Research On Stereoscopic Visual Attention Model For 3D Video With Machine Learning

Visual saliency is a fundamental problem in both computational science and image processing. Research on the visual saliency of videos and images can help us understand human's visual attention. And it can be applied to image cropping, video compression and seam carving. With years of effort, researchers have achieved many important goals in 2D visual saliency research. However, the rapid development of stereoscopic display techniques along with the emerging 3D applications brings the depth information which makes the viewer's visual attention become a little different.Researchers are always trying to get a model which can get the visual saliency precisely. Cognitive theory(feature integrate, guided search) shows that the key to forming a good visual saliency model is feature extraction and feature fusion. This paper extracts the global features like DCT features, Itti features, subband features, color space features, motion feature, depth feature and centre bias feature. And it also extracts the local saliency feature with the help of CNN. After feature extraction, we analysis all the features and choose different feature combination for different fusion method. This paper uses two fusion methods, the support vector machine and the convolutional neural network which we remove the subsample layer to improve the learning efficiency.We compare the proposed models with several other stereoscopic visual attention models with direct evaluation, ROC and P-R curve, AUC, F-measure, PLCC and KLD. Experimental results demonstrate that the proposed models are efficient and reasonable, and are also robust for different scenes.