Exam question type: multiple choice 10/2' | fill in the blank 10/1' | true or false 10/1'|short answer 6/5'|comprehensive 3/10'
!!PS: Note that this article is for learning reference only, if there is any error, please contact me by private message for modification~~
End of term review knowledge point 1
1. Short answer questions
1. What are the main research contents of digital image processing? (P6-8)
Answer:
1) Image digitization—the purpose is to represent an image in digital form, and it must be free from distortion and easy to be processed by a computer.
2) Image enhancement - enhance the useful information in an image while suppressing useless information to improve the observability of the image.
3) Image geometric transformation - changing the size or shape of an image.
4) Image restoration - restore the original information of the degraded and blurred image to achieve the purpose of clarity.
5) Image reconstruction—constructing an image of a three-dimensional object based on two-dimensional plane image data.
6) Image hiding - hiding an image or some digitizable media information in an image.
7) Image transformation——refers to a mathematical method that converts image information in the air domain to spaces such as frequency domain, time-frequency domain, etc. for analysis through a mathematical mapping method.
8) Image coding - mainly to compress image data. The purpose is to simplify the image representation method and compress the data representing the image for easy storage and transmission.
9) Image recognition and understanding - refers to the extraction of the desired target after the quantitative description of various object features in the image, and a certain quantitative analysis of the proposed target.
2. Concepts of image digitization, sampling, quantization, and histogram.
Answer:
Image digitization refers to the discretization of analog images to obtain images represented by numbers.
Image digitization includes two processes of sampling and quantization .
Sampling (Sampling) : It is the operation of converting a continuous image in the spatial or temporal domain into a discrete set of sampling points (ie, pixels). That is: discretization of spatial coordinates. In fact, the sampling method is determined by the sensor device that generates the image.
Quantization : The conversion of the gray value of each pixel obtained after sampling from analog to discrete is called the quantization of image gray. That is: discretization of gray scale. Determined by grayscale. Quantization can be divided into uniform quantization and non-uniform quantization.
Histogram refers to the grayscale histogram, that is, the distribution of pixel grayscale values in a digital image. The gray level histogram is a function of the gray level distribution, which is the statistics of the gray level distribution in the image. The gray histogram is to count all the pixels in the digital image according to the size of the gray value and count the frequency of occurrence. It reflects the frequency of a certain gray level in the image.
3. What is image enhancement? The role of image enhancement?
Answer: Image enhancement is to enhance the useful information in an image while suppressing useless information to improve the observability of the image. The purpose is to improve the visual effect of the image, or to facilitate the analysis and processing of humans or machines.
4. Classification of geometric transformation?
Answer: Geometric transformation includes: image position transformation and image shape transformation .
Image position transformation:
1. Translation - the scene after translation is the same as the original image, but the "canvas" must be enlarged. Otherwise information will be lost.
2. Mirroring - horizontal mirroring and vertical mirroring - because the matrix coordinates representing the image cannot be negative, it is necessary to translate the coordinates after the mirroring calculation.
3. The
pre-rotation processing includes : expanding the canvas, rounding processing, and translation processing.
The purpose of canvas expansion: to avoid the loss of image information; the principle of canvas expansion: to carry all the picture information with the smallest area.
Empty solution ideas : The core of the problem is that the connection between pixels is discontinuous.
The angle of adjacent pixels cannot be changed, so the problem can only be solved as a whole by increasing the resolution. Use some kind of filling method to fill the void.
Hole solution : Row interpolation (column interpolation)
Step 1: Find the coordinates of the smallest and largest non-background points (target objects) in the current row, denoted as: (i,k1), (i,k2).
The second step: perform interpolation within the range of (k1, k2), the interpolation method is: the pixel value of the empty point is equal to the pixel value of the previous point.
Step 3: Repeat the same operation to all rows.
Image shape transformation
1. Reduction
Implementation ideas: Image reduction is actually to select or process multiple original data to obtain data that is expected to be reduced in size, and try to keep the original features without losing them.
