Research On High Dynamic Range Scenes Visualization

With the development of computer and multimedia technology, medical imaging, video surveillance, satellite remote sensing and other various fields such as computer vision ask high requirement for image quality. High quality images can provide more abundant information and more real visual perception to human eyes. It is the basis of many practical applications. How to display high dynamic range (HDR) scenes on common devices is a difficult problem to hinder these applications. In order to solve this problem, this dissertation focuses on the visualization of HDR scenes, and carries out research from two aspects:tone mapping and exposure fusion.The final receiving object for displayed images is human visual system (HVS), therefore, using the HVS characteristics to guide the dynamic rang compression for HDR images has a great significance. According to this conclusion, we introduce a lightness perception theory called "double anchoring" into the tone mapping algorithm. The effectiveness of our proposed algorithm is verified by means of an evaluation standard called HDR visual difference predictor (HDR-VDP). It can keep more details and the resulting images also accord with the perception of HVS to the real scenes.We know that the surface of object has low dynamic range; the illumination intensity of scenes usually has large dynamic range. According to this phenomenon, researchers present a new thinking for HDR images:decomposing the HDR images into reflection and illumination layers. Because each layer has different information and dynamic range, carrying out different approach can achieve the compression of dynamic range, and maintain more detail information. Based on this theory, we propose a tone mapping method, which is carried out using YUV color space instead of the commonly used RGB color space. Bilateral filter is used to decompose the HDR images into different layers in the logarithmic domain with adaptive base value. In addition, we propose an improved center/surround method to maintain more details in the dark and highlight regions.For most exposure fusion algorithms in the spatial domain, single image feature is usually served as a standard to evaluate the quality of the image. However, the single feature is not comprehensive. In order to overcome this problem, we propose an exposure fusion method, which is based on support vector regression. It can establish the mapping relation between multiple image features and the standard of the image. On the other hand, existing exposure fusion methods usually treat the R, G and B channels separately, which are very time consuming. In this dissertation, we present a novel way to get more chrominance information of the scene.In frequency domain, most exposure fusion algorithms are put forward on the assumption that the source images are aligned prior to fusion. As a result, some artifacts, such as haloing, may be caused due to the slight misalignment in the source images. In order to reduce the influence induced by the misalignment, a novel shift-invariant and rotation-invariant steerable pyramid-based exposure fusion (called SPBEF) algorithm is proposed. In addition, fusion is performed in a hierarchical fashion. It can increase the contrast of resulting images. Experiments show that SPBEF can give comparative or even better results compared to other exposure fusion algorithms.