Theory And Practice On Tone Reproduction Of Color Photos

With the development of relative technology and theory, photogrammetry is no longer limited in merely "surveying", but pays more intension on "vision". As a carrier of most information from 3-d world, the processing of photos developes not only in the direction of feature information extraction and geometry reconstruction, but also in the direction of generating photos which are close to the subjective perceive as a base of further process. Different areas require differently on such object. The researches in Graphics attemp to render a real 3-d scene from geometry and physics information of the scene, its a proceed from 3-d to 2-d. While the researches in computer vision attempt to approach real cognition of 3-d scene from 2-d photos. Samely, the ordinary applications of photogrammetry is also a mapping from 2-d photo to 3-d scene. In such a process, more and more attention is payed on physical information of the scene such as surface texture structure, lightness & color appearance except traditional geometry reconstruction to approach vision cognition as real as possible at the same time of obtaining exact geometry model. Based on such phenomena, a great deal of researches are discussed in the view of constructing an illumination model. However, such researches are merely reflect the process of lights radiating from illuminants to cameras or human eyes through medium and imaging surface, which is only a physical description. The cognition on color and lightness is always affected by mental and physilogical features. At the same time, constructing a physical model means illustrating each part of the imaging process such as illuminants, media, reflectance of imaging surface with strict description, which is difficult in real scene and presents only as an experimental method. So, when the illumination model is hard to handle tone reproduction maturely, the method based on imaging model indicates a direction. With some basic suppositions, the imaging model can solve the problem of tone reproduction in some photomrammetry applications through separating basic components of illumination in the scene and simulating the human vision system. But, there are some limitations of such method. Firstly, such models surppose that the illumination varies smoothly in the scene and the scene accords to "gray world" assumption. Secondly, most imaging model methods ignore the difference on perceiving the world between the human eyes and cameras. Hence, aimed at practical applications, takes into account of such problems based on traditional computational color constancy, approach effective color and lightness reproduction is exactly the objective of the dissertation.The holistic cognition on a photo by human vision system is focused on two important aspects: lightness distribution and main tone, where the lightness distribution includes space distribution to illustrate lightness difference in spatial domain and statistcal distribution to illustrate dynamic range of image lightness. The main tone means the holistic color appearance of the scene which influenced by environment and the chroma of illuminants. In photogrammetry, different imaging systems are influenced differently by atmosphere, illuminant, camera, geometry position et al, which can easily caused the imbalance of lightnessspatial distribution, the mismatch of dynamic range between imaging equipments and real scene, holistic color cast, et al. Because such problems in different imaging scale are similar in description, request and solution, all can be looked as reproduction of lightness spatial distribution and color appearance, the process is called tone reproduction in the dissertation.The concrete objective of tone reproduction is different when the imaging scales or acquiring methods are different. Aimed at remote sense images and aero photos, the main consideration is on the lightness and color spatial distribution, color matching between adjacent photos. Such requests are usually in dodging and mosaicking in photogrammetry applications. In a direct view, a kind of simple, direct, high efficiency method which is named adaptive template based on local means is suggested to realize the dodging of remote sense images and aero photos. The experiments indicate that such method can approach the basic objective of dodging, but it can decrease the local contrast which can lose some detail of the images. The further analysis on the local mean template indicates that such method is a kind of method based on imaging model, which trates the curve facet fitted by LSM of local means as a kind of low-pass filter. Through the analysis on low pass filter and introducing the idea of "scale space", a multi-scale imaging model method is suggested, which can dodge the lightness and color by synthetizing filtering result with small, moderate and large scales to improve the detail of the photos at the same time. A lot of experiments indicate that most photos accord with the assumption of "gray world" can be dodged and improved by such method. At the same time, aimed at the difference of main tone between adjacent photos and holistic color restriction on dodging result which is processed by three channels separately, the Wallis algorithm which usually focuses on local process and inage enchancement is developed as a global algorithm, which can realize the matchof main tone between photo sequence and holistic color restriction on single photo. The dodging and mosaicking experiment system is developed based on such method, and a great deal of experiments validates its effect. What's more, the comparison between multi-scale imaging model and multi-scale retinex indicates that in the machine vision