An Employment of Neural Network Classifiers to evaluate the Performance of Color Feature Descriptors in an Image Retrieval System: An Experimental Survey

Abstract

Content-based Image Retrieval (CBIR) is a system utilized to index and redeem images from a massive image repository, according to the user’s preference. This paper focusses on an experimental analysis on varied color descriptors namely Color Moment, Color auto-Correlogram, Color Histogram, Color Coherence Vector and Dominant Color Descriptor utilized for color feature extraction in the two phases. In the first phase, Average Precision is computed for each technique one by one and in the second phase, a Cascade forward back propagation neural network (CFBPNN) is used in combination with each color descriptor to again calculate the performance of each technique. The classification accuracy of each color descriptor is computed by using both CFBPNN and Patternnet neural network and the results of these neural network classifiers are compared. The results of these analytical experimentations depict that a framework of Color auto-correlogram and CFBPNN outperforms the other color descriptors by obtaining average precision of 97%, 90.5% and 89.5% on Corel-1K, Corel-5K and Corel-10K benchmark datasets respectively.