Feature extraction algorithm of pattern recognition

Description: Here briefly introduces a variety of feature extraction algorithms, follow-up.

In the pattern recognition, the identification is based on the image characteristics when the matching recognition or classifier classification is identified. The extracted features are used to represent the whole image content, match or classify the image target according to the feature. Common feature extraction algorithms are divided into the following 3 categories:

① based on color characteristics: such as color histogram, color set, color moment, color aggregation vector, etc.;
② based on texture features: such as Tamura texture feature, autoregressive texture model, Gabor transform, wavelet transform, MPEG7 edge histogram, etc.
③ is based on shape characteristics: such as Fourier shape descriptor, invariant moment, wavelet contour descriptor, etc.

The following main introduction of commonly used feature extraction algorithm (the first three is the Classifier Application Common 3 feature extraction algorithm ): 1, LBP feature extraction algorithm

A: LBP (local Binary Patterns, partial two value mode) is the extraction of local features as a discriminant, an efficient texture description operator, to measure and extract local texture information of the image, the illumination is invariant. There are many improved types, LBP combined with BP neural network has been used in face recognition and other fields. The basic idea of LBP is that it is defined in the 8 neighborhood of the pixel, with the gray value of the center pixel as the threshold, and the value of the surrounding 8 pixels compared with it, if the surrounding pixel value is less than the gray value of the center pixel, the pixel position is marked as 0, otherwise marked as 1. Each pixel gets a binary combination, just like 00010011. Each pixel has 8 adjacent pixels, which is a combination of 2^8 possibilities. As follows:

Improvement and optimization of LBP:

(1) Improved circular neighborhood LBP: In order to adapt to the texture characteristics of different scales and achieve the requirements of gray-scale invariance, the relative scholars have improved the LBP operator, extended the 3x3 neighborhood to any neighborhood, and replaced the Square neighborhood with the circular neighborhood. So if a point on the circle is not on the image coordinate, it is interpolated with the pixel value around it. As follows:

(2) Rotation unchanged LBP: As can be seen from the definition of LBP, the LBP operator is invariant in grayscale, but not in rotation. The rotation of the image will get different LBP values. In this paper, the LBP operator is extended and the LBP operator with rotational invariance is proposed, that is, a series of initial defined LBP values are obtained by rotating the circular neighborhood, and the minimum value is used as the LBP value of the neighborhood. The following figure shows the process:

Application: LBP is mainly used in texture classification, face analysis and so on. LBP feature Extraction is also an image of the same size, but the LBP map is not used as a feature vector for classification recognition, but the statistical histogram of LBP feature spectrum is used as eigenvector for classification recognition.

Invariance: Improved LBP has dimensional and rotational invariance.

More detailed reference: http://blog.csdn.net/SoaringLee_fighting/article/details/52699381

2. Hog feature Extraction Algorithm

A: The directional gradient histogram (histogram of oriented Gradient, HOG) feature is a feature descriptor used for object detection in computer vision and image processing. It is characterized by calculating and statistic the gradient direction histogram of local region of image. Hog feature combined with SVM classifier has been widely used in image recognition, especially in pedestrian detection. It is to be reminded that the method of pedestrian detection is HOG+SVM French researchers Dalal in 2005 of CVPR, and now although a lot of pedestrian detection algorithms continue to be proposed, but the basic is based on HOG+SVM thinking.

Invariance: Has the illumination invariance, does not have the dimension and the rotation invariance.

Application: The hog algorithm extracts the statistical histogram of each pixel gradient of the image, and usually transforms these gradient histograms into a vector for the training input of the classifier.

Specific reference: http://blog.csdn.net/abcjennifer/article/details/7365651 and Http://blog.csdn.net/soaringlee_fighting/article /details/52693843.

3. Haar Feature extraction operator

A: Often combined with adaboost to identify faces. Haar features are simple, divided into three categories: Edge features, linear features, central features and diagonal features, combined into feature templates. The feature template has both white and black rectangles, and defines the template's characteristic values as white rectangle pixels and minus black rectangle pixels and. The Haar characteristic value reflects the gray-level change of the image. For example: Some features of the face can be described by a simple rectangular feature, such as: The eye is darker than the cheek color, the nose bridge on both sides of the bridge than the color to deep, mouth than the surrounding color to be deep. But the rectangle feature is sensitive to some simple graphic structures, such as edges and segments, so it can only describe the structure of a particular trend (horizontal, vertical, diagonal).

The Haar 4 feature extraction template rectangles are shown below (the color block matrix gesture includes horizontal, vertical, oblique 45 degrees):

Specific references: http://blog.csdn.net/xizero00/article/details/46929261 and http://blog.csdn.net/xiongchao99/article/details/78807876.
4. Log feature extraction algorithm

A: LoG (dog is a first-order edge extraction) is a Higerapuras-Gaussian edge extraction algorithm, first Gaussian filtering and then the pull edge extraction. The Laplace operator is sensitive to the discrete points and noises when the image is manipulated to realize edge detection. So, firstly, the Gaussian convolution filter is used to reduce the noise, and then the Laplace operator is applied to detect the edge, the noise immunity ability of the operator can be improved, and the Gaussian-Laplace (Laplacian of Gaussian (LOG)) Edge detection operator is born.

Specific reference: http://blog.csdn.net/tonyshengtan/article/details/43794359.

5. Harris Corner feature extraction algorithm

Answer: Refer to http://www.cnblogs.com/ronny/p/4009425.html and http://www.cnblogs.com/zhchoutai/p/7182438.html

6. Sift feature extraction operator

A: The SIFT operator is an algorithm for detecting local features, which obtains the characteristics and matches the image feature points by finding the feature points and their relative dimensions and directions in a picture, and obtains good results. The SIFT feature of each feature point is a 128-dimensional vector, so the computational amount is enormous.

Invariance: With dimensional and rotational invariance.

Improved type: Pca-sift, as the name of the "principal component SIFT feature", the main extraction of 128-dimensional eigenvector of the 20 features, greatly reducing the calculation.

Specific references: http://www.cnblogs.com/liuchaogege/p/5155739.html and http://blog.csdn.net/abcjennifer/article/details/7639681.

7. Surf Feature extraction algorithm

A: Surf is an improved version of the SIFT Corner detection algorithm, mainly embodied in the speed, surf is the sift speed of 3 times times. Sift in the case of scale and rotation transformation, the matching effect is better than surf, and surf is better than the matching effect under the change of brightness. Specific reference: Http://www.cnblogs.com/tornadomeet/archive/2012/08/17/2644903.html.