Machine vision daily exercises

Lecture 2 Overview of Image Processing

1. Geometric transformation: Also known as image space transformation, it maps the coordinate position in one image to a new coordinate position in another image.
2. Image segmentation: The technology and process of dividing images into distinctive regions and extracting objects of interest.
3. Image edges:

1. Smooth, enhance, restore .
2. Generally include: image enhancement, image restoration , image coding, image segmentation, image reconstruction , etc.
3. Generally includes two parts:
   spatial operation , which realizes the position mapping of the input image and the output image and
   grayscale interpolation , which is used to calculate the grayscale value of the pixel at the new position.
4. Point operations and template operations
5. Smoothing and sharpening filters
6. Keep low frequency components and remove high frequency components
7. smooth
8. sharpen
9. Numerical similarity : such as grayscale difference, regional grayscale distribution
   spatial proximity : such as Euclidean distance, regional density
10. One is a region-based method ;
    the other is a contour estimation method using edge detection

11. Generally, first-order or second-order derivatives are commonly used to detect edges.
12. Step-shaped
    Pulse-shaped
    Roof-shaped
13. First-order differential edge detection operator
14. Single threshold segmentation
    Multiple threshold segmentation
15. Edges are the result of grayscale discontinuities and are the most significant changes in local intensity of the image . Such discontinuities can often be easily detected using derivatives
16. Gray threshold segmentation method
17. Region Growth
    Merging and Splitting
18. Additive noise and multiplicative noise

1. The main purpose of image preprocessing is to eliminate irrelevant information in the image, restore useful real information, enhance the detectability of relevant information and simplify the data to the greatest extent, thereby improving the reliability of feature extraction, image segmentation, matching and recognition.

2. Generally includes: digitization, geometric transformation, normalization, smoothing, enhancement, restoration, etc.

3. The purpose is to perform some related operations on the image to simplify subsequent image analysis (image description, image recognition) work

4. Generally include: image enhancement, image restoration, image coding, image segmentation, image reconstruction, etc.

5. Convenient to save images in the computer
   Convenient to calculate and process images

6. Linear smoothing filter + nonlinear smoothing filter

7. Convert the image from image space to frequency domain space (Fourier transform)
   to enhance the image in frequency domain space;
   convert the enhanced image from frequency domain space to image space (inverse Fourier transform).

8. Low pass filter, high pass filter, band pass filter, band stop filter

9. Noise: internal noise, external noise; Gaussian noise, salt and pepper noise
   Motion: image smear problem.

10. Numerical similarity: such as grayscale difference, regional grayscale distribution
    Spatial proximity: such as Euclidean distance, regional density

11. Edges are the result of grayscale discontinuities and are the most significant changes in local intensity of the image. This kind of discontinuity can often be easily detected using derivatives. Generally, first or second order derivatives are commonly used to detect edges.
12. The threshold is selected from the histogram grayscale distribution. For the grayscale image, the grayscale distribution is displayed using the image grayscale statistical information. The segmentation threshold is selected in different valleys, and the one-dimensional histogram thresholding method is generally used. Just select an appropriate threshold from the histogram for image segmentation.
13. Image segmentation is the technology and process of dividing images into distinctive areas and extracting objects of interest.
14. Boundary-based segmentation technology: first-order differential operator, second-order differential operator
    Region-based segmentation technology: parallel region technology, serial region technology

1. B. Median filtering has excellent filtering effect on salt and pepper noise.
2. B
3. A
   smoothing filter: It can weaken or eliminate the high-frequency components in the image, but does not affect the low-frequency components. It is mainly used to eliminate noise.
   Low-pass filtering
   principle: retain the low-frequency components and remove the high-frequency components.
   Result: Noise is removed or eliminated, but edge contours are also blurred.
4. B
   High-pass filtering
   principle: retain high-frequency components and remove low-frequency components.
   Result: Most of the information in the image is removed, leaving only edge contours.
   Sharpening filter: It can weaken or eliminate the low-frequency components in the image, but does not affect the high-frequency components. It is mainly used to enhance the blurred details or the edges of the target.
5. A
   
6. B thresholding is the most common parallel region segmentation algorithm

7. A
8. B (uncertain)
9. B
   filtering is a typical template operation
   that uses the grayscale relationship between the pixel itself and its neighboring pixels for enhancement.
10. The A
    image spectrum gives the global properties of the image, so frequency domain enhancement is not performed pixel by pixel.

