An efficient method for texture-based image retrieval

The extraction of significant features from an image is a primary process in content-based analysis of image data. It is also a computationally intensive process. Using an appropriate approximation of the original image with effective signal processing can decrease the computational cost of the feature extraction process by orders of magnitude. This thesis shows that a pertinent approximation for texture images when used with a proposed distance measure can lead to a significant reduction in the cost of texture image analysis without introducing any significant error in the retrieval performance. An evaluation of the hypothesis is conducted in two popular texture analysis methods, Gabor wavelets and GLCM. Empirical results on Gabor wavelets show that approximate textural feature with the proposed distance measure works 5 times faster and equally robust, when compared to the previously reported methods. Using the GLCM, with approximate textural features works 13 times faster without any significant loss in retrieval performance. A normalization technique proposed for multi-channel distance measure provided 2% improvement in overall retrieval performance.