Image enhancement and image quality improvement

Generally, the source image information acquired by the image information input system contains various noises and distortions. The purpose of image enhancement is to improve the visual effect of the image, the adaptability of the process, and to facilitate the analysis and embedding of humans and computers to meet the requirements for image reproduction or reproduction. Image enhancement and image quality improvement are mainly caused by color conversion and gradation conversion. Image sharpening, noise removal, geometric distortion correction, and image size conversion. In short, some operations are performed on the grayscale and coordinates of the image.

In the process of image enhancement and image quality improvement, the emphasis on one part of information and the loss of another part of information are always paid for. Therefore, under the premise of reducing the image quality under the premise of improving the image quality for image processing, there is no uniform quality evaluation standard for image enhancement and image quality improvement at present, that is, the lack of image appearance. The mathematical metrics of subjective and objective judgments, so the means for image enhancement and image quality improvement have yet to be further studied.

First, the color transformation

Color conversion is a key means of color image processing. It can be known from the foregoing that a color image is generally represented by three independent colors (for example, R, G, B, Y, M, and C), so that three color-shaded images can be obtained after color decomposition. The three color separation images are mutually independent and mutually restrictive, each of which can be processed using a grayscale image processing method, but must comply with a common setting, that is, to ensure that the three images are synthesized. The color reproduction of the physical image is correct. Therefore, a color image is different from a one-dimensional gray image and is a three-dimensional image. Therefore, in color conversion, the key to color image input is the accuracy of the color decomposition and the accuracy of the positioning (fitting) between the three color separation images. The key to color image output is correct color reproduction and accurate fit.

According to the theory of colorimetry, any color can be expressed by the light of three kinds of color fish, which are called three primary colors. There are two color mixing methods. The first is the additive method. The additive primary colors are red (R), green (G), and blue (B), respectively; the second is subtractive. The subtractive primary colors are yellow (Y), magenta (M), and cyan (C). For different purposes and different application areas, different color systems are used in color conversion. The following are commonly used:

(1) 1931 CIERGB system

(2) 1931 CIEXYZ system

(3) 1960 CIE uniform chromaticity coordinate system

When a tristimulus value x,y,z is given:

When given the chromaticity coordinates x, y:

Only uniform coordinates of the two spaces of hue and saturation are given in the chroma space, and no brightness is set.

(4) 1964 CIE W* U* v* uniform color space system

Among them: W* is lightness, U*, V* are chromaticity coordinates.

u,v is the uniform chromaticity coordinates of the CIE 1960 VCS of the sample color.

U0, v0, the uniform color coordinates of the light source.

In 1964 CIEW*U*v*, the color difference is

Like f(x,y) or P(D) (for discrete digital images f(i,j)). The grayscale histogram reflects the ratio of the area or the number of pixels of different grayscale values ​​D in the entire image, and can further reflect the amount of information contained in the image. However, the histogram has nothing to do with the position of the pixel and does not involve the shape of the image.

(1) Gray-scale histogram of continuous image f(x,y)

Let the area of ​​an image be A and A(D) be the sum of the areas of the image where the gray value is less than D, then the probability p (D) is:

And

Then make a Dp(D) curve as shown in Figure 3.12(a), which is a grayscale histogram of f(x,y).

(2) Gray-scale histogram of digital image f(i,j)

Let M×N digital image f(i,j) be the pixel grey value of the image and the probability is:

And

Thus, the distribution of P(D) can be easily calculated by the computer from f(i,j), and since D is discrete, the histogram of the histogram is discontinuous. As shown in Figure 3.12(b).

2. Grayscale enhancement

Gray scale enhancement, also known as gray scale conversion, refers to a method of changing the gray value of each image cable in the original image point by point according to certain target conditions. That is, if the gray value D=f(x,y) of the pixel of the original image and the gray value D′=g(x,y) of the image pixel after processing, the gray scale enhancement can be expressed as:

Or D'=T(D)

When the gradation transformation relationship D′=T(D) is determined, a specific grayscale enhancement method is determined. D'=T(D) is usually a single-valued function. Grayscale conversion is also referred to as gradation correction in electronic color separation.