1. (To be written, I wrote it wrong before, I’m so sorry)

No test

2. Using convolution, I won’t say much about this (the following results are not verified, the main purpose is to learn the calculation method)

This matrix is ​​special. It is very symmetrical. We can also add the corresponding positions and finally multiply by 1/4. The
edges are not considered in the question, so we will get a 3*3 edge matrix. Finally, just add the edges and the
result will be good. :

There are two methods for edges. The first is to fill the edges with 0, and the second is edge mapping, which maps edges
to edges.

handwriting:

Chapter Four

1. The characteristics of image objects are abstractions of the characteristics of image objects and are conceptual descriptions used to distinguish different types of objects.
2. Texture is the recurring local patterns in an image and their arrangement
3. 特征选择是指从一组原始特征中挑选出最有效的特征子集，以达到降低特征空间维数、简化分类器设计、提高分类速度的目的。

1. 手工设计特征（Hand Crafted, 传统的方法）
   特征学习（Feature Learning, 深度学习的方法）
2. 视觉特征、统计特征、代数特征、变换系数特征、其他物理特征
3. 视觉特征指人类视觉对目标的感觉特征，其主要包括：颜色、边缘、轮廓、纹理、形状，等等
4. 色彩感觉不仅与物体本来的颜色特性有关，还受时间、空间、外表状态以及该物体周围环境的影响，同时还受人的主观因素的影响。
5. 自然界的颜色分为彩色和非彩色两大类
6. RGB是一种相加模型，主要用于显示设备。
7. CMY是一种相减模型，主要用于印刷行业。即：混合的油墨能反射和吸收什么光。
8. 纹理反映了图像中像素亮度变化的一种趋势

1. 视觉特征、统计特征、代数特征、变换系数特征、其他物理特征

2. 视觉特征指人类视觉对目标的感觉特征，其主要包括：颜色、边缘、轮廓、纹理、形状，等等

3. 色是光作用于人眼引起的除形象以外的视觉特性，是人的一种心理反映。色彩感觉不仅与物体本来的颜色特性有关，还受时间、空间、外表状态以及该物体周围环境的影响，同时还受人的主观因素的影响。

4. RGB color model, CMY color model, HSV/HSI color model, YUV/YCbCr color model, CIE-Lab/L a b* color model
   RGB is an additive model mainly used in display devices.
   CMY is a subtractive model mainly used in the printing industry. That is: what light can the mixed ink reflect and absorb?

5. Shape factor, appearance ratio, expansion ratio, fullness, eccentricity, sphericity, Euler number

6. 1. Scale of texture: The same object has different texture characteristics at different scales
   2. Roughness of texture: Different substances may have different roughness of texture
   3. Regularity of texture: regular or irregular
   4. Texture Regionality: Different regions may have different textures

7. Mainly divided into two categories:
   texture features based on grayscale histograms, including: average brightness, average contrast, smoothness, third-order moment, consistency, entropy, etc.
   Texture features based on gray level co-occurrence matrix, including: energy, contrast, uniformity, correlation, entropy, etc.

8. Two termination conditions can be set:
   1) The specified maximum number of evolution generations MaxGen has been reached, or
   2) the optimal feature subset has not changed in consecutive generations.
   MaxGen and Gen are empirical constants.

1. A
2. B
3. C
4. B
5. B
6. A gray level co-occurrence matrix is ​​a common method to describe texture by studying the spatial correlation characteristics of gray levels.
7. D feature selection refers to selecting the most effective feature subset from a set of original features to achieve the purpose of reducing the dimensionality of the feature space, simplifying the design of the classifier, and improving the classification speed.
8. B Genetic algorithm is suitable for searching large feature spaces and does not require the evaluation function to be monotonic. Therefore, it has been successfully used in many fields such as feature selection, parameter optimization, and system control.

Calculation problems:

1. This is an 8-level grayscale image, so we first draw an 8 8 image, here I use xlsx to represent it.
   
   Then add one to the (previous number, next number) coordinates on this 8 8 matrix.
   Such as (0, 7), (7, 5), (3, 1), etc.
   
   Final:
   
   Fill in other positions with 0

Comprehensive questions:

Chapter 5 Invariant Characteristics (to be written...)

Glossary

1. The global invariant feature treats the entire image as a whole and treats every pixel data in the entire image indiscriminately, regardless of whether the data represents the target or the background.
2. The local features of the image are local structures composed of some pixels with large brightness changes. These local structures contain rich image information and are highly representative.

fill in the blank

1. According to the degrees of freedom of geometric invariance, it can be divided into
   translation invariant features, rotation invariant features, affine invariant features, scale invariant features and projection invariant features ;
2. According to the different levels of features, they can be divided into: point invariant features and region invariant features;
3. According to the size of the area during feature extraction, it can be divided into: global invariant features, local invariant features, and global and local invariant features.
4. The global invariant feature treats the entire image as a whole and treats every pixel data in the entire image indiscriminately , regardless of whether the data represents the target or the background.
5. origin moment
6. in

