Research On Depth Map Generation Algorithm Based On Image Scene Classification And Visual Attention

Three-dimensional movies(3D movies) and three-dimensional television(3D-TV)are nowadays popular among people. However, conventional 3D technologies cannot satisfy the demands of 3D film and television market for their high cost and inefficiency.In this case, researchers attempt to cope with 2D images and convert them into 3D images. The 2D-to-3D conversion process is subdivided into two segments: the extraction of depth information and the rendering of the depth images. The extraction of the depth information segment is of vital importance because its accuracy directly affects the ultimate result of the conversion.In this thesis, a depth extraction method based on scene classification and visual saliency is discussed. Via preliminary analyses, the pictures were roughly classified by the algorithm. Next, different depth extraction methods were utilized for different scenes on purpose of a better fitting. Eventually, the saliency attention method was managed to acquire the foreground regions from the images. The final depth map was obtained by fusing the depth map and the visual attention regions. This thesis follows three aspects in detail:First, we discuss the visual attention method based on the GMM model. Then by combining with the GrabCut method, we can extract the foreground regions of the image.Second, by applying the BOW model, the images in different scenes were classified and the depth information of the images were extracted. This step was initiated via classifying the images into three categories: outdoor, city and indoor. For outdoor images, the depth maps were generated by means of identifying the outdoor regions and considering the region rules. For city images, depth information was extracted through analysing the vanishing point and linear perspective clues. Whilst for indoor images,depth maps' generation was based on indoor floor recognition.Third, the image depth information extraction of complex images which cannot be identified by scene classification was implemented by studying their GIST features and mapping them to similar images. With the aid of the SIFT flow, the images weremapped on the pixel level. Finally, through the IRLS method, the depth information for the images were estimated.In this thesis,Sun Database is applied to discuss the accuracy of the scene classification algorithm. Besides, the depth information of some images was extracted in the Make3 D database, aiming at showing the results of the depth generation methods in different scenes and comparing them with the true depth information.