Research On Key Technologies Of Intelligent Stereoscopic Image Processing

Since 2009 we have witnessed the flourishing of 3D movies represented by A-vatar. Consequently, the industries like stereo camera and 3DTV have experienced a rapid growth. The research for stereoscopic media has also been boosted. Stereo im-age processing, which manipulates binocular stereoscopic images, plays the key role to support and lead the development of stereoscopic industries. To improve the effect and efficiency of stereo image processing, binocular characteristics needs to be fully utilized. Compared to monocular scenarios, the key for stereo image processing lies in the exploitation of binocular correlation. Besides, a combination of traditional image processing methods with binocular correlation could lead to better processing result-s. According to the application requirements, we make an in-depth investigation to the characteristics of stereo images and conduct our research on depth acquirement, saliency analysis and interactive object segmentation for stereo images. The novelty and major contributions of this thesis are as follows.1. A novel local stereo matching method based on color and correlation, which effectively solves the accuracy and robustness problem of current method-s caused by over-reliance on color distribution. Current local methods are mainly based on color similarity. The basic assumption relies heavily on the consistency of color and depth distribution, which cannot be always satisfied in real applications. Af-ter studying the correlation between views, we make a combination of correlation and color cues. The proposed method improves matching accuracy compared to current methods, and shows better robustness in real applications.2. A novel saliency analysis method based on stereopsis, which outperforms state-of-the-art methods. Most previous methods work on 2D images. The power of depth information has not yet been effectively exploited. Besides, the lack of dataset has limited the research on depth-aware saliency analysis. We focus on how to employ the implicit depth cue from stereopsis to detect salient objects. By investigating the characteristics of depth cue and the spatial structure of salient objects, we propose a novel method based on anisotropic center-surround difference. We further combine the depth-aware method with current 2D saliency methods. The experimental results show that our method outperforms state-of-the-art by 15% on detection accuracy and F-measure. Furthermore, considering the lack of depth-aware saliency datasets, we build and open the largest evaluation benchmark on depth-aware salient object detection for related research.3. A contour based consistent image segmentation method, which substan-tially improves the computational efficiency. Most of the consistent image segmen-tation methods are based on region, which neglect the view similarity and thus lead to high computational redundancy. To overcome the problem, we propose a contour based consistent segmentation method for stereo images. The efficiency of our method is 10 times faster and has a better consistent segmentation accuracy than state-of-the-art. Besides, our method solves the consistent segmentation problem independently, which enables our method to be easily combined with any single view methods to handle stereo scenarios.4. An object segmentation method based on adaptive color and depth fusion, which significantly improves the accuracy of segmentation results. Current image segmentation methods are mainly based on color information. However, the character-istics of the depth cue have not been fully utilized. By investigating the characteristics of depth maps, we propose an object classification model based on geodesic distance. Furthermore, we make an adaptive fusion of color and depth cues according to their advantages in segmentation. The evaluations show that the proposed method makes an improvement of 2% on F-measure, and the precision and recall both surpass 95%.Based on the above research, we give a few applications of intelligent stereo image processing. The examples show that our research plays a well supporting role and has a favorable application prosect.