Article Preview
Top1. Introduction
Efficiency of a classifier depends on its performance at decision level. For example, study on face recognition shows that video description of face is preferable over using still images. As demonstrated in Knight and Johnston (1997), motion helps in recognition of familiar faces better even when the images are negated or reasonably distorted. Bruce et al. (1998) showed that neuropsychological basis of different types of face recognitions (familial and non-familial) were different. The work of Wallis and Bulthoff (2001) also showed that temporal description of face helped human cognition better than a single snapshot of the same. While these works pointed out need of a group of multiple inputs for better recognition of an object, it was also appeared to be supported by the Bayesian Hypothesis for the following reason. Gose et al. (1996) showed in an example the need of multiple test data to enhance accuracy of ELIZA test based diagnosis following Bayesian Classifier. Taking cue from these works, in this study effort was put to incorporate the human cognition criteria at decision level of the classifier to build a successful object recognition system. A cognition analogue was developed to mathematically adopt the face recognition system based on temporal description of object as defined by Wallis and Bulthoff (2001). In this paper, it was shown that judicious clustering of decisions obtained from a group of multiple inputs can lead to considerable improvement in overall classification accuracy over that obtained through single input.
This work was completely different from the commonly known concept of “decision clustering” for improving classification accuracy in cases of Optical Character Recognition (Wilson et al., 1996) and Face Image Recognition (Ebrahimpour et al., 2005). In these works mainly a binary decision clustering tree were formed by merging large number of classes into smaller binary cluster tree at feature space. The purpose of doing so is to replace a single classifier that was incapable to handle large number of classes, by a set of classifiers where each classifier had to deal with smaller number of super-classes.
In case of automatic recognition of most of the natural objects of interest within an image, especially, face, which is actually embedded in a non-linear sub-manifold within image space, features based on Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are found to yield moderate result of efficiency as described by (Li et al., 2008) and (Ahuja & Chhabra, 2011). Therefore, in this work PCA feature was selected to check capability of our cognition derived approach to increase efficiency of face recognition using even this feature.
As for the survey of past research in the area of application of Fractal for object recognition gives following examples. Furthermore, since calculation of fractal dimension of an object gives an idea of its spread within the support space (Feder, 1989) therefore it appears that one can try with a fractal dimension based feature to make up the gap created after PCA based study. A Fractal based image coding has been referenced in Face recognition (Tan & Yan, 1999; Ebrahimpour-Komleh et al., 2001; Kouzani et al., 1997). Fractals have been successfully used in Texture classification (Ferens & Kinsner, 1995) in which they measured generalized information content in the gray level. Evidence on direct use of Fractal dimension in image classification on Face, Brain MRI and Retinal Images can be found in Kouzani and Sammut (1999), Deaton et al. (1994), and Shu-Chen and Yueh-Min (2003). In this regard, Lahiri et al. (2009 & 2011), Lahiri and Dutta (2002), and Singh et al. (2005) showed that a scale and rotation independent multifractal feature, Intensity Level based Multifractal Dimension (ILMFD) could categorize very rough objects like protein aggregates. We found that ILMFD was derived from binary forms of an image constructed from its different intensity segments. The set of Fractal Dimensions calculated from each of these binary forms gives ILMFD. Further, ILMFD was found to capture, not only the similarity at a particular intensity segment to group some objects on the basis of that; but also the discrimination between objects on the basis of data derived from multiple intensity segments. Face having an inherent complexity of natural object, was considered as fit case to be represented by ILMFD feature by us.