"2018.5.27 meeting Record"--[algorithm principle]: The concept problem of manual feature extraction.

1, the difference between extracting feature points, feature descriptors and extracting eigenvector:

(1), feature points : refers to a picture on the more representative of the ' location ', extracting feature points is the picture of these representative position to mark out.

　　

(2), feature description : When the feature point is extracted, since the feature point is a position of the picture, in order to be able to perform mathematical calculations, we need to give these "positions" in a mathematical way to describe, so you can use a vector v to represent each position. And this vector is called the feature descriptor .

(3), eigenvector : Refers to the image by doing some transformations on the pixel level (LBP, color characteristics, Sift, surf), the generation of a vector V, with this vector can represent the entire picture, this vector is called the eigenvector .

2, understand the relationship between sift image retrieval and clustering algorithm.

(1) Sift feature extraction process:

① Image Description Sub-establishment : A picture of the feature point extraction (find a representative position in the picture), and each feature in this picture is described by the description of the description (set this image has 1000 features, each descriptor is a 128-dimensional vector).

　　　　　　

The core idea of the algorithm: 73740612

② Introduction of the word bag model : 51475550

-now extracts all the feature points (with 50 images) of all the images in the image library (with each feature point number 1000) and creates descriptors (so that there are a total of 1000 descriptors, each with a sub-dimension of 128). And these 1000 descriptors can be called "words."

-now introduce the K-means clustering algorithm to divide the 1000 descriptors into 64 classes, that is, set the cluster center to 64, and then learn which of these x 1000 descriptors should be in each category.

-so that each of these 64 classes has a > different number of;> from different pictures.

-When clustering is complete, each dimension of the 64 dimension is equivalent to a description sub-set, which refers to each dimension as a " word bag ".

　　　　　　

③ the formation of eigenvectors:

-Create a corresponding 64-dimensional histogram for the 64-word bag.

-Take a look at all the feature descriptors in a picture corresponds to the 64-dimensional word bag which word bag, as long as a certain dimension of the word bag contains a feature description of the picture, then the corresponding dimension on the histogram is added one, know that the image of all the feature points in the description are traversed completely. So the final 64-dimensional histogram is the eigenvector of this image.

　　　　　　

(2), Image retrieval: That is, compare the distance between each image in the image library to be retrieved and set a threshold , which is less than the threshold value of the image retrieved.

"2018.5.27 meeting Record"--[algorithm principle]: The concept problem of manual feature extraction.