Select data at equal intervals, select data at non-equal intervals
2. Zoom
in 3. Miscut: In fact, it is the non-perpendicular projection effect of the plane scene on the projection plane,
because most of the images are obtained by the projection of three-dimensional objects on the two-dimensional plane , so it is necessary to study the miscutting phenomenon of the image.
Affine transformation of the image;
note: the x direction and y direction are the row and column directions of the matrix.
5. Noise suppression is sometimes called smoothing, image smoothing is also called noise suppression, what is image smoothing?
Answer:
The function of filtering is to filter out unnecessary components according to needs and keep the required components.
Image smoothing (smoothing): denoising, blurring details, weakening high-frequency components in the image, and having the characteristics of low-pass filtering.
Image sharpening: enhance the edges and details, weaken the low-frequency components in the image, and have the characteristics of high-pass filtering.
6. What is noise? What are common noises? Are there common denoising models or filtering models?
Answer:
Noise is one of the most common image degradation factors, and it is a kind of external interference; it is the unwanted part of the image, the unwanted part of the image, and the part that is expected to be filtered out. Image noise is a random interference signal received during image acquisition, image digitization, or image transmission.
Common noise types: Gaussian noise (the position of appearance is certain (at each point), but the amplitude of the noise is random, also known as normal noise), pulse (salt and pepper) noise (the amplitude of the noise is basically the same, but the noise The position of occurrence is random), Rayleigh noise, exponential noise, uniform noise, periodic noise.
Common denoising models: mean filtering and median filtering.
Mean filter : Features: The coefficients of the mask are all positive; the gray value range remains unchanged (the sum of all coefficients is 1).
When the size of the template increases, the effect of noise removal is enhanced, but the image becomes blurred, that is, the edge details are reduced
Median filter : blunt image, remove noise.
Because of the appearance of noise (such as salt and pepper noise), the pixel at this point is much brighter (darker) than the surrounding pixels. If in a template, the pixels are rearranged from small to large, then the brightest or darkest points must be arranged on both sides. The purpose of filtering noise can be achieved by replacing the value of the pixel to be processed with the gray value of the pixel at the middle position in the template.
For salt and pepper noise, the median filter works better than the mean filter.
7. What is image sharpening? Boundary-based image segmentation, the core idea of image sharpening? What are the common sharpening operators?
Answer: Image sharpening refers to enhancing the details, edges and outlines of the scene in the image. Its essence is filtering, which removes areas with relatively gentle grayscale changes, that is, removes low-frequency signals, and retains areas with relatively drastic grayscale changes, that is, retains high-frequency signals.
The effect is to enhance the grayscale contrast.
Realization: Because both the edge and the contour are located at the place where the grayscale changes. So the implementation of the sharpening algorithm is based on differential equations.
Common sharpening operators:
Unidirectional first-order differential sharpening refers to enhancing edge information in a specific direction. Because the image is composed of horizontal and vertical directions, the so-called unidirectional sharpening actually includes sharpening in the horizontal and vertical directions.
Non-directional first-order differential sharpening: cross-differential (Roberts) sharpening, Sobel sharpening, Priwitt sharpening
From the gray distribution characteristics of the scene details of the image, the description of the first-order differential of some gray-scale change characteristics is not very clear. Therefore, using the second order differential can obtain richer scene details.
Laplacian Algorithm, Wallis Algorithm
Comparison:
First-order differential produces wider edges
Second-order differential has a stronger response to details, such as thin lines and isolated pointsSecond-order differential
has a stronger response
The degree step change produces a double response.
Take the Sobel and Laplacian algorithm as an example for comparison:
The boundary obtained by the Sobel operator is a relatively rough boundary, which reflects less boundary information, but the reflected boundary is relatively clear; the
boundary obtained by the Laplacian operator Boundaries are finer boundaries. The reflected boundary information includes a lot of detailed information, but the reflected boundary is not very clear.
8. Binary image segmentation? The purpose of image segmentation? connected domain? How to label? Morphological operations - erosion, dilation, opening, closing.
Answer: Image segmentation is divided into: non-continuous segmentation (boundary-based segmentation); similarity segmentation (region-based segmentation).
Image segmentation extracts the target of interest from the image. Image segmentation processing is actually to distinguish the "foreground object" and "background" of the image, so it is usually called image binarization processing, also known as object detection.
