Stability and reduction of statistical features for image classification and retrieval: Preliminary results

Content-based image retrieval (CBIR) and image classification are challenging problems in computer vision. In both fields, feature extraction plays an important role in ensuring effectiveness and stability of the results. In this paper, we present an experimental study to test the robustness of the features we proposed earlier. We demonstrate that the extraction of statistical features locally using small block size and different color models makes the features more robust and stable. The features perform well under dimensionality reduction and for different validation approaches and several classification algorithms. Discrete wavelet transform (DWT) is used for dimensionality reduction and re-sampling. The experimental results show that the extracted features are stable enough for classification and the classifiers perform well after dimensionality reduction. In addition to reducing the training time significantly while maintaining almost the same system performance, and therefore, such statistical features can be efficiently used for CBIR and image classification.