Assimilation of Pair-Wise Comparison Method to Decide Weights of Features Based on the Content of Image

Computer-based image recognition systems rely on training with an initial training set to recognize similar images. However, when such training set is not available, individual features within a given image can be used to identify and compare with congruent features in database to find similarity among images. An experiment is implemented for extracting similar images using this technique of computer-based image recognition. For this study multiple features color, shape and texture have been extracted from Corel 10k database images. After manual selection of individual feature weightage, all features with equal weightage have been used to identify similar images from database. After observing class wise average precision and average recall of Corel 10k database images, it has been found that equal weightage approach yields satisfactory results and also better as compared with individual feature approach in maximum number of classes. In some of the classes, results are not satisfactory with equal weightage approach. In these classes, results vary with weights of feature according to nature of image. The average precision and recall of entire classes are 96.62% based on all feature equal weight. Now, the problem is to decide weights of features. For this, this study proposes the automatic determination of the weightage for features using pair-wise comparison method. The problem is formulated as multiple criteria decision-making (MCMD) problem and solved through analytic hierarchical process (AHP). Graphical user interface (GUI) is also designed. The proposed approach determines color dominant, shape dominant and texture dominant based on importance on respective features. The average precision and recall of 95 classes are 100% in all dominant features. The average precision and recall of remaining five classes are 59%, 53% and 46% in color dominant, shape dominant and texture dominant, respectively. Performance measure shows that the proposed method archives better results as compared to manual assignment of weights of features.