Machine Vision machine learning weapons Library

Record some knowledge points, anyway, this is the draft version. A lot of things are not carefully studied. A lot of things are a pitfall first. Let's take a look at it later!

 

1. GMM: Gaussian mixture model. As a more universal statistical model.

2. Markov Random filed: Markov Random Field. Markov describes a kind of action in the world. Future events are only related to the current status and are irrelevant to the past. A random field provides a place for application. It is similar to a place where each small area is located, and a phase space is located at each location. You can choose which crops to plant. An image is actually a random field. Each pixel is a piece of land, and the value obtained on each pixel belongs to phase space. The most closely related relationship between an image and each pixel is its neighbor. Therefore, MRF can be used to model many problems on the image.

3. Bayes estimation. Use data to estimate the parameters in the model. The maximum likelihood estimation produces the best estimation results. From the transformation of the anterior probability to the posterior probability, I feel that this tool has changed the idea of solving the problem just like the transformation from algebra to equations in elementary school.

In Liu weipeng's blog there is a very popular introduction of Bayes, it is interesting that he mentioned why we often use L2-norm to measure the distance between data, the answer is that most often this is determined by the normal distribution of data.

4. L1-norm. The recent boom in the release of Buddha seems to have a lot to do with the sparse matrix solution. Sparse Matrix has played a major role in face recognition. msra's Ma Yi uses sparse matrix to produce a huge result for face recognition.

5. . In the solution of many models, smooth convergence is required. In this case, we can use a pair of unique parameters.

 

6. PCA and principal component analysis are the main contradictions.

7. Support Vector Machine (SVM), which separates two types of data to the maximum extent.

8. SVD and matrix decomposition. One application is to convert a large matrix into two small matrix operations.

9.