Related Work

In order to obtain an image with proper exposedness and contrast, an image contrast enhancement technique is needed. There are plenty of researches proposed for image contrast enhancement. These methods can be classified into four major categories:

(1) Histogram-based methods [1-7];

(2) Transform-based methods [1], [8], [9];

(3) Exposure-based methods [10], [11];

(4) Image fusion based methods [12-14].

The most common and well-known histogram-based method is histogram equalization (HE) [1]. HE adjusts the input image histogram by using a non-linear mapping function to yield a histogram which approximates uniform distribution. It will spread the gray levels with high occurring probabilities and compress the gray levels with low occurring probabilities to obtain an image having better contrast.

However, HE was proved to produce some unwanted artifacts, including (1) False contour;

(2) Amplified noise;

(3) Washed-out appearance.

Various advanced HE based approaches have been developed [2-7]. Pizer et al.

3

[2] proposed a local HE method. First, the input image is divided into several non-overlapping blocks. Then, HE is applied on each block. The HE enhanced blocks are finally fused by using bilinear interpolation to reduce the blocking effect. Kim [3]

proposed a subimage independent HE method named brightness preserving bi-histogram equalization (BBHE). In BBHE, the input gray image is first

decomposed into two subimages based on its mean luminance, μ. Then, HE is applied to the histograms corresponding to these two subimages independently. The subimage with luminance lower than the mean value is mapped into the range [lmin, μ], where lmin denotes the minimum gray level. The subimage with luminance larger than the mean value is mapped into the range [μ, l_max], where l_max denotes the maximum gray level. Then, the composition of these two equalized subimages is the output image.

Wang et al. [4] proposed another bi-histogram HE method in which the input image is decomposed into two subimages by using the threshold value which yields maximum entropy of the processed image. Then, HE is applied to two subimages independently and the composition of HE enhanced subimages will form the output image. Chen et al. [5] extended the former two methods and proposed minimum mean brightness error bi-histogram equalization (MMBEBHE). In MMBEBHE, the threshold with

minimum absolute mean brightness error (AMBE) is found to divide the input image

into two subimages. Then, HE is applied independently to each subimage and the

4

composition of the HE result is the output image. Wang et al. [6] used histogram specification to yield the target histogram which maximizes the entropy under the constraint that the mean brightness is fixed. Chen et al. [7] proposed a method based on BBHE. They recursively divide each subimage into two new subimages and finally perform HE on each portions independently.

Transform-based methods [1], [8], [9] were widely used in electrical devices and computer software. These methods use a function to map original image luminance values to another ones. To get a pleasing image, some user-specified parameters are needed. That is, these methods require some user interactions and thus are not fully automatic. Transform-based methods can well handle either under-exposed images or over-exposed images if appropriate parameters are selected. However, they cannot produce pleasing images when the input images have both under-exposed regions and over-exposed regions. Moroney [8] proposed a local color correction operation which uses non-linear masking and a pixel-by-pixel gamma correction to enhance the image quality. Schettini et al. [9] presented a local and image-dependent exponential correction method which uses bilateral filter instead of Gaussian filter to avoid halo effects. However, the global contrast is reduced as well.

Exposure-based methods [10], [11] adjust the exposedness of images by using a function between the light quantity and the image gray values. Battiato et al. [10]

5

proposed a method which first identifies the information carrying regions and then adjusts the exposure levels using a “camera response”-like function. In their algorithm, contrast and focus are used as the measures to identify the information carrying regions. In addition, skin pixel identification method is applied to find the skin regions. Then, the mean gray values of those pixels in informative regions are used as reference values to adjust the exposure levels. Since the technique is designed specifically to regions of interest, it can produce proper results in those interested regions. However, other regions may yield poorer illumination. Safonov et al. [11]

provided an exposure correction approach based on contrast stretching and alpha-blending which considers both brightness and the estimated reflectance of the input image. The main problem of this method is that it may exhibit unsatisfied illumination in some regions.

Image fusion based methods [12-14], [22], [26] tried to extract and merge relevant information from several images taken in the same scene in order to form a fused image which contains more information and has better visual quality/contrast than each input image. Hsieh et al. [12] used a linear function to fuse the input image and a HE enhanced image to produce a fused image. Pei et al. [13] performed HE and sharpening to the input image and fused together these two enhanced images in the wavelet transform domain. Mertens et al. [22] used contrast, saturation and

6

well-exposedness as image quality measures to evaluate the contribution of each pixel to the fused image. First, for each input image, a corresponding weight map is computed. Then, the Laplacian pyramid of the input image and Gaussian pyramid of the weight map are built respectively. The Laplacian pyramids of the input images are blended with the corresponding Gaussian pyramid as the weights. Finally, the output image is produced from the blended Laplacian pyramid. Malik et al. [26] proposed an image fusion method performed in the wavelet transform domain. The output image is produced by taken the inverse wavelet transform.

The aforementioned contrast enhancement methods [1-14], [22], [26] often cannot produce pleasing images for a broad variety of low contrast images or cannot be automatically applied on all images. That is, some user-specified parameters are needed to obtain satisfied pictures. Therefore, we tried to design an image contrast enhancement algorithm which can automatically enhance the contrast without taking any user-specified parameters for any low contrast images.

In this thesis, we will propose a classified image fusion method for image contrast enhancement. First, several virtual images having different intensities are generated. In addition, the input image pixels are classified to several classes according to their luminance values. Finally, a classified image fusion method, performed in DWT domain, will be used to combine relevant information in the

7

virtual images and produce a fused image which is well-exposed in every region.