Machine vision Algorithms and applications [notes]

1. The magnification of the dual telecentric lens is independent of the position of the analyte and the position of the image plane. In the object side telecentric lens, the magnification is a constant, corresponding to a fixed image plane.
2. Aberration of the lens: spherical aberration (symmetry): Aspheric lens instead of spherical lenses, large f, small light-through aperture
Poor (asymmetric): Use a smaller f, larger light aperture
3. Line scanning applications require very strong illumination, the lens usually uses a smaller F-value, limiting the depth of field
4. Lens selection should not be less than the sensor size, such as 1/2 ' lens cannot use 2/3 ' sensor
5.Image is a two-dimensional array
6.Region with stroke-coded records, only points within the storage area, multiple available travel-coded linked lists or arrays
7. Sub-pixel data can be derived from sub-pixel threshold segmentation and sub-pixel edge extraction
8.Image Enhancement: Gray value Linear transformation f (g) =ag+b; Generally, the way to find table Lut, gray value normalization is the best linear transformation of adaptive computing, but there is no obvious effect for the image with highlight and dark angle. ; Robust normalization removes the histogram pixels from the highlighted and dark corners, which can produce noticeable effects for feature extraction
9. The Fourier transform can be used to calculate the convolution of any linear filter in the frequency domain, such as with the mean filter or Gaussian filter
10.2D plane scaling, translation, rotation, tilt can be corrected by affine transformation; 3D spatial transformations require projection transformations, and affine is a special projection transformation .
11. Polar transformations are often used to correct a circular object in an image or an object that is contained in a ring.
12. Global value threshold segmentation; Histogram global threshold segmentation; Filter histogram robust global threshold segmentation, Filter gray scale local threshold value of dynamic threshold segmentation, smoothing filter size determines the size of the object can be divided, by experience, the width of the mean filter must be greater than the width of the recognized object. The same applies to median and Gaussian filters, in general, if the size of the mean or Gaussian filter is larger, then the filtered result will be more representative of the local background
13. When the defect is detected, the deviation model of multiple sets of images is taken into account in order to guarantee the tolerance.
14. Extracting connected domains, 4-connected and 8-connected, connected-domain travel codes using depth-first search
15. Sub-pixel accuracy result is contour edge, cannot be region, region can only be pixel precision
16. Area, moment, normalized moment, long R and short r of Ellipse;
17. The method for calculating the minimum bounding rectangle and minimum circumscribed circle of any azimuth is first based on the convex hull of the calculation area, where the convexity is defined as the ratio between the area of an area and the area of the convex hull of the area, the range [0,1], the compactness of the measuring area
18. Regional features, grayscale features, contour features
19. An expansive treatment of the foreground equates to a corrosive treatment of the background
20. If the camera is not geometrically calibrated, it is meaningless to extract the edges in sub-pixel accuracy. In addition to the use of extremely high-quality lenses in practical applications, the error caused by lens distortion is approximately 3 pixels
21. Robust minimum circle fitting a circle is first fitted with a normal minimum square and then the distance from each contour point to this circle is used to calculate the weights for each point that will be used in subsequent iterations. The RANSAC algorithm may be required in extreme cases because large outliers may hinder the eventual convergence of the correct solution of the algorithm .
22. The smaller the angle of the arc, the worse the accuracy of the fitting parameter
23. Transparent plane calibration plate is easy to operate, the size of the production can be very accurate, very convenient back light
24. Mean filter is the preferred filter for creating image pyramids 2x2
25. Each additional layer of image pyramid, image points and template points are reduced by 4 times times, that is, the operation speed increased by 16 times times.
The high-level match results map the next level of location.

Machine vision Algorithms and applications [notes]