Abstract:
Texture analysis has been an active and important area due to significant applications in the field of image processing, machine vision and pattern recognition. It further has many potential sub-areas that have extensive scope of applications ranging from remote sensing to robotics path tracking, industrial material categorization and testing to quality control, satellite imagery to natural scene imaging and biomedical diseases detection to document analysis. This thesis aims to comprehensively analyze several state of the art texture descriptors for classification of images and based on this analysis to propose an efficient and computationally simple set of descriptors. We performed texture classification utilizing descriptors which have been widely used including Gabor Wavelets (GW), Local Binary Patterns (LBP) with its most recent extensions and Segmentation based Fractal Analysis (SFTA). These descriptors have invariance qualities against scale, orientation and illumination in textured images. Some of the benchmark texture databases including Brodatz album, USC-SIPI, Outex-10, Outex-12, KTH-TIPS and UIUC were used for feature extraction. Classifiers like K Nearest Neighbors (KNN), Support Vector Machine (SVM), Naïve Bayes, Random Decision Forest (RDF) and Multilayer Perceptron (MLP) were used for testing performance of classification. We compared performance of all descriptors and among all descriptors it was observed that LBP’s most recent extension Dominant Rotated Local Binary Pattern (DRLBP) gives better performance. Based on this comparison we proposed a novel descriptor that outperformed all descriptors on four out of five datasets. The proposed approach is a combination of GW, SFTA and uniform LBP descriptors that gives highest accuracy while having very low and simple feature dimensions. Our proposed work also tested using multiple classifiers and datasets and got same results which distinguish it from state of art descriptors and other classification approaches with only single classifier and limited datasets.
