Image retrieval based on data mining-essays

1) Image Retrieval

The algorithm extracts the underlying features of the image in the database, takes the image and the underlying features extracted as training data, and carries on the semi-supervised learning of the class area to realize the semantic Association of the image and category.

algorithm 1 semi-supervised learning of image class area

input The image data set.

Output the feature library and class area of the image.

The first step : to read the image set of images, stored in the image library. The image in the image set is preprocessed, the underlying feature of the image is extracted, and the feature library is deposited.

The second step : Compute the class Region center of each image class through the underlying features of the image set.

The third step : Judging the Image class area according to the distance from the center to the image. The user retrieves an image that can be retrieved with a sample image, or it can be retrieved using a keyword, as shown in procedure 1. Keyword retrieval method matches the keyword entered by the user to the knowledge base, finds the retrieved image category tag, returns the image in the class area, and arranges the output in ascending order with the center distance of the class region.

algorithm 2 image retrieval

input new Query Sample image Q.

Output Image Q a similar set of images and class tags.

The first step is to preprocess the sample image Q, divide the area of interest, extract the 72-D color feature vector of the image and the 28-dimensional texture feature.

The second step is to calculate the similarity of the image and various centers in the image features and features space dis (I,J). Output similar images in ascending order of similarity dis (I,J) and judge the class label of query image Q. When Q satisfies more than one class region, the class label of the image Q is selected by the query to return the highest category of the image.

The third step : The user feedback query results whether satisfied, satisfied with the image Q into the image library, its characteristics into the feature library, add the collection of image classes, and recalculate the category of the regional Center.

Figure 1 Image Retrieval System

2) Multi-source color migration

here, we take a multi-source image of the template Image color migration, Multi-source images, with said, K is the total number of source images. M is the number of segmented regions of the shape image. it guarantees that any source image should provide a reference color for at least one target area. The goal of area matching is to pick the most matching reference color from a multi-source image for any target area, including the following steps:

① Luminance remapping

In the region matching process, the brightness distribution of the color image and the pixel points of the shape image is very different, and the result of the match will be greatly affected, so we for each source image of the the channel implements the luminance remapping, the formula is as follows:

(3-1)

here, and the is the source image in the mean and standard deviation of the channel; and is the mean and variance of the target image channel. is the new matching brightness.

② pick the best matching area for

according to the above The K-mean algorithm divides color images and shape images into m-color regions. By calculating the color image area , the mean vector and the shape image area , the Euclidean distance between the color mean vectors is judged by the basis, The two regions with the smallest distance are the best match area pairs. The distance between the color image and the color mean of the shape image area is calculated as follows:

(3-2)

through the equation (3-3), we can determine the most matched area pairs , where the K candidate regions are similar to the source area, described below:

(3-3)

by collecting the pixels in the , we can get a composite source image S

③ result Combinations

after the area is matched, the color of each color image area is migrated to the corresponding matching shape image area. We calculate The transformation value of the color channel:

(3-4)

(3-5)

here,the statistics of the channel with a superscriptto indicate thatthe channel is marked with a superscriptrepresentation. ,Is the target image, respectively .,the pixel point of the channel;,target image, respectively .J Block area.,the mean value;,color images, respectively .I-block area,the mean value;,target image, respectively .J Block area.,the standard variance of the channel;,target image, respectively .J Block area.,the standard variance of the channel.

we use the mean and standard variance of all pixels on the target image and the combined source image respectively. The conversion values for the calculated channel are as follows:

(3-6)

you can combine each area color migration result image into a complete image. Finally, the synthesized image is converted back to the RGB color space.

When coloring a grayscale image, we only have a luminance channel to use. So for the pixels of the grayscale target image, we first assign the same distribution of the channel to the absent channel before the last color migration .

Image retrieval based on data mining-essays