applications based on low dynamic range images, the further non-linear attenuation of input signal should be avoided to avoid the loss of detail information.For the close range photos with smaller imaging scale, the most obvious limitation for imaging model to reproduce tone of photos lies in the violation of two basic presupposition, which means the "gray world" assumption and illumination component in the scene varies smoothly. Secondly, for the photo with simple texture, the Gauss filter in the imaging model will form the reversion of local contrast near clear edges, which will cause artifacts like "halo". Based on above considerations, aimed at close range photos, a relatively integrated framework is suggested, which discusses each part of the processing, which includes holistic color cast removal, separation of illumination component and reflectance component, reproduction of reflectance component and illumination component. The framework can realize tone reproduction of such kinds of photos effectively.In the holistic color cast removal, based on diagonal model suggested by von Kries and "gray world" supposition, method on holistic color cast judgement and removal is discussed and the approach based on gray balance and white balance is suggested for the scene based on different reflection surfaces in the scene. Such approach can remove the holistic color cast effectively.In the separation of illumination component and reflectance component, bilateral filter is introduced which can constrain the edge diffusion in the process of low-pass filtering as an anisotropic filter to avoid the appearance of "halo" artifact. The relation between edge-stop parameter of bilateral filter and window size of filter is discussed simply.For the reproduction of reflectance component, because it's a product of high-pass filtering, the noise will be exaggerated. Furthurly, although such method can avoid the influence of specular reflectance, the influence of shadow is inevitable. To contrain the noise and shadow edges, an approach based on controlling edge-stop parameter of bilateral filter through definition descriptor is suggested. Such method can obtain a reflectance component with smooth local area and weak shadow edges.In the reproduction of illumination component, based on the characteristic of data set, a region growing method based on bilateral filter and double criterion on local area is suggested firstly, associates with binary close operation, each kind of texture area can be segmented rapidly and robustly, which can avoid the influence of shadow edges simultaneously. Such method can be extended to similar region segmentation in other applications. Secondly, based on the difference of achromatic and chromatic color, the color statistic is performed with different strategy. Finaly, the color based on the statistic data or interactive input is rendered in current region.The theory and experiments of each part indicates that the framework for close range photos can realize tone reproduction effectively. It provides a reference method for tone reproduction of real scene which violates the basic presuppositions of imaging model.Except the tone reproduction introduced above, with the development of technology on hardware and software, high dynamic range photos is more and more comprehensive in the photogrammetry applications. Subsequently to the review of dynamic range compression research, a multi-scale bilateral filtering method is suggested. Through analysizing the influence on gauss filrer by the parameter of edge-stop function, the bilateral filter that be viewed as a filter has no concern with filter size for a long time is extended to a multi-scale method. Furthurly, the edge-stop function is rearranged as a non-linear function reflect the contrast character of the vision system in the intensity domain. The function can be looked as a descriptor of local adaptability to provide vision theory proof for our method together with the multi-scale retinex theory. Experiments on 12-bit remote sensing photos, aero photos, and HDR images indicate that the method is better than other methods in content integrality and detail representation ability based on the data set provided.As an entire digital image processing technology, it's necessary to evaluate the experiment results to validate the techniques. Aimed at the request of tone reproduction, quantitative indicators on the degree of uniformity of the lightness and color spatial distribution and the degree of color cast are discussed primarily. For the former, two quantitative indicators are suggested, which are the variance of lightness template and shape descriptor of binary sequence. The two indicators both can be used as absolute indicator and relative indicator. For the latter, color difference in relative even color space is calculated based on gray balance and white balance as a absolute and relative quantitative indicator for difference of main tone and color cast.Aimed at the limitation of imaging model, the difference between human vision system and machine vision system, solutions for three kinds data sets and quantitative quality evaluation framework are suggested. Experiments indicate that the methods and indicators can realize tone reproduction of different applications effectively and evaluate the result quantitatively. When the physical model is still not able to realize tone reproduction in such applications conveniently, the technology suggested here can approach the objective of tone reproduction in photogrammetry applications as a relative method effectively, which provides a referencable, practicable solution and promote the relative technology in a certain extent for vision technology in photogrammetry applications.