybrid techniques, one based on the orthogonal combination of local binary patterns and histogram of oriented gradients, and the other based on gabor filters and Zernike moments. A hybrid metaheuristic classifier is also investigated for classification based on the integration of genetic algorithms (GA) and support vector machines (SVM), where GA has been used for the optimization of the SVM parameters. This is crucial since the optimal selection of SVM parameters ultimately governs its recognition accuracy. Experimental results and comparisons prove the suitability of the proposed frameworks as compared to the other baseline and previous works.

The authors develop an advanced hand motion recognition system for virtual reality applications using a well defined stochastic mathematical approach. Hand gesture is a natural way of interaction with a computer by interpreting the primitive characteristics of gesture movement to the system. This concerns three basic issues: (1) there is no physical contact between the user and the system, (2) the rotation of the hand gesture can be determined by the geometric features, and (3) the model parameter must be optimized to improve measurement of performance. A comparative analysis of other classification techniques used in hand gesture recognition is carried out on the proposed work hybrid with the bio-inspired metaheuristic approach, namely the cuckoo search algorithm, for reducing the complex trajectory in the hidden Markov model (HMM) model. An experimental result is as to how to validate the HMM model, based on the cost value of the optimizer, in order to improve the performance measures of the system.

Image enhancement is an essential subdomain of image processing which caters to the enhancement of visual information within an image. Researchers incorporate different bio-inspired methodologies which imitate the behavior of natural species for optimization-based enhancement techniques. Particle Swarm Optimization imitates the behavior of swarms to discover the finest possible solution in the search space. The peculiar nature of ants to accumulate information about the environment by depositing pheromones is adopted by another technique called Ant Colony Optimization. However, termites have both these characteristics common in them. In this work, the authors have proposed a Termite Colony Optimization (TCO) algorithm based on the behavior of termites. Thereafter they use the proposed algorithm and fuzzy entropy for satellite image contrast enhancement. This technique offers better contrast enhancement of images by utilizing a type-2 fuzzy system and TCO. Initially two sub-images from the input image, named lower and upper in the fuzzy domain, are determined by a type-2 fuzzy system. The S-shape membership function is used for fuzzification. Then an objective function such as fuzzy entropy is optimized in terms of TCO and the adaptive parameters are defined which are applied in the proposed enhancement technique. The performance of the proposed method is evaluated and compared with a number of optimization-based enhancement methods using several test images with several statistical metrics. Moreover, the execution time of TCO is evaluated to find its applicability in real time. Better experimental results over the conventional optimization based enhancement techniques demonstrate the superiority of our proposed methodology.

The goal of the segmentation techniques called deformable models is to adapt a curve in order to optimize the overlapping with another image of interest with the actual contour. Some of the problems existing in optimization involve choosing an optimization method, selecting parameters, and initializing the curve. All these problems will be discussed within this chapter, with reference to metaheuristics, and are designed to solve complex optimization and machine learning problems. We discuss image segmentation techniques which depend on active contour models using metaheuristics. Similarly, histological image segmentation techniques are elaborated using a level set approach based upon metaheuristics.

This chapter presents a new hybrid optimization method based on the univariate marginal distribution algorithm for a continuous domain, and the heuristic of simulated annealing for the parabola detection problem. The hybrid proposed method is applied to the DRIVE database of retinal fundus images to approximate the retinal vessels as a parabolic shape. The hybrid method is applied separately using two different objective functions. Firstly, the objective function only considers the superposition of pixels between the target pixels in the input image and the virtual parabola; secondly, the objective function implements a weighted restriction on the pixels close to the parabola vertex. Both objective functions in the hybrid method obtain suitable results to approximate a parabolic form on the retinal vessels present in the retinal images. The experiments show that the parabola detection results obtained from the proposed method are more robust than those obtained by the comparative method. Additionally, the average execution time achieved by the proposed hybrid method (1.57 s) is lower than the computational time obtained by the comparative method on the database of 20 retinal images, which is of interest to computer-aided diagnosis in clinical practice.

Segmentation is a crucial stage in the image analysis process, whose main purpose is to partition an image into meaningful regions of interest. Thresholding is the simplest image segmentation method, where a global or local threshold value is selected for segmenting pixels into background and foreground regions. However, the determination of a proper threshold value is typically dependent on subjective assumptions or empirical rules. In this work, we propose and analyze an image thresholding technique based on a fuzzy particle swarm optimization. Several images are used in our experiments to show the effectiveness of the developed approach.

Evolutionary computation has much scope for solving several important practical applications. However, sometimes they return only marginal performance, related to inappropriate selection of various parameters (tuning), inadequate representation, the number of iterations and stop criteria, and so on. For these cases, hybridization could be a reasonable way to improve the performance of algorithms. Electrical impedance tomography (EIT) is a non-invasive imaging technique free of ionizing radiation. EIT image reconstruction is considered an ill-posed problem and, therefore, its results are dependent on dynamics and constraints of reconstruction algorithms. The use of evolutionary and bioinspired techniques to reconstruct EIT images has been taking place in the reconstruction algorithm area with promising qualitative results. In this chapter, we discuss the implementation of evolutionary and bioinspired algorithms and its hybridizations to EIT image reconstruction. Quantitative and qualitative analyses of the results demonstrate that hybrid algorithms, here considered, in general, obtain more coherent anatomical images than canonical and non-hybrid algorithms.