Image Texture Analysis and Feature Extraction Using Multi-scale Decomposition and Supervised Learning

Book excerpt: « Texture analysis is an active research area in image processing and computer vision. Analyzing images with powerful feature extraction methods can lead to the successful design and implementation of machine intelligence applications such as content-based image retrieval, image classification, object detection, image segmentation, face recognition, abnormality detection, etc. In this thesis, we address the issue of texture analysis and discrimination with a new methodology based on parametric statistical modeling of multi-scale image representations. A novel multi-scale image decomposition, named RCT-Plus, is proposed. It is a variant of the contourlet transform that is redundant, rich in directional information, and applicable to grayscale and color texture images. We also propose a hybrid approach for modeling texture data in the multi-scale space by a combination of suitable parametric statistical models such as Generalized Gaussian Distribution (GGD) and multivariate Gaussian Mixture Model (GMM). This approach along with adapted similarity metrics resulted in the development of new feature extraction methods that capture relevant texture information, provide highly compact features, allow for a joint exploitation of texture and color texture features and enhance texture discrimination in applications such as content-based image retrieval (CBIR) in texture datasets and abnormality detection in dermoscopic images of human skin tissue. Furthermore, supervised machine learning algorithms (KNN and SVM) are integrated into the processing system as key techniques of feature learning and multi-class classification to infer texture types on the extracted features and achieve improved performance in terms of texture discrimination. Various experimental setups are conducted using six well-known texture datasets. We successfully increased the image retrieval rate up to 97.10% for the Stex dataset while the size of the feature vector is reduced to only 67 elements. In the case of abnormality detection, moving from grayscale texture features to joint color texture features improved the Precision of detection by up to 21% in the ISIC-42 dataset. A comparison with state-of-the-art methods, including deep learning, showed that our proposed texture feature extraction methodology yields more successful results. »--Page 15.