[Reading notes] computer vision and algorithm application

"Computer vision and Algorithm application" this book actually I have read the fourth chapter, has been used notes in the notebook, and now directly continue to remember here, the previous also no longer added ~

Chapter 4 feature detection and matching

• Key feature or point of interest or corner point
• Edge

4.1 Points and blocks

• There are two main ways to get the characteristic points and their correspondence: 1. In the first image, look for features that can be tracked accurately using a local search method, such as correlation or least squares 2, to detect feature points independently in all the examined images and then match them based on their local apparent view.
• Key point detection and matching pipeline: 1) feature detection (extraction) 2) Feature description 3) feature matching 4) feature tracking

4.1.1 Feature Detector

• Compare two blocks of images

• Calculate the stability of the matching result--"autocorrelation function" or autocorrelation surface

Matrix A gives a lower bound of the position uncertainty of the matching block, and the uncertainty can be visualized by eigenvalue analysis.

The uncertainty matrix of eigenvalue analysis corresponding to autocorrelation matrix A

• Basic idea: Use the local maximum value of rotation invariant scalar measurement derived from autocorrelation matrix to locate key points to achieve the purpose of matching sparse features. Use the Gaussian weights window instead of a square block of images
• Eigenvalue minimum λ0 is not the only amount that can be used to find the key points, others are:

, ,

• The steps of a basic feature detection algorithm

• Adaptive non-maximum suppression

Most feature detectors only find the local maximum value of the interest function, which results in the non-uniform distribution of the feature points, and the more dense regional feature points with large contrast. The workaround is to detect only those features that are both local maximums and whose responses are significantly larger than the response within the radius r area around them. It is first sorted based on the response strength of the feature points, and then a second sorted list is established by decreasing the suppression radius.

• Measurement of reproducibility

Repeatability: The frequency at which the key detected in one image is found within the ε-pixel range of the corresponding position in another transformed image.

"Available information" for each feature point (information content): A shang of a rotation-invariant local grayscale description sub-collection.

1. The scale is the same:
2. Rotation invariant and direction estimation
3. Affine invariant

[Reading notes] computer vision and algorithm application