The basic idea of threshold segmentation : Assume that the image is composed of two types of regions (target and background) with different gray levels. According to the difference in the grayscale characteristics between the target and the background to be extracted in the image, an appropriate threshold is selected to determine whether each pixel in the image should belong to the target area or the background area, thereby generating a binary image.
The scope of application of the threshold segmentation method : It is suitable for situations where there is a strong contrast (obvious difference) between the object and the background, and the gray level of the background or object is required to be relatively single.
Common methods of threshold segmentation : simple global threshold segmentation, semi-threshold segmentation (only the background of the image is represented as black or white, while the objects in the image are still multi-valued images), local threshold segmentation Threshold selection is the key to threshold segmentation
technology .
Common methods for determining the threshold : the threshold is obtained by manual selection (p parameter method), the threshold is obtained by the histogram, and the threshold is obtained by iterative calculation.
The basic idea of threshold value obtained by histogram : If an image is assumed to be composed of only two parts of the object and the background, its grayscale histogram will form obvious double peaks; in this case, select the gray at the bottom of the valley between the double peaks The degree value T is used as the threshold to separate the object and the background well.
The connected domain generally refers to the image area composed of foreground pixels with the same pixel value and adjacent positions in the image.
attach a label: Because different connected domains represent different goals, in order to distinguish them, different connected domains need to be identified.
- Initialization: Set the label number as Lab=0, the number of labels N=0, the label matrix g is all 0 arrays, and search for unlabeled target points in the order from top to bottom and from left to right;
- Check the status of adjacent pixels: perform corresponding processing according to the status of adjacent pixels in the template;
2.1 If the scanned pixels are all 0, then Lab=Lab+1, g(i,j)=Lab,N=N +1;
2.2 If the scanned pixel label numbers are the same, then g(i,j)=Lab;
2.3 If the scanned pixel label numbers are different, for example: Lab2>Lab1, then g(i,j)=Lab1, N=N-1, modify all the pixel values of Lab2 to be Lab1; - Process all the pixels according to step 2 until all the pixels are processed;
- Determine whether the final Lab satisfies Lab=N: if yes, the labeling process is completed; if not, it indicates that there is a non-consecutive number in the label. At this time, a code arrangement will be carried out to eliminate the situation of discontinuous numbering.
Erosion is a process that eliminates the boundary points of the connected domain and shrinks the boundary inward.
Design idea: Design a structural element. The origin of the structural element is located on the target pixel to be processed. By judging whether it is covered, it is determined whether the point is corroded.
Algorithm steps:
1. Scan the original image to find the first target point with a pixel value of 1;
2. Move the origin of the structural element with a preset shape and origin position to this point;
3. Determine the coverage of the structural element Are all the pixel values of 1:
3.1 If yes, the pixel value at the same position in the etched image is 1;
3.2 If not, the pixel value at the same position in the etched image is 0;
4. Repeat 2 and 3, until all pixels in the original image are processed.
Dilation is a process of combining the background points of the target area with the target to expand the boundary of the target to the outside.
Design idea: Design a structural element. The origin of the structural element is located on the background pixel, and judge whether the target point is covered to determine whether the point is expanded as the target point.
Algorithm steps:
1. Scan the original image to find the first background point with a pixel value of 0;
2. Move the origin of the structural element with a preset shape and origin position to this point;
3. Determine the coverage of the structural element Whether there is a target point with a pixel value of 1:
3.1 If yes, the pixel value at the same position in the expanded image is 1;
3.2 If not, the pixel value at the same position in the expanded image is 0;
4 .Repeat 2 and 3 until all pixels in the original image are processed.
The opening operation is to corrode the original image first , and then expand it. The opening operation can basically maintain the size of the original target while separating the sticky target.
The closing operation is to expand the original image first , and then corrode it. The closing operation can basically maintain the size of the original object while merging the fractured objects.
