Image Enhancement And Analysis For Street Views

Street Views is more attractive than 2D maps.Users can enjoy beautiful sceneries without the need to going out.It is a useful application,as users can browse the famous landmarks just in front of a computer,or preview the nearby scenes of the destination.Street Views is composed of 2D panorama and 3D point clouds,captured by moving vehicles every few metres.There are several post processing steps such as stitching,color correction and detail enhancement.Recently,autonomous driving is a hot topic in industry,and high resolution map is a very important part of the autonomous driving system.Useful information can be extracted from the Street View images.Thus,image enhancement and analysis are important techniques in Street Views,as they are the key to the user experience of the Street View application and important information source for autonomous driving.Compared with traditional images,images in Street Views have much larger resolution and suffer from distortion.Due to these properties,traditional image processing techniques can not be directly used for Street View images.This thesis focuses on image enhancement and analysis for Street Views,mainly the unsolved problems of image completion for Street Views,content analysis of Street View images.The main contributions of this thesis are:1.A faithful image completion approach.It can automatically find similar candidate images for the target image,and warp the candidate image to the target image,matching the key points and straight lines.While traditional image completion approaches aim at visually pleasing completion results,our approach can generate factually correct ones.When applied to Street View images,our approach can be used as a panorama updating technique which can avoid re-capturing.2.A panorama completion approach for Street Views.An algorithm to complete for panoramas of 360 degrees was proposed.It considers the distortion that exists in panoramas while traditional image completion can not handle.To tackle the distortion,a structure rectifying warp that can simultaneously preserve local shapes and straight lines was proposed.Completion can be performed on the transformed image via the proposed warp.The completed result is further transformed back to the original,and then final completion result is obtained.3.An optimization approach for traffic sign candidate localisation.An localisation refinement approach for candidate traffic signs was proposed.Previous traffic sign localisation approaches which place a bounding rectangle around the sign do not always give a compact bounding box,making the subsequent classification task more difficult.The localisation refinement was formulated as a segmentation problem,and prior knowledge concerning color and shape of traffic signs were incorporated.4.An simultaneously detection and classification approach for traffic signs in the wild.It can handle small traffic signs in large images,and has higher recall and accuracy compared with state-of-the-art object detection approaches.