Article Preview
TopFeature subset selection is essentially viewed as an optimization problem, in which it involves searching the possible feature subsets that are optimal or near optimal with respect to certain criterion. There are two main components in every feature subset selection system: (i) Search strategy: used to pick the feature subsets and (ii) Evaluation method: used to test their goodness based on some criteria. The different search strategies are Sequential Forward Selection (SFS) and Sequential Backward Selection (SBS), branch and bound, Genetic Algorithm and hybrid approaches.
M. A. Herráez et al. (2013) have described a hybrid multi objective approach using scatter search based on Non-dominated Sorting Genetic Algorithm (NSGA) for feature selection and relevance feedback for retrieval. They claim that, this approach yields better result in the progressive iterations compared with other methods. In El Alami (2011), SFS and GA are utilized for feature selection and discrimination respectively. The GA is used to obtain the optimal boundaries of the numerical invariant texture and color features for effective discrimination. C.H. Lin et al. (2014) have described the feature subset selection based on GA and two class SVM for classification and retrieval. In this work, two texture features such as Adaptive Motifs Co-occurrence Matrix (AMCOM) and Gradient Histogram for Adaptive Motifs (GHAM) and color feature of an Adaptive Color Histogram for k-means (ACH) were used. In addition with GA, other two approaches such as SFS and (SBS) have also been adopted for feature selection and reported that GA provides robust result with high computational cost when compared with others. S. F. Silva et al. (2011) have introduced the concept of utility function (ranking quality evaluation function) for feature selection. In this work, a set of ranking evaluation functions are employed in the feature selection algorithm based on GA and they also act as a quality measurement for retrieval than the current approaches based on classification error. From the literatures, it is found that GA combined with any one of the classifier model (Lin et al., 2014; Jiang et al., 2006; Huang and Wang, 2006) is found to be robust than other methods at the expense of more computational cost. Though feature subset selection improves the performance of the CBIR system, in order to further increase the retrieval efficiency and to bridge the semantic gap and also incorporate human into loop, relevance feedback mechanism is used.