Article directory
color characteristics
Quantize color histogram
Applicable color space: RGB, HSV and other color spaces
- HSV space calculation is efficient, but there are quantization problems and sparsity
operate
- Color space quantization, the unit (bin) is represented by the center of the unit
-
- Count the number of pixels that fall on the quantization unit
Geometric Features
Edge Edge
edge definition
- The region where the pixel value function changes rapidly -> the extreme value region of the first derivative
edge extraction
- Gaussian denoising first, and then use the first derivative to obtain the extremum
- Derivatives are sensitive to noise
Keypoint-based Feature Descriptors
introduce
Geometric Features: Keypoints
Geometric features: Harris corners
FAST corner detection
Geometry Features: Blobs
- –
- When sigma is small, more detailed edges will be identified
Local features: SIFT
- Based on scale-space invariant features
- Features:
- good invariance
- good uniqueness and rich information
- multiple measures
- fast calculation
- calculation steps
- The operating results of scale space
LoG and DoG are similar, but the operation efficiency of DoG (difference) is much higher than that of LoG (differentiation).
- Gaussian Pyramid
Supplement:
1. Why build a Gaussian pyramid?
The Gaussian pyramid imitates the different scales of the image. How should the scale be understood? For an image, when you observe the image at a close distance, the effect of the image you see is different from that when you observe it one meter away. The former is clearer, the latter is blurrier, the former is larger, and the latter is smaller. Through the former, you can See some details of the image, through which you can see some outline information of the image, this is the scale of the image, the scale of the image exists naturally, not artificially created. Well, here we understand that in fact, the processing of an image was relatively monotonous before, because our focus was only on two-dimensional space, and the concept of "image depth" was not considered. If these Taking the content into consideration, will we get more information that was not obtained in the two-dimensional space before? So the Gaussian Pyramid was born, it is to squeeze out another dimension that naturally exists in the image on the basis of the two-dimensional image: scale. Because the Gaussian kernel is the only linear kernel, that is to say, using the Gaussian kernel to blur the image will not introduce other noise, so the Gaussian kernel is selected to construct the scale of the image.
2. The construction steps of the Gaussian pyramid:
In the process of building the Gaussian pyramid, the image is generally doubled first, and the Gaussian pyramid is constructed on the basis of the enlarged image, and then the image under the size is Gaussian blurred, and several images after blurring The collection constitutes an octave, and then the process of down-sampling the most fuzzy image under the octave, the length and width are shortened by one time, and the image area becomes a quarter of the original.
-
SIFT-feature point direction estimation
The rotation process is more complicated, and when rotating a specific angle, the point grid coordinates are difficult to express -
SIFT-calculate feature point descriptor
Preliminary knowledge: Haar-like features
-
Haar-like features are divided into: edge features, linear features, center features and diagonal features. They are combined into feature templates.
The Haar feature reflects the grayscale variation of the image. -
Fast Computation of Haar-like Features: Integral Maps
Local features: SURF
Improvements to SIFT
using Hafen matrix
- small shape
Feature point → main direction → 64-dimensional vector
Analyze the Hessian matrix to form a template
- ORB feature description
BRIEF
ORB calculates the centroid.
LBP
-
Improved LBP
grayscale invariance and rotation invariance -
Gabor filter