PERFORMANCE STUDY OF CONTENT BASED IMAGE RETRIEVAL SYSTEM

A plethora of researchers through their research work proposed various Content Based Image Retrieval (CBIR) systems which retrieves images from data base similar to the query image. Since it is an evolving field, there is lot of scope for improvement in the performance parameters of Precision, Accuracy and Recall. The main objectives of the present work is to improve the performance of Content Based Image Retrieval System. Colour, straight line, outline sketch and texture signatures of images stored in the database and query images are extracted and categorized with SVM neural network. Signatures of images of database which are assigned same category as that of query image are compared with the signatures of query image. Images are ranked based on the similarity. Most similar images are retrieved as the relevant images of the database. This is achieved by the methods such as Region Colour Signature (3RCS), Combined Colour and Straight Line Signature (CSLS), Combined Colour, Straight Line and Outline Signature (CSLOS), Improved CBIR with combined Colour , Straight Line and Outline Signature (ICSLOS), Combined Texture, Colour , Straight Line and Outline Signature (TCSLOS), Colour Co-Occurrence Matrix with combined Colour , Straight Line and Outline Signature (CCMCSLOS) ,Support Vector Machine Classifier Based CBIR with combined Colour , Straight Line and Outline Signature (SVMCSLOS) and Combined Colour, Straight Line and Outline Signature (CSLOS) using Mahalanobis distance measure methods. Performance of these experimented methods are studied by measuring metrics such as False Positive Rate, Precision , Recall, FI Score , G measure , True Positives, True Negatives ,Accuracy and Balanced Accuracy . The performance of SVMCSLOS method was better than all the other experimented methods.