Characteristics-Based Image Classification

As an exciting branch of cognitive and computer science, digital image processing has attracted more attentions and made great progress during the decades. As the demand on visual information processing increases, it becomes more active of the research on developing algorithms applicable to various fields and tasks. Though great efforts have been made on it, few of them have been performed on evaluating the performance of difference algorithms. This mainly because that less emphasis is put on investigating the object of processing algorithms, that is an image, further no scientific and objective description on an image exists. Consequently, the research on image processing is striving on a "spontaneous" way in which work is conducted blindly, instead of a "conscious" way.The research in this thesis concentrates on intrinsic characteristics based image classification, that is, aims at systematically classifying images by way of defining an objective metric to describe the intrinsic characteristics of images related to human vision perception. The motivation of the research lies in: 1) only if objects of processing is categorized in a mathematical sense, would it be possible for predicting the results of processing algorithms. It is required for a scientific theory system to provide an evaluation criterion and methodology for processing algorithms. 2) the research on an objective metric for instinct characteristics of images makes it possible for relaxing the impractical assumptions and for developing adaptive processing systems, which adaptively choose appropriate algorithms for a specific application and tune the parameters of the algorithms. 3) a quantitative and objective metric for image assessment would midwifery scientific methodology and rules for algorithm evaluation. "Little progress will be made, unless the scientific assessment metric, systematic performance evaluation, and the ability of re-employed for algorithms have developed". In one word, the research on the evaluation (assessment) metric will not only improve the performance of the existing algorithms, but also direct thedevelopment of new algorithms.Concentrating on the three tasks of image processing, namely image enhancement and restoration, image compression, and image segmentation, this thesis attempts to seek, investigate, and probe into intrinsic characteristics of images. Moreover, a formulation is constructed for describing the characteristics, and then applied to categorize images according to the characteristics for specific tasks. In this thesis, the intrinsic characteristics is termed as those of images related to developing and evaluating adaptive algorithms, and those reflecting human perception to images. The contributions of the thesis are as follows:A characteristics-based image model and a formulation for image characteristics are proposed. Derived from the methodology of Extenics, a basic image element, termed VISXEL is defined. Motivated by the hierarchy of human vision, Visxel is hierarchized as pixel-Visxel, region Vixel, and image Vixel. The property of Vixel and the relationship of Visxels between different layers are analyzed. It is the formulation of image characteristics unifies any kinds of images into one framework, which facilitates to study and categorize in further.For image compression, a concept to measure and classify images by use of edge information in images is proposed, and an algorithm to implement the concept is also presented. The distribution of wavelet high-frequency coefficients in images is exploited while edge active measure is defined to describe the spatial redundancy of images in this algorithm. Thus the classification of images with various content is achieved by the value of EAM, and the results of compression algorithms can be forecasted based the classification. Additionally, the dependency between EZW algorithm and the optimal level for wavelet decomposition is studied. The experiments show that the results of classification algorithm and the forecasting results make sense, and these results are accordant to human vision perception. Prior to employing EZW algorithm, an optimal value of wavelet decomposition level can be chosen according to the class of the image to be processed and the desired compression ratio.For image denoising, a wavelet based algorithm that identifies the type of imagenoise is proposed. The algorithm is effective on discriminating Gaussian and Salt&Pepper noise. Based on the work of identifying noise type, two algorithms are respectively proposed to estimate the noise level of the two types of noise. One achieves estimating the density of Salt&Pepper noise, the other is capable of estimating the level of Gaussian noise with small variance, which can hardly be estimated accurately by the conventional methods. By use of the proposed methodology, the noise that degrades an image can be measured completely and objectively without any prior knowledge. Consequently, the denoising algorithms can be implemented free of any subjective assumptions about the noise.For image restoration, we propose a wavelet based algorithm to identify the type of image blurring operator, those are motion blurring along horizontal, vertical, and diagonal directions and defocused blurring. Especially for defocused blurring (Gaussian), we define an objective metric accordant to human perception for image blurring. The definition of the metric takes account into the different sensitivity of human vision for smoothing and non-smoothing region in an image, thus leads to the results consistent to those obtained by human rating.For edge detection, we define the measures to characterize the edge properties determined by the high frequency information arising from wavelet decomposition. These measures include Edge Intensity Measure that indicates the intensities of edges in an image, and Edge Diffusion Measure that describes the distribution of strong and weak edges. Because all the cases of edge degradation (e.g. bluring and noise corruption) are considered, the defined measures can be taken as objective basis for choosing detection algorithms and tuning parameters in the chosen algorithms.