Abstract

Contrast enhancement is essential to improve the image quality in most of image pre-processing. A histogram equalization process can be used to achieve a high contrast. It causes, however, also noise generation. Involving a low-pass filtering process is an effective way to achieve a high-quality contrast enhancement with low-noise, but it leads to the conflict between noise removal and signal preservation. To perform discriminative low-pass filtering operations with the presence of noises and signal variations in different regions, it is thus necessary to develop good algorithms to classify the pixels.
In this thesis, two classification algorithms are proposed. They aim at low-contrast images where gradient signals are severely degraded by various causes during the acquisition process. They are to classify the pixels according to the initial gray-level homogeneity of their regions. The basic classification method is done by gradient thresholding, and the threshold values are generated by means of gradient distribution analysis. To tackle the problems of various gradient degradation patterns in low-contrast images, image pixels are grouped in a particular way that, in the same group, pixels in homogeneous regions can be easily distinguished from those in non-homogeneous regions by the basic method of simple gradient thresholding. Two algorithms based on different grouping methods are proposed. The first algorithm aims at high dynamic range images. The pixels are first grouped according to their gray-level ranges, as the gradient degradation is, in such a case, gray-level-dependent. The gradient distribution of each sub-range is obtained and a pixel classification is then made to adapt to their original gray-level signals in the sub-range. The other algorithm is to tackle a wider range of low-contrast images. In this algorithm, a gray-level histogram thresholding is performed to divide the pixels into two groups according to their likelihood to homogeneous, or non-homogeneous, pixels. Thus, in one group a majority of homogeneous pixels is established and in the other group the majority is of non-homogeneous pixels. The classification done in each group is to identify those in the minority.
Both proposed algorithms are very simple in computation and each of them is incorporated into the contrast enhancement procedure to make the integrated low-pass filters effectively remove the noise generated in the histogram equalization while well preserving the signal details. The simulation results demonstrates, by subjective observation and objective measurements, that the proposed algorithms lead to a superior quality of the contrast enhancement for varieties of images, with respect to two advanced enhancement schemes.