High Dynamic Range Image Tone Mapping Using Intensity-Layer Decomposition And Energy Optimization

Dynamic range is the ratio between the maximum and minimum intensity value in an image or natural scene. Traditional image formats store each color channel in eight bits as an integer value, and thus have a dynamic range of 100:1. However, the dynamic range of real world far exceeds this area, so natural scenes can not accurately described via traditional image formats. To overcome, high dynamic range image is proposed using floating point numbers to represent color values, which has a much higher dynamic range. Tone mapping is the technique that converts high dynamic range images to low dynamic range, making them able to be displayed in current output devices such as monitors and printers.High dynamic range image was originally used to render scenes in computer games, while other uses also emerged. Some consumer devices already provide high dynamic range imaging, and multi-exposure synthesis is supported by many software. To fully exhibit the advantage of high dynamic range, good tone mapping is necessary. In this paper we propose tone mapping algorithms that utilize intensity-layer decomposition based filtering and energy optimization. Tone mapping can be completed automatically, while interactive adjustment is also supported to achieve subjectively satisfying results. The main contributions of this paper are as follows:1. We proposed a fast trilateral filter using intensity-layer decomposition. The proposed approach accelerates the gradient filtering stage of trilateral filter. We decompose an image by pixel intensity values into several layers. For each layer, we compute two intermediate images, and the gradient filtering is achieved by Gaussian filtering on these intermediate images. The experiment results shows that , our method is about 10 times faster than the original one in the application of tone mapping, and the visual quality of obtained low dynamic images are quite similar.2. An interactive tonal adjustment method is presented. An energy function is defined to interpret user interaction, which take into consideration the affinities between each pair of image pixels. Affinity matrix, which consists of affinity values, is close to low rank, and a fast solution for the energy function is obtained via low rank approximation for the affinity matrix. This approach does not require precise user interaction, while the results are accurate and edge-preserving. We discuss the bleeding problem and give a practical solution. Diffusion maps and diffusion distances are introduced, which can also be computed using low rank approximation. An experiment shows that, diffusion distance can provide better edge-preserving than Euclidean distance. An intuitive interpretation is shown, based on which the possibility of finding image features making use of image segmentation is discussed.3. We design an effective work flow for, which enables users to add or cancel interaction in real-time manner, while a lot of intermediate results are not stored.