Research Of Texture Classification Based On Shearlet-Transform

Texture is a fundamental property of objects’surface in nature, and also a visual feature which is important but hard to define. As a basic research area of computer vision, texture analysis nowadays has been playing a significant role in many disciplines and applications. Being an indispensable component of texture analysis, texture classification is an enduring research topic that intrigues so many scholars both at home and abroad. As a crucial factor affecting the classification accuracy, texture features of robust distinctiveness have long been drawing the attention of researchers. In this thesis, a novel texture descriptor based on Shearlet-Transform is introduced, which has outstanding performance in multi-scale and multi-directional signal analysis. A series of experimental results show that Shearlet has a significant advantage in term of feature extraction of texture.First, this thesis introduces properties and research backgrounds of texture, and then reviews the methods of texture features extraction in four categories:statistical method, geometric method, model method and signal processing method. Moreover, the development of wavelet transform is elaborated in detail as well, together with the limitation of wavelet when analyzing multivariate data. Moreover, a brief introduction of Shearlet will be presented, including its properties and advantage. Meanwhile, an approach of discretized Shearlet-Transform in frequency domain is provided. Additionally, the properties of Shearlet-Transform are analyzed comprehensively, including directional sensitivity, reconstruction accuracy and sparsity.In this thesis, to capture the anisotropic information of texture, such as edge, contour, curve, etc., Shearlet-Transform is employed to extract texture features, and an analysis is presented about its feature computing time. Furthermore, gray level co-occurrence matrix and color moment will be computed from the Shearlet-Transform coefficient matrix. This further extracted information is fused with original Shearlet-Transform coefficient thus a new texture descriptor is obtained at last. With the statistical information in the fused descriptor, so it’s much more distinctive than ever before.Finally, a series of experiments are conducted in 10 public datasets and the method introduced in this thesis will be compared with other frequently used descriptors. The results show that texture features based on Shearlet-Transform outperforms other conventional methods, besides, when it’s fused with the statistical information of texture, this advantage will be more remarkable.