Rotation-invariant and spatial orderless deep feature representation for texture image classification

Abstract

therefore, a texture descriptor should encode those patterns regardless of their locations. Hence, in this thesis, we propose a new network architecture that is robust to the orderless property when encoding texture patterns. The key to our proposed method is a histogram pooling layer that pools convolutional neural network features in an orderless manner. To summarize, we have studied handcraft and deep feature based texture classification and the deep feature gave better performance compared to the handcraft feature. However, we believe that the findings of the handcraft feature have gave motivations and clues to the deep feature approach. Moreover, both handcraft and deep feature approaches will influence each other for the texture classification area.

In this thesis we propose texture classification method by two approaches: rotation-invariant handcraft feature and spatial orderless deep feature. The rotation-invariant handcraft approach has been a long-standing research issue for texture classification problem. Since the texture images include arbitrary rotational changes in every spatial positions so that invariance to the rotational change is most important property. However, despite of the importance of the rotation-invariant, the encoding process to consider the rotation-invariant property has caused some limitations including loss of some patterns and locality information. Hence, in this thesis, we tackle the limitation of the conventional methods by suggesting new texture feature descriptor. We first propose a sorted consecutive local binary pattern (LBP) for texture classification to resolve the problem with regard to the loss of some patterns. Second, we present a locality-preserving descriptor that encodes rotationinvariant features while retaining locality information. The deep feature approach has also studied to use deep neural network for the spatial orderless property. The deep feature has showed highly improved performance compared to handcraft feature in most domain of computer vision area. However, the texture classification area has certain characteristics that existing pre-trained models do not consider: specifically, the ‘orderless’ property of texture images. A texture image contains distinctive patterns that appear at arbitrary positions