3. Short answer questions

1. During the image acquisition process, there are always factors such as scale scaling, rotation, translation, noise interference, changes in observation viewpoint, and lighting changes. These factors affect the robustness of the system. Therefore, when performing target recognition or tracking, features that are invariant to the above changes must be extracted from the image, and then feature matching is performed.
2. During the image acquisition process, there are always factors such as scale scaling, rotation, translation, noise interference, changes in observation viewpoint, and lighting changes. These factors affect the robustness of the system.
3. According to the degrees of freedom of geometric invariance, it can be divided into translation invariant features, rotation invariant features, affine invariant features, scale invariant features and projection invariant features;
4. as follows
   

4. Single choice

1. D
2. A
3. B

1. Center of mass coordinate formula:
   $$\begin{gathered} m_{(}pq)=\sum _{x=1}^C\sum _{y=1}^R{x}^p{y}^{q}f(x,y) \\ p,q=0,1,\mathrm{2...} \\ {x}_0=\frac{m_{10}}{m_{00}}{y}_0=\frac{m_{01}}{m_{00}} \end{gathered}$$

1. Fill in the blanks

1. Object classification mainly studies the question of what
2. Object detection mainly studies the problem of where
3. as follows:
   
4. small big
5. Based on interest point detection and dense extraction method
6. This can generally be achieved through clustering , or using spatial feature aggregation.

2. Short answer

1. as follows

Face recognition, pedestrian detection, intelligent video analysis, pedestrian tracking, etc. in the security field.
Traffic scene object recognition (car obstacle avoidance, etc.), vehicle counting, retrograde detection, license plate detection and recognition in the transportation field. Content-based
image retrieval in the Internet field, Automatically categorize photo albums

2. as follows

1. Instance level
   refers to a single object instance. Usually due to differences in lighting conditions, shooting angles, distances, non-rigid body deformations of the object itself and partial occlusion of other objects during the image collection process, the apparent characteristics of the object instance have great differences. Changes have brought great difficulties to visual recognition algorithms.
2. Category hierarchy
   First, there is large intra-category difference, that is, the apparent characteristics of objects belonging to the same category are quite different. The emphasis is on the differences between different instances within the category. For example, the appearance of the same chair is vastly different. Second, there is inter-category ambiguity, that is, differences between categories
   . Object instances of a class have certain similarities, such as wolves and huskies. It is difficult to separate the two in appearance. Again, the
   interference of the background is the reason. In actual scenes, objects cannot appear against a very clean background. Often, on the contrary, the background It may be very complex and interfere with the objects we are interested in, which greatly increases the difficulty of the identification problem.
3. Semantic Level
   Difficulties and challenges are related to the visual semantics of images, a typical problem called multistability

3. as follows

Simple to complex, special to general, small-scale to large
-scale. At first, the focus was on image classification in specific problems.
Then the focus was on classification and detection of general targets.
Now the focus is on larger-scale classification and detection.

4. as follows
   
5. as follows

Deep belief network DBN
DBN is a hierarchical undirected graph model.
The basic unit of DBN is RBM (restricted Boltzmann machine).
First, the original input is used as the visible layer, a single-layer RBM is trained, and then the first layer RBM weight is fixed, and the response of the RBM hidden layer unit is used as the new The visible layer trains the RBM of the next layer, and so on.
Through this greedy unsupervised training, the entire DBN model can get a better initial value, and then label information can be added to perform supervised fine-tuning of the entire network through production or discriminant methods to further improve the network. performance.
The multi-layer structure of DBN enables it to learn hierarchical feature expressions and realize automatic feature extraction.

6. as follows

The target positioning
task is to determine whether there is an object of a specified category in the input image; if it exists, give the position and range of the object, usually using a square bounding box.
The target classification
task is to determine whether objects of the category of interest appear in the selected image area (Proposals). The output label with a score indicates the possibility that the object of the category of interest appears in the selected area.

7. as follows

Main components of traditional detection algorithms

1. Selection of detection window
   Sliding window method (violent search)
   Selective Search area extraction method
2. Feature design
   Haar, HOG, LBP
3. Classifier design
   AdaBoost (SVM), SVM, decision tree

8. as follows

The general composition of the target detection algorithm based on deep learning

1. The feature learning part
   mainly implements feature learning through the stacking of several convolution, pooling, and activation layers.
   Attention modules will also be embedded to improve the learning effect.
2. Classification part:
   Target classification can be achieved through SoftMax classifier
3. Post-processing part (optional)
   non-maximum suppression

3. Single choice

1. A
2. D

4. Calculation

1. It is the convolution operation. This does not consider the edges and calculates directly. Finally, the edge values ​​​​of the original image are added.
   
   5. Comprehensive questions

slightly