Feature-Preserving Image Transformations In Different Domains

Accompanied with the development of electric equipments, the problems solved by image processing and the applications developed become deeper and wider. Researchers are interested not only in2dimensional signals, but also3dimensional ones, because of the tight connection between image processing and mesh processing. This thesis endeavors to put the current topics into an ensemble:Image Transformation. Image transformation can be divided into four parts:equal-dimension spatial transformation, various-dimension spatial transformation, equal-dimension value transformation and various-dimension val-ue transformation. Spatial transformation means geometrical change during the process-ing while value transformation means the change of pixel values. Meanwhile, the equal-dimension transformation produces changes only in one dimension, and various-dimension transformation maps signals into a different dimension, such as2D and3D spatial space, or1D grey and3D color space. Furthermore, feature preservation is another hot research topic partly due to the growth of saliency map. Several questions like finding the connec-tion of feature in different research fields, the definition of feature according to processing requirement, finding an appropriate operation to detect the feature, and the preservation of feature during the transformation are the most difficult parts of feature-preserved prob-lems. This thesis introduces three algorithms in two fields of image transformation, which discuss the definition of image feature and illustrate how they are adopted to ameliorate images.The second chapter is mainly related to the various-dimension value transformation: color-to-gray conversion of color images. This topic has a huge demand for the services in black-and-white print, image algorithm pre-processing and black-and-white photography. Taking the luminance channel as the grey result is the simplest way. And then some global and local methods are invented focusing on the protection of important gradient or salient region in color images. We analyze the incompatibility of global and local target and then propose two methods both in global and local perspectives. For the global part, based on some perceived intensity experiments, we present a global nonlinear tri-mapping which could assure mapping consistent from color to grayscale. This mapping fully considers the hue, luminance, and saturation information, and further preserves feature discriminability of color gradient by an optimized tri-mapping. For the local part, we capture the perceived appearance of color images in our color-to-gray conversion. Based on the Filter Theory, a novel measurement of channel-level distinction is formulated, Channel Salience, to depict the filter level of three color stimuli. This salience metric guides a contrast adjustment process to enhance the perceived grayscale in the final output with a two-steps conversion which elaborates the grayscale on selected points first and then propagates to pixels all over the images.The third chapter discusses a reversed problem, colorization of grey images. Image colorization is a process of adding color to a monochrome image or movie, which is an active and still challenging research topic. One of the applications is an improvement in the appearance of black-and-white cartoon images. Conventional methods for natural images begin with color transfer and luminance continuity optimization framework. In the carton cases, segmentation seems to be necessary. We propose a novel approach for colorizing black-and-white cartoon images and preserving their tone properties. First, we use an edge preserving decomposition method to compute a tone map. In order to prevent boundary leakage, we compute a pattern energy term by using the pattern feature vector. Later, the tone map and the pattern map are both applied to an optimization algorithm. Only a few of color strokes are required to be specified on the input image by the user and segmentation of the image is not required.Finally, the fourth chapter belongs to equal-dimension spatial transformation. For the sake of displaying high resolution images on a limited screen, image retargeting is getting more attention among academic and industrial communities in recent years. The former methods mainly used gradient or salience map to describe important information. Recent researches put repetitive patterns into consideration but random textures are not included. A new image retargeting method is built upon the image redundancy analysis which reduces the sole dependence upon salience map. We introduce an image decomposition frame-work to separate redundant information from an image, to avoid the disturbance from high gradient redundant regions. Then we define a texture gradient by analyzing the texture re-dundancy which indicates the detail richness and the importance of a region. We also show that the proposed algorithm is suitable for texture-rich natural images, especially cartoon images.