Image-decomposition Based Tone-mapping Algorithm For High Dynamic Range Images

High dynamic range images, which record the physical value of luminance or radiance in the real scene that can be perceived ranging from 10-6cd/m2 to 108cd/m2, provide more abundant details and more real perception. However, common output devices such as CRT or LCD displays and printers have very limited dynamic range(that actually is about ranging from 1cd/m2 to100cd/m2), that results HDR scene must be compressed significantly in order to be displayed. Tone mapping algorithms was proposed for this purpose, of which, algorithms based on image-decomposition are deemed to the most promising due to their great superiority on contrast reproduction and halo elimination.This dissertation focuses on image-decomposition based tone-mapping algorithms for HDR images, and three improved algorithms are proposed. More details are as follows:An image-decomposition based tone-mapping algorithm aimed at detail-preserving is proposed. In this algorithm, in order to solve problems of over-compression by a non-adaptive mapping function and changes of perceived contrasts for luminance shift during mapping, the luminance-response curve adapting to each local luminance in high dynamic range images, as a mapping function, is used to map luminance of the base layer. Then, compensation coefficients of the detail layer, for stretching or compressing details, are computed according to values of luminance shift based on Stevens’ effect. The experimental results show that the proposed algorithm has good performance on preserving perceived details.An improved decomposition algorithm is proposed for eliminating halo or gradient reversal artifacts, which appear on salient edges during HDR images tone-mapping based on improper image-decomposition. In this algorithm, firstly by analyzing base layer image and detail layer image obtained via image decomposition, improper signal characteristics which cause halo and gradient reversal artifacts are confirmed. Then according to the improper signals, the artifacts are recognized and corrected in the gradient term of detail layer image, and corresponding image regularization equation is established based on the corrected gradient subsequently. At last, conjugate gradient method is adopted to solve the regularization equation. The experimental results show that the proposed algorithm has good performance on eliminating halo and gradient reversal artifacts and preserves visible details as well as original algorithms do.A new image-decomposition algorithm based on variational method is proposed for decomposing HDR images more improperly, so that avoiding halo or gradient artifacts caused by edge blur or sharpening in base-layer-images which occurs in many existed algorithm. This algorithm is based on generalized TV image-decomposition model with selecting the Tukey’s bi-weight function as potential function to build regularization terms of the variational model. And then, numerical method is given for solving this variational model, including Graduated Non-Convex algorithm for global optimizations and half-quadratic algorithm based on alternate minimizations strategy for local optimizations. The experimental results show that the decomposed base-layer-images obtained by proposed algorithm has the same sharpness as the original images and avoids halo and gradient reversal artifacts totally.The main innovations of this dissertation include: the problem of perceived contrasts shift due to luminance changing is solved by establishing detail-layer adjustment model; and the halo and gradient reversal artifacts are eliminated successfully from the perspective of image decomposition.