9. Color image processing - conventional color models, such as: RGB, CMYK, HSV?
Answer:
In order to use a computer to express and process colors, quantitative methods must be used to describe colors, that is, to establish a suitable color expression model to correctly and effectively represent color information. There are three types of color models widely used at present:
computational color models, also known as colorimetric color models, are mainly used in pure theoretical research and calculation derivation, such as XYZ, LAB;
industrial color models focus on the realization technology of practical applications, such as color display Or the color model used by hardware devices such as printers, such as RGB (including most graphic displays such as computer monitors and color TVs), YUV, YIQ, CMYK (this color system is used in the printing industry. It is a reduction Color system, the color obtained after filtering out the three primary colors from white light is used as the three primary colors CMY of its color system. K is black, and black ink can be used for printing), YCbCr (this is commonly used in color image compression A color model at the time); the
visual color model is used for occasions that directly interface with people and aim at color processing, such as the HS* series, including HSL, HIS (H: represents hue, saturation component S, brightness component I ), HSB. HSV?
RGB to HSI conversion
ps: Various color models can be converted mathematically .
10. Image transformation, Fourier transform - nature, what is high-pass and low-pass?
Answer: Fourier transform properties - translation, distribution law, proportional transformation, separability, rotation, mean value, energy conservation principle, periodicity and conjugate symmetry, convolution, correlation.
Translation properties: the translation of the image in the spatial domain does not affect the spectrum amplitude (the amplitude remains unchanged), and only corresponds to the phase shift in the frequency domain (only the phase spectrum is changed). High-pass means that signals higher than the set frequency are allowed to pass
through . Signals with a frequency lower than the specified value are prevented from passing through, and the lower the frequency, the greater the relative blocking effect.
Low-pass corresponds to high-pass, that is, there is little resistance to signals below the specified frequency, and a large attenuation for signals above the specified frequency.
Summary of Fourier transform:
It is time-consuming to perform complex calculations; and when M and N are both integer powers of 2, using FFT can greatly improve the calculation speed; the
origin of the spectrum calculated by FFT is not located in the center of the spectrum, and it needs to Translation processing; the spectrum after "shifting" is symmetrical about the origin; the
main energy of the image is concentrated in the low frequency, and the high frequency component decays too quickly, so using logarithmic transformation to display the spectrum of the image can achieve better results;
11. How to register images? Purpose - for fusion
Answer:
Image registration seeks a kind of (or a series of) spatial transformation for an image to make it spatially consistent with the corresponding points on another image. This agreement means that all points on the object (or at least all points of interest) have the same spatial position on the two registered images.
Image registration is the premise of image fusion. Only when image registration is done well, can the later image fusion achieve good results.
Four steps of image matching: feature detection (features include: point, line and surface), feature matching, transformation model estimation, image resampling and transformation
12. What is data redundancy? Redundant classification?
Answer:
Some data must represent useless information, or repeatedly represent information that other data has already represented, which is data redundancy.
There are three basic types of data redundancy in digital image data, namely inter-pixel redundancy, psychovisual redundancy and coding redundancy.
Inter-pixel redundancy: Redundancy caused by the correlation between adjacent pixels of an image. Including:
Spatial Redundancy: The signal changes slowly in most areas of the image, especially the background. Most of the visual contribution of a single pixel to an image is redundant, and its value can be properly predicted from this pixel's
neighborsVideo) Correlation coding redundancy
between adjacent frames in the image sequence
Redundancy, the gray level with high probability of occurrence should be expressed with as few bits as possible. This method is called "variable length coding".
13. How to compress images, the concept of compression? Compressed classification? Encoder, decoder? Conventional compression coding method (lossless: compression method corresponding to BMP file - run-length coding) (lossy: compression coding method corresponding to JPEG file - discrete cosine transform in transform coding), look at the compression effect - compression ratio?
Answer:
Image compression can save storage space and transmission time; in short,
compression for storage and transmission can be divided into two categories. The first type of compression process is reversible, that is, the compressed image can be completely The original image is restored without any loss of information, which is called lossless compression; the second type of compression process is irreversible, and the original image cannot be completely restored, and some information is lost, which is called lossy compression. Generally, under normal observation conditions, there is no obvious loss (visually lossless);
usually the compression ratio of lossy compression is higher than that of lossless compression; the higher the compression ratio of lossy compression, the greater the distortion.