UCAS-AI Academy-Computer Vision Special Course- Lecture 2-Course Notes

Image feature extraction

• The basic steps of feature extraction computer vision
• Features: able to reflect the details of the image content
  • Edges and contours
• Reliable edge and key point extraction can solve many visual problems
• edge
  • Object boundary
  • Changes in surface orientation
  • different color
  • Dark changes in light illumination

Edge extraction

• Definition of edge
  • The area in the image where the brightness suddenly changes
  • Steep area composed of gray
  • A collection of step changes in grayscale or roof changes
• Edge type
  • Stepped edge
  • Ridged edge (gradient)
  • Linear edge
• motivation
  • The most basic image features
  • Insensitive to image changes (geometry, grayscale, lighting)
• Ideas
  • Noise suppression (LP)
  • Edge feature enhancement (HP)
  • Edge positioning
• Differential operator
  • First-order differential operator: gradient, detection maximum
    • The amplitude indicates the strength of the edge, and the direction indicates the direction with the fastest grayscale change
    • Sensitive to noise-noise suppression (filtering-convolution filtering after operator differentiation)
    • Prewitt: Approximate first-order differential, denoising + edge enhancement
    • Sobel: Approximate first-order differential, with greater weight in the four-neighborhood
  • Second-order differential operator: Laplace, detecting zero crossing
    • Laplacian: The direction attribute is lost and it is very sensitive to noise
    • LOG: Gaussian smoothing first, filtering in Laplacian-filtering after differentiating the operator
      • Straw hat filter
      • Can be implemented in two parts for greater flexibility
      • Accurate positioning, but produces many closed contours, and zero-crossing detection also requires complex algorithms
• Canny edge detection
  • Optimal principle: edge detection (edge ​​is more obvious than noise) + good localization (maximum suppression) + low error (single extreme point)
  • Calculate gradient-local extremum
  • Non-maximum suppression-only the maximum point in the gradient direction is retained
  • Double threshold extraction edge
    • Large threshold-a small number of edge points, a large number of gaps
    • Small threshold-a lot of edge points, a lot of errors
    • Edge links: large threshold results extend along small threshold results
  • Less parameters, high calculation efficiency, continuous and complete edges

Feature point extraction

• Feature points are the basis

• Good corner detection algorithm

  • True corner
  • Accurate positioning
  • Good stability
  • Robust to noise
  • Higher efficiency

• Harris corner detection: Observe image features from a small window

  • Corner point: There is obvious grayscale change in any direction of the window
  • Edge: No obvious change in the direction of the edge
  • Window pan $[u, v]$ resulting grayscale change $E(u, v) = \sum_{x,y} [I(x + u, y + v) - I(x, y)]^2$
  • Taylor expands the quadratic term to get the form $E(u, v) = \left[ \begin{array}{c} u & v \end{array} \right] \left[ \begin{array}{c} I_x^2 & I_x I_y \\ I_x I_y & I_y^2 \end{array} \right] \left[ \begin{array}{c} u \\ v \end{array} \right]$
  • For small translations, it can be expressed as $E(u, v) = \left[ \begin{array}{c} u & v \end{array} \right] M \left[ \begin{array}{c} u \\ v \end{array} \right]$
  • Use its elliptical form to turn the problem into $M$Eigenvalue analysis of$M$ (the largest and smallest eigenvalues ​​correspond to the fastest and slowest direction, respectively)
  • Flat area, both eigenvalues ​​are small
  • Edge, one of which is significantly larger than the other
  • Corner points, both are very large, and the value is equivalent
  • Corner corresponding function $R = \det M - k (\operatorname{trace} M)^2 = \lambda_1\lambda_2 - k (\lambda_1 + \lambda_2)^2$
    • $k$ generally 0.04 to 0.06
    • Corner point, R is a large positive number
    • Edge, R is a large negative number
    • Flat area, R is a small value
  • algorithm
    • Correct $R$ thresholding
    • Extract the local maximum after processing
  • nature
    • Rotation invariance (eigenvalue unchanged)
    • The affine transformation of the gray level is partially invariant (the translation and addition are unchanged, the scale number is changed)
    • Geometric scale changes

• ORB feature detection

  • FAST (feature detection) + BRIEF (feature description)
  • Fast and accurate
  • FAST: The gray value is larger or smaller than the gray value of enough pixels around it, which is the corner