Research On Key Technologies Of Exposure Fusion

The luminance range of current conventional display devices are much smaller than the real-world scenes and the human perception, so it is impossible to show the real scene observed by human eyes in conventional devices. But the HDR imaging technology can achieve to display more details of real HDR scenes on common display devices and conform to the human perception as far as possible. With the rapid development of high-definition digital industry, HDR display technology has become one of the hottest topics in the digital domain. The technology not only has a very important theoretical value, but also has very large economic and commercial values.Conventional HDR display technology strategy is to recover the HDR radiance map by a series of images taken with different exposures in the same scene according to the camera response function, and then to compress the dynamic range of the radiance map by the tone mapping algorithms, but this strategy has some flaws. For example, a variety of exposure parameters are needed for input. Instead, the exposure fusion technology can directly fuse these images with different exposures into a single image that can be displayed on a common display device, without the camera response function and a variety of exposure parameters. How to preserve clear details of the bright and dark areas in the fused image without distortion and how to automatically remove ghost artifacts of exposure fusion for dynamic scenes are the hardest problems of current exposure fusion technology.We make an in-depth study of the technical characteristics and research status of HDR imaging and exposure fusion technology. We propose a new exposure fusion method for preserving details of the bright and dark areas in the fused image without distortion, and do some exploratory study on HDR image quality evaluation. Besides, in order to remove the ghost artifacts for exposure fusion of dynamic scenes, we propose an efficient ghost detection and removal algorithm based on a reference image. Our main works and contributions are as follow:1. A multi-resolution exposure fusion algorithm with weight maps correctedWe propose a Weight Modification Function to modify the basic weight map for each image, and to enhance the weight of the informative pixels in ultra-bright or ultra-dark areas. We maximize the details by minimizing the distortion probability function. Compared with previous methods, our approach preserves more useful details for scenes with very large dynamic range and minimize the distortion as far as possible. Several objective quality metrics prove the advantages of our method.2. An interactive exposure fusion method for display multi-level detailsWe propose an interactive exposure fusion method by introducing Local Laplacian Filtering (LLF) for edge-aware image processing to multi-resolution Laplacian pyramid weighted blending. Our method can not only show multi-level detail manipulation, but also selectively show the details in different regions of the original scene by adjusting a simple parameter. Compared with previous methods, our interactive method has a greater flexibility on detail display for the requirement of different users.3. An deghosting exposure fusion method for dynamic scenes based on linear photometric relationIn this paper, a new effective and high-efficiency ghost detection method is proposed. According to the existing linear photometric relation between luminance and exposure, we normalize each input image to the luminance of the selected reference exposure. Before normalization, we first replace each underexposed or overexposed pixels by the matched available exposure. Then the moving objects are detected using modified difference method and the detection binary images are modified with morphological operations. Finally, we redefine weight maps of exposure fusion according to the modified binary images to obtain the fused image. Compared with the previous deghosting methods, our method can offer high-quality deghosting result with natural textures, even if the selected reference image contains saturated regions. Our method is efficient without complicated iteration process. We demonstrate our performance advantage by comparing with previous reference based methods.