A Study Of Texture Classification Based On GMM Matching

With rapid development of artificial intelligence and computer technology, texture classification has been playing a crucial role in medical treatment, satellite, robots and so on, which exists everywhere in our life. Texture classification has attracted a lot of attentions as a hot topic of image classification in computer vision for a long time. Different from generic images classification problem, texture classification faces more specific and fine-grained images, which results in smaller inter-class difference. Meanwhile, images from same class suffer from drastic changes due to variations in scales, viewpoint and illumination. Both above reasons make the design of robust texture classification method being very challenging.As Gaussian Mixture Model (GMM) has strong modeling capability, it is widely used for representing images, especially for texture images. GMM matching based methods have shown their superiority in texture classification. Such methods include two main components: feature extracting and GMM matching. It is well known that image features play a key role in image classification. With great success of deep learning, deep convolutional neural networks (CNN) is successfully used in image classification, image retrieval and semantic segmentation. The existing GMM based image modeling methods just focus on classical handcrafted features. Inspired by the success of deep CNN, this paper makes the first attempt to propose the CNN features based GMM modeling method. Meanwhile, this thesis makes a comprehensive evaluation of different features, which include two kinds of high level CNN features and five kinds of classical handcrafted features widely used in texture classification task, i.e., patch, Gabor filter, LBP, SIFT and covariance descriptors. Note that a comprehensive evaluation of various features in GMM modeling based texture classification has not been well studied. For GMMs matching, this paper employs the efficient and robust Sparse Representation based Earth Mover's Distance (SR-EMD). In computation of SR-EMD, the ground distance between Gaussian components plays a very important role, which however, is still a difficult and open problem. The existing ground distances are first roughly divided into statistics based ones and Riemannian manifold based ones in this thesis. Based on the analysis of relationship between different ground distances and inspired by Gaussian embedding distance (GE) and product of Lie Group distance (PLG), this thesis proposes an improved Gaussian embedding distance (IGE), which not only improves the classification performance, but also verifies some important conclusions.The experiments are conducted on three challenging texture databases:KTH-TIPS-2b, FMD and UIUC Material. The results show that the CNN features based GMM is very suitable for image representation. Moreover, combination of the proposed ground distance with the CNN features based GMM can obtain very competitive results on all three texture datasets. In addition, the proposed CNN features based GMM matching method has a good generalization ability, and it can be naturally applied to image retrieval. On both two popular datasets, i.e., INRIA Holidays and UKB, the proposed method achieves the state-of-the-art results.