Image search based on shape feature vectors

Image search based on shape feature vectors

Reprinted from: http://blog.csdn.net/huohunri2013/article/details/7965760

Certificate ------------------------------------------------------------------------------------------------------------------------------------------

Image Search items are as follows:

Simple image set: MPEG image set (which contains 20 classes and 20 binary images for each class ). Image Set features: The image is a single object, so it is easier to highlight the characteristics of the image, so that it is easier to use the evaluation of the quality of the PAF algorithm.

The following is the retrieval of MPEG image sets:

 

 

Another

 

 

Next, we combine the extracted shape feature vectors with other feature vectors (such as color and texture) to perform image search for standard image sets. Standard image set: corel5k image set features: perfect image set, taking into account color, texture, shape and other factors. Search: search interface:

 

On the search page, click submitquery image on the left. The search image page is displayed:

 

Select the image page and click one of the images:

 

You can click search in the pre-search phase to perform image search directly or enter search in the query key words on the right.

Word.

 

Joint search: Enter sun in the search term

 

The evaluation function is used to evaluate the PAF algorithm. Result chart:

The top curve in the figure is the recall-precision curve of the shape feature vector algorithm extracted from the distance between the PAF and the center of the center. The center is the center distance. At the bottom of the list is the self-generated group. The following describes the project:

The top curve in the figure is the recall-precision curve of the shape feature vector algorithm extracted from the distance between the PAF and the center of the center. The center is the center distance. At the bottom of the list is the self-generated group. The following describes the project:

Image Search is implemented based on an improved method for extracting shape feature vectors

 

I. Purpose and content

Objective: To find a better method for extracting shape feature vectors.

Work Description: The perimeter areafunction (PAF) method mentioned in the "shape retrieval using combined fourierfeatures" paper is evaluated by the recall rate and precision rate statistical function written by Jin Xin, the function used to evaluate the search results) is compared with the previous feature vector extraction method. And integrates it with other image processing technologies to achieve image retrieval.

Ii. Principle of perimeter area function (PAF)

2.1 Introduction to variables: Take n consecutive points (points) in clockwise direction on the Shape boundary ). And are the coordinates of the boundary points. N indicates the number of border points. Because the Shape Contour boundary is closed, so. The O coordinates of the center point are given by formulas and formulas.

 

 

The specific algorithm steps of the 2.2paf method are as follows:

1. Calculate the center distance for each boundary point.

2. Tracking the border counterclockwise Based on the basis point, and finding the point with the distance as the arc-length. Similarly, the point that finds the distance along the clockwise tracking boundary is recorded as (fig2 ).

Figure 1 fig2

Because the shape boundary is a digital curve, it must be located on a line segment connected to adjacent points (assumed to be points and points. If so, the abscissa and ordinate. If the horizontal and vertical coordinates of the points are calculated according to the formula and respectively, the formula is as follows :. Similarly, we can obtain the abscissa and ordinate of a point. Here, three points form a triangle, the area of the triangle (note that the order of each row of the determinant cannot be reversed)

Through the above steps, will generate a triangle area as the element sequence, where k =, 2, ·, N-1. This sequence is a brand-new feature vector (that is, PAF ). It is worth mentioning that the PAF depicts the partial information of the graphic contour boundary, each of which is obtained by the determinant, so it contains positive and negative values. The center distance not only depicts the global information of the Shape boundary but also the local information of the Shape boundary.

Iii. Project Task Allocation (I am deeply impressed by the poor project owner)

The program is divided into four parts:

Part 1: feature vector extraction ----- three feature vectors are extracted:

1. centroid distance feature vector (Fourier transformation after extraction)

2. Paf feature vectors (area feature vectors mentioned in this article, which are extracted and then undergo Fourier transformation)

3. Combine feature vectors (a hybrid feature vector consisting of the first 16 Dimensions of FFT and the first 16 Dimensions of FFT (CD)

Note that the method for extracting feature vectors is

A Represents the CD feature vector After Fourier transformation, and B represents the PAF feature vector After Fourier transformation.

Part 2: Use the three feature vectors generated in the first part for loop search and generate three search matrices respectively. Each row in the matrix represents the result sequence of an image ID Search. The result matrix generated in this section is used to count function statistics.

Part 3: Perform cyclic statistics on the search results. Different recall-precision statistical results are obtained based on the number of statistical images. For the MPEG-7 image set, you can get a 1400-to-Child recall-precision value.

Part 4: Use the recall-precision value obtained in Part 3 to describe the recall-precision statistical result graph.

Iv. Search and evaluation results

And the evaluation results are shown at the beginning of this article. The top curve in the evaluation chart represents the recall-precision of the joint feature vector (CD + PAF. The curve in the middle represents the recall-precision result graph of the center-of-center distance. The third curve represents the statistical result graph of PAF. By analyzing the result graph, we can find that we are not doing well for the moment, and there is still a gap in the results given in the paper. We hope to improve it in the future.

5. My suggestions and new ideas

(1) One of our previous ideas is feasible, that is, to think of the origin, except for finding and, we can also find the other two sets of vertices for the arc lengths (assuming they are and) respectively, so that we can form three different groups of feature vectors for searching, the search effect has been significantly improved. I have verified it in other image sets, but there is no relevant experiment in the MPEG-7 and leaf image sets, because the basic algorithms of Cd and PAF have not yet met the requirements in the paper.

(2) I have made a few guesses about why I have not achieved the effect in the paper:

1. fourier transformation is wrong. After the formation of the CD feature vector, we directly use the FFT () function in MATLAB. However, I have observed the normalization problem mentioned in this paper, it may be that we have not been normalized.

2. There may be problems with the statistical function, because the statistical function is not made by myself, but I have changed the statistical function about word search that Jin Xin has prepared to the statistical function of Graph Search. A bug may be introduced during my change.

3. After reading the MPEG-7 image set into MATLAB, I found that the image is a binary image, I directly used the tracking function to obtain the image information, and did not perform any smooth processing on the binary image.

This is the family happiness of my project team members. The first child shoes on the left are called Duan Cong (omnipotent cong), a hacker. I was abused when I was a freshman. The second one is me. At that time, I was strong and I am fat. The third is Yue qingyu, who gives a sense of sureness. The fourth is Jin Xin (he wrote the evaluation function of this project, which is mentioned below ). The last one looks awkward, but we have to count on him for a lot of things.

This is a photo of the studio. in a twinkling of an eye, we have to go to different places.