KR20080020257A - Bio-image retrieval method using characteristic edge block of edge histogram descriptor and apparatus at the same - Google Patents

Abstract

A method and an apparatus for bio-image retrieval using the characteristic edge block of an edge histogram descriptor are provided to generate a characteristic vector using the characteristic edge block of an image and local dispersion and perform similarity comparison using the characteristic vector, thereby providing the location information of a main object existing in the image and improving search performance by reducing the error value of similarity comparison caused by the comparison of an edge block with no characteristic. A similarity comparison process for bio-image retrieval comprises the following steps of: collecting original bio-images and constructing a DB(Database) with bio-images generated by conversion processes(S501); storing the local and global edge histogram of reference images inputted to the DB in the DB(S502); generating a characteristic vector for an inquiry image(S503,S504); selecting the same block of the reference image existing in the DB for the selected characteristic edge block of the inquiry image to generate the characteristic vector of the reference image and accumulating the absolute error of two characteristic vector values of the inquiry image and the reference image(S505,S506); and arranging the reference images in an order in which the size of the sum of accumulated errors is arranged in a small size and providing parent reference images to a search system user as result(S507).

Description

Bio-image Retrieval Method Using Characteristic Edge Block Of Edge Histogram Descriptor and Apparatus at the same}

1 is a block diagram showing a bio-image retrieval apparatus according to the present invention;

FIG. 2 is a flowchart illustrating a process of generating a feature vector using characteristic edge blocks in the bioimage retrieval apparatus according to the present invention shown in FIG.

3 is a flowchart illustrating a process of reselecting a characteristic edge block using the local partial mountain of FIG.

4 is an embodiment illustrating a method of calculating a local partial acid of FIG. 3;

5 is a flowchart illustrating a process of performing a similarity comparison for searching for a bio image according to the present invention;

6 is a graph comparing the search performance of a bio-image according to the present invention with the search performance of an existing search based on ANMRR.

According to the present invention, a bio-image composed of a monotone background and a main object is effectively used to generate a feature vector and perform a similarity comparison using a descriptor suitable to meet the configuration characteristics of the bio-image. An apparatus and a method for improving are provided.

Moving Picture Experts Group (MPEG) is a leading organization responsible for multimedia standards. It sets standards for the compression and multiplexing of video and sound. MPEG standardization works with MPEG-1 (compression for storage media), MPEG-2 (compression for broadcast, communication, and storage media), MPEG-4 (compression for mobile communications), MPEG-7 (multimedia information retrieval), and MPEG-21. The process was divided by field of application (e-commerce).

With the development of information compression and transmission technology, more and more information is produced in multimedia form rather than text form, and the desire to find and use various multimedia information such as photographs, 3D models, audio, and video increases. As the quantity increases exponentially, it becomes more difficult to find the information desired by the user, and thus, the technology for searching and searching for the desired information easily and quickly is required. This demand and necessity led to the emergence of a standard called MPEG-7.

In the medical field, with the introduction of PACS (Picture Archive and Communication System) based on information and communication, networking, database, information management, user interface technology and information storage management technology, vast bio and medical images have been digitized. You are building a database. With the introduction of such a system, the necessity for effective retrieval of bio images having characteristics different from general natural images has emerged. Accordingly, researches applying MPEG-7 have been actively conducted.

The MPEG-7 standard standardizes descriptions (descriptions and descriptions) of various types of multimedia information and is also called a multimedia content description interface for multimedia contents. The term Description refers to pasting metadata that describes the content itself so that the required multimedia information can be found easily and accurately.

In the MPEG-7 standard, a visual descriptor has been usefully applied to natural image retrieval, and its performance has been proven. In addition, the visual descriptor may describe a unique visual characteristic of an image, and thus search performance may vary according to characteristics of the image.

Therefore, in the search for a bio image using a visual descriptor, when the image is a black and white image or a representative color varies depending on the dyeing material, the utility of the color descriptors is inferior, and in the case of an image having a complex geometry forming structure such as a neural network, Descriptors have low search performance. Among them, Edge Histogram Descriptor, which is widely used in the texture descriptor, has four types of directional edges and two non-directional edges in a local region of an image, and a spatial distribution of edges of an image. Express. In particular, when it is used for the search of the bio-image may have a high search performance for the directional image.

However, the characteristics of the bio image composed of the background and the main object, which are monotones, may cause errors due to the comparison of the blocks corresponding to the background, and thus may be a factor that degrades the search performance when the edge histogram descriptor is used.

