A Comparative Evaluation of Feature Extraction and Similarity Measurement Methods for Content-based Image Retrieval
International Journal of Image Graphics and Signal Processing
Content-based image retrieval is the popular approach for image data searching because in this case, the searching process analyses the actual contents of the image rather than the metadata associated with the image. It is not clear from prior research which feature or which similarity measure performs better among the many available alternatives as well as what are the best combinations of them in content-based image retrieval. We performed a systematic and comprehensive evaluation of several
... isual feature extraction methods as well as several similarity measurement methods for this case. A feature vector is created after color and/or texture and/or shape features extraction. Then similar images are retrieved using different similarity measures. From experimental results, we found that color moment and wavelet packet entropy features are most effective whereas color autocorrelogram, wavelet moment, and invariant moment features show narrow result. As a similarity measure, cosine and correlation measures are robust in maximum cases; Standardized L 2 in few cases and on average, city block measure retrieves more similar images whereas L 1 and Mahalanobis measures are less effective in maximum cases. This is the first such system to be informed by a rigorous comparative analysis of the total six features and twelve similarity measures. features values are extracted for query image and images in the database. All the features of an image form a feature vector. Then these feature vectors are evaluated using similarity measures to find the similar images. Our research objective is to find out the most effective feature extraction method(s) and similarity measurement method(s) in case of content based image retrieval, which will help the future researchers. To do this, we undertook a rigorous comparison of two different color feature extraction methods-color moment and color autocorrelogram; three different texture feature extraction methods-wavelet moment, wavelet energy, and wavelet packet entropy, and one shape feature extraction method-Hu's seven invariant moments. Again, we focused here on twelve measures, they are: L 1 , L 2 , Standardized L 2 , Normalized L 2 , Mahalanobis, City block, Minkowski, Chebyshev, Cosine, Correlation, Spearman's rank correlation, and Relative deviation. These are chosen because of their extensive and successful applications to many datasets in CBIR. Several methods       ,     exist that use some of the RST invariant feature extraction methods and/or similarity measurement methods mentioned above, but a comprehensive evaluation applying all these methods does not exist. Thus it is not clear from the literature which feature or which similarity measure performs better among the many available alternatives. It is also currently poorly understood, what are the best combinations of feature extraction and similarity measurement methods in CBIR. The present study allows determinations of the features as well as similarity measures that perform best for similar image retrieval among the many available alternatives. It also allows identification of the best performing combinations of them. For experiments, we used mostly applied Wang database  which contains ten categories of image, hundred images in each category. From the experimental results, we found that color moment and wavelet packet entropy features are most effective features whereas color autocorrelogram, wavelet moment, and invariant moment features show worse performance among all features used here. Again, as a similarity measure, cosine and correlation measures show the best accuracy in maximum cases; Standardized L 2 in few cases and on average, city block measure retrieves more similar images, whereas L 1 and Mahalanobis measures show the worse result in maximum cases. The remainder of the paper is organized as follows. In section II, the materials and methods are described in details. The Results and comparative evaluations are outlined in Section III and we draw our conclusion in the last section.