Recognition of Dynamic Texture Patterns Using CHLAC Features and Linear Regression

In this paper, we propose a statistical scheme for recognizing three-dimensional textures shown in motion images, e.g., ultrasound imaging, which we call “dynamic textures”. The texture characteristics emerge as the distinct movement in the motion images, and the dynamic cues would be useful especially for recognizing ambiguous texture patterns in noisy images. We apply cubic higher-order auto-correlation (CHLAC) to extract features both of the textures and their movements, and then linear regression to evaluate (recognize) the texture. For the linear regression, we extend ordinary multiple regression analysis so as to reduce within-class fluctuations. In the experiment for estimating quality of beef meat by using ultrasound motion images, the proposed method exhibits the favorable performances which are close to ground truth given by the experts.