(1) Linear transformation of gray scale If D′=T(D) is a linear or piecewise linear single-valued function, then the gray scale transformation that it walks is called gray scale linear transformation, or simply linear transformation.

(i) Linear transformation

Let the image gray value D = f (x, y) possible value domain, but under certain conditions, so that its value range is reduced to and shown in Figure 3.13, this state often appears in the continuous image f Since the (x, y) value has a small dynamic range or insufficient exposure for photography, the corresponding histogram P(D) is concentrated in a small gray scale interval as shown in Fig. 3.13(c). However, people generally want the grayscale histogram to be evenly distributed throughout the entire grayscale region. The simplest way is to convert the low-contrast grayscale to obtain a high-contrast image, that is, a linear transformation, as follows:

Equation (3-22) can expand the value range of the transformed grayscale D'. When it is, the curve of the linear transformation relation DD′ is as shown in Fig. 3.13(b). After the transformation, the histogram of the grayscale D' is transformed from Fig. 3.13(c) to Fig. 3.13(d).

From the foregoing, it can be seen that human vision can be distinguished only when the gradation values ​​(brightness values) of two adjacent pixels are different to a certain extent. If the gray value D is only in a small interval, the total number of levels of brightness difference that can be distinguished by the human eye is also small, resulting in that the target image gray value is close to the background gray value, and the human eye cannot Distinguishing detection. However, when D′=T(D) is transformed, it can be made larger as shown in Fig. 3.13, so that the total number of luminance differences visually resolved by the transformed image is increased, resulting in a target image. The increase in the difference in brightness between the background and the background makes it possible to detect the target image that could not be detected by the human eye, and the image sharpness after the conversion is greatly improved.

At the same time, it should be noted that the results of grayscale enhancement before and after digitization of the image are not the same. If the digitized image f(i,j) is enhanced, only the k different values ​​are taken in the interval because of the pixel gray value D. After the transformation, the gray value D′ is in the enlarged range. It can only take k different values, but the contrast increases. If the difference between the target to be detected and the background gray value in the continuous image f(x,y) is small and the respective quantized value enters the same gray level, the target has actually disappeared in f(i,j) At this time, with grayscale enhancement, the target image cannot be highlighted. If you enhance the image f(x,y) before digitizing, you have to:

From (3-23), it can be seen that when C>1, an amplifying effect is generated, and the slight difference between the target image and the background is magnified. If c = 2 and then digitized, f(i, j) will have 2k different grayscale values, and the difference between the same target image and its background will increase by a factor of two. As long as there is a difference, it can be highlighted with grayscale enhancement means. In the foregoing situation, the goal has disappeared during the quantification process. It can be seen that the digital image f(i,j) does not completely reflect the original image f(x,y), and the choice of the quantization order is very important.

Figure 3.14(a) shows a linear transformation of the positive slope of the curve of DD′. Its expression is:

The function of this formula is to invert the grayscale even if the maximum (minimum) value in the grayscale value of f(x,y) is transformed into the smallest (maximum) value in g(x,y). That is, the larger the gray value of the pixel value in f(x,y), the smaller the gray value of the pixel corresponding to g(x,y), and vice versa. It is used for positive/negative transformations, ie, positive or negative negatives.

(ii) Piecewise linear transformation

A piecewise linear transformation is an algorithm that divides the range of values ​​in the range of the image as shown in Figure 3.15 and performs different linear transformations. Partition linear transformation can be used to compress a part of the grayscale interval according to the purpose, and expand into a part of the grayscale interval. Its transform expression is:

Piecewise linear transformation can use a multi-segment polyline to form a unit function, and can also approximate a curve to complete a more detailed transformation.