Basis of Pattern Recognition --- mqdf classifier = mqdf improved secondary Classifier

Qdf assumes that the sample conforms to the Gaussian distribution and uses the Estimation Mean and covariance matrix to train the classifier. However, due to the high feature dimension, the time-space complexity is high. (The dimension of the covariance matrix is feature dimension * feature dimension ). In addition, the covariance matrix is often less than the rank to obtain the inverse (number of samples, feature dimension ). Mqdf has the following improvements: 1. Add a small constant to the diagonal of the covariance matrix to ensure the full rank (non-singularity) of the matrix ). 2. decompose the feature values of the covariance matrix and sort the feature values in reverse order (the EIG function of MATLAB is already in descending order by default). Select a certain number of feature values (truncation dimension ), replace small feature values with constants. This reduces the space complexity of the classifier. If the original feature dimension is too high, PCA Lda is usually used first ).

 

Classifier Training: The larger the covariance, the more information it contains, and the more accurate the final classification. The so-called secondary classifier is like a quadratic function in elementary mathematics. The covariance matrix reflects the distribution between features. Each training sample can be considered as an observation solution of the equations, for example, in elementary algebra, three independent equations are required to solve a three-element one-time equation. Therefore, only a certain number of samples can be obtained during training.

 

 

What are the solutions for multiple types of problems in OCR? (License Plate Recognition is a small Character Set Problem)

Are similar characters categorized? (Such as numbers 0 and letters O)

Http://blog.sina.com.cn/s/blog_49d1bc3601011w3w.html

Basis of Pattern Recognition --- mqdf classifier = mqdf improved secondary Classifier