Image retrieval based on Gabor texture feature

one, Gabor textureIn the image retrieval, the textures of different images differ greatly in different roughness and directionality. The Gabor filter has the flexibility and directionality to select the need to use a single filter that cannot satisfy different types of texture analysis at the same time. Gabor texture features are widely used in texture analysis, face recognition and image retrieval by using Gabor filtering on different scales and different directions, and then feature extraction of filtering results.
Gobar Filter

Gabor textures are obtained and computed by Gabor filtering. The 2-D Gabor function is a complex sine curve modulated by a Gaussian function, which is expressed as:

Among them, Σ is the standard deviation of the Gaussian function, which is wavelength and direction respectively. Specify different filters for different directions and sizes.

Gobar Texture Representation

In order to extract the Gabor texture better, we filter the image in 4 different scales and 6 different directions respectively. The energy values in each scale and direction can be obtained after filtering. We use the mean and mean variance of the energy values of all the filtering results to describe the texture information. Assuming the energy values of the filter and image convolution, the energy mean and mean variance are:

second, the search experiment 1. Image Sample

The image database used in the experiment has 350 images, a total of 7 categories. Each class is a common activity image. Its specific composition is shown in table 1.

The composition of the image database

Kinds	Number
Filament	50
Sunspot	50
Active region	50
Coronal Jet	50
Emerging Flux	50
Flare	50
Oscillation	50

An example of each type of image in the database, as shown in the following figure.

Figure 1 Sample Library Image example 2, search experiment

First, all the images in the database are feature extracted, and the eigenvector database is built to save to local backup. Secondly, the feature extraction of the query image. Finally, based on the selected three distance calculation methods, the similarity measure of the query image and all the images in the database is made, and the results are returned and queried. The test set has 70 images, which correspond to 7 categories in the database, and 10 test images for each category. One query is shown in Figure 2. In Figure 2, the first image of the first row is the query image, the rest is the returned result of the retrieval, and the results are sorted according to the decreasing similarity. The number of images returned each time is 20 sheets.

Figure 21 Retrieval process