New Feature Extraction Method for Scene Depth Level Classification

Abstract

Eye fatigue caused by 3D contents having unnatural scene depth level assigned by conventional 2D to 3D conversion system has been issued as one of the worst side effect of 3D contents. For generation of more realistic 3D contents to overcome eye fatigue and other side effects, it is need to study depth control method which has not been studied enough. In this paper, new feature extraction method for image scene depth level classification are introduced. Based on natural phenomenon, we found a direct relation between natural image statistics, and depth pattern, and proposed new texture related features for scene depth level classification. And, to overcome the effect of illumination color on color related feature, we suggested color feature extraction method with illumination color estimation based on conditional color temperature adjustment. Proposed features represent scene depth level efficiently. Finally, proposed features and existed features for indoor-outdoor classification are concatenated to generate the feature vectors and fed into the SVM classifier for the scene depth level classification. To justify the efficiency and robustness of the proposed method, the evaluation is conducted over 600 images.