Depth Maps Generation Algorithm Based On Machine Learning

As the development of display techniques, a great step has been made in the improvement of 3D display equipment, which makes realistic impressions for 3D images and videos. However, as the complex and long-periodic producing procedure of 3D filming techniques, which is also very expensive at the same time, 3D resources are quite deficient in present. Estimating depth information of 2D image using computer vision and related image processing algorithms, converting it to 3D, can not only solve the requirement of 3D resources, but provide people a way to review former pictures and videos in 3D form. Depth maps generation based on machine learning has advantages in adaption, which makes no restriction in the types of scene. It can produce depth maps with high quality which are able to represent the depth changes of targets properly, has become a key research area of depth maps generation. This paper mainly focus on how to improve the speed and quality of depth maps generation with learning-based methods. The major work can be sumarized as follows:1. After studying the depth maps generation based on kNN searching algorithm and cross bilateral filtering(CBF), we proposed a depth maps generation algorithm based on scenes clustering and guided image filtering. The proposed algorithm clusters all images in datasets to classes with some certain structural features, which can turn the one-by-one similar images searching to the searching of similar scene, speeding up searching process greatly. On the other side, we adopted guided filtering with higher speed to replace CBF, which can shorten elapsed time used in depth maps filtering. From the simulation experiments, we can notice that both methods produced depth maps with similar quality, but compared to the method which based on kNN searching and CBF, our method has improved the efficiency of depth maps generation.2. In this paper, we proposed a training method of visual dictionary for depth maps generation. For the application of depth maps generation specially, we first took a research on the primary concepts of visual words, and confirmed three properties for them: representative, discriminative and informative. After that, we designed a training algorithm for visual words which included steps like random sampling, initialized clustering, cross validation, etc. Experiments results show that our training algorithm can mine visual words which represent some basic structures from dataset of depth images.3. Based on the work introduced above, we proposed a depth maps generation based on visual dictionary. This method detects visual word on multiple resolution of target image, and recovers depth map of target using the detecting results. Experiments show that the quality of depth maps generated by visual dictionary-based method exceeds the traditional methods based on depth cues or images searching.4. With all the research result of depth maps generation in this paper, we developed a 2D-3D conversion experiments platform with plenty functions, which can complete various tasks include image preprocessing, depth maps generation, 3D conversion, image filtering, etc.