Study On The Vector Filters And Hypercomplex Matching Appraoch For Color Images

With the increasing popularity of all kinds of color image application, it is more and more important to improve its processing quality and computational efficiency in the field of image processing currently. Compared with the monochromatic-based techniques in which the pixels are separated and combined, vector or hypercomplex approach for color image describe the intrinsic relationship of different channels and illustrate strong application and research significance.Due to the noise affection during the acquisition or transmission process, the removal of such noise is often an essential pre-processing step for many digital image systems and its processing quality directly affects the visual performance and recognition accuracy. The classical vector median filter and its extensions avoid the artifacts when removing the multichannel impulsive noise, whereas it have the shortcoming such as low computational efficiency and bad detail preserving ability. Meanwhile the hypercomplex Fourier transform extends the processing signal field compared with the classical Fourier transform. As a new theory in color image processing field, its application range should be extended. The main contributions of this thesis are as follows:1. A new vector median filter based on the central pixel's rank value in filtering window is presented based on the relation analysis between the rank value and the noise ratio for color images. On the basis of the variational condition of rank value, a new switching vector filter with ranked thresholds is also suggested. Through the training process on the specific image database using the image quality metrics, the ranked parameters are optimized. Simulation experiments show that the new scheme can suppress the impulsive noise effectively and preserve the details and edges of original image concurrently.2. To improve the computation speed of the classic vector median filter and their extensions, two new fast algorithms are proposed to fast obtain the vector distance matrix or fast peer group in filtering window respectively. By analyzing the relation of current pixel's and neighbors'filtering windows, the new vector distance computation methods are redefined to save the whole computation. At the same time the new fuzzy peer group of each image pixel is determined by means of a novel fuzzy logic-based procedure to achieve better processed image quality.3. Considering the condition that Gaussian and impulsive mixed noise exist during acquisition or transmission process concurrently, we propose a bilateral vector filer based on texture analysis. The new method carry out noise detection, texture information analysis and bilateral filter with clustering results and show the ability to suppress the two kind of noise simultaneously.4. To estimate the parameters of translation, rotation and scale transform of two color images, the paper proposes a new approach based on hypercomplex Fourier-Mellin transform. The hypercomplex is utilized to present the color component of color image and avoid the hue loss when using the classical Fourier-Mellin transform registration technique. Several equivalent formulas are presented to compute the phase correlation between only-translated color images and the integrated algorithm is given to estimate the translation, rotation and scale parameters. Finally the experiments using simulated and real images show that the new algorithm is more accurate compared with the classical Fourier-Mellin transform approach.