The present invention has been made to solve the above problems, by generating a feature vector using the characteristic edge block and local partial of the edge histogram descriptor for the bio-image to be searched, by using the similarity comparison, An object of the present invention is to provide a bio image retrieval method and apparatus for improving image retrieval performance.

In order to achieve the above object, according to an aspect of the present invention, generating a feature vector of the query image, generating a feature vector of the reference image, the feature vector of the query image and the feature vector of the reference image Accumulating the sum of the absolute error of the liver, and sorting the reference image in the order of the smallest difference of the accumulated difference, and provides the upper image to the user as a result of the search for a bio-image search method.

The method may further include classifying and defining a set of ground truths related to the query image.

The method may further include constructing a database of reference images by collecting the original bio image and including the image generated through transformation of rotation, distortion, and deformation of the original bio image.

The generating of the feature vector of the reference image or the query image may include inputting a local edge histogram and a global edge histogram of the image, and an edge having the largest size among five edges of the input global edge histogram. Selecting as a representative edge, calculating a representative edge average for all blocks of the image, iterating over each block of the image, and determining the size of the edge and the representative edge of the same type as the representative edge in each block. Comparing and selecting the block as the characteristic edge block if the edge value of the compared block is larger than the representative edge average value, combining the local and global edge histograms with the edge histograms of the selected characteristic edge blocks Generating a step.

In addition, the total number of the first selected characteristic edge blocks is obtained, and when the total number of the first selected characteristic edge blocks is less than or equal to the threshold value, all of them are selected as characteristic edge blocks, and the first selected characteristic edge blocks are selected. If the total number of cross-sections exceeds the threshold, a local partial value that considers neighboring blocks of the primary selected characteristic edge block is calculated, and the second characteristic is as much as the threshold value in the order of the higher blocks in which the calculated local partial sizes are larger. And reselecting to an edge block.

In addition, calculating a local partial using an average of the four neighboring blocks of the first selected characteristic edge block and the error of the first selected characteristic edge block, and according to the size of the calculated variance And aligning the first selected characteristic edge block and re-selecting the upper block having a large variance by a threshold value in all the aligned characteristic edge blocks as the secondary characteristic edge block.

The threshold may be determined in consideration of the characteristics of the object and the background of the reference image and the number of characteristic edge blocks.

The global edge histogram may be obtained by using an average of five edges of a local edge histogram in the input image.

The method may further include generating a feature vector of the reference image by selecting a block of the reference image that is the same as the selected characteristic edge block of the query image.

According to another aspect of the invention, the object orientation determination unit for selecting the largest edge among the five edges of the global edge histogram of the input bio-image as a representative edge, and the average of the entire block of the image with respect to the representative edge And an average calculating unit for calculating a, and comparing each block of the image with edges of the same type as the representative edge and the size of the representative edge in each block, and the edge value of the compared block is larger than the representative edge average value. In this case, a feature vector is generated by combining a first selector which selects the block as a feature edge block, an edge histogram of the selected feature edge block, and a local edge histogram and a global edge histogram of the input bio image. And a bio image search including an image database storing the generated feature vector. Provide value.

In addition, the feature block number determination unit obtaining the total number of the characteristic edge blocks selected by the first selector, and when the number of the characteristic edge blocks exceeds a threshold, the average of the four blocks neighboring the characteristic edge block A local partial calculation unit for calculating a local partial mountain for the characteristic edge block by using the second partial selecting unit for reselecting the characteristic edge block by the threshold value in the order of higher blocks having the largest size of the calculated local partial mountain; It features.

The threshold may be determined in consideration of the characteristics of the object and the background of the reference image and the number of characteristic edge blocks.

The global edge histogram may be obtained by using an average of five edges of a local edge histogram in the input image.

The method may further include a similarity comparison unit configured to compare the absolute error of two feature vector values of the query image and the reference image with respect to the bio image to determine similarity.

Hereinafter, with reference to the accompanying drawings, a preferred embodiment of a bio-image retrieval apparatus and method according to the present invention will be described in detail.

1 is a block diagram showing a bio image retrieval apparatus according to the present invention.

Referring to FIG. 1, a bio image retrieval apparatus according to the present invention generates a feature vector using characteristic edge blocks of an edge histogram descriptor (EHD) corresponding to a feature of a bio image, and performs a similarity comparison. It is a device to search video.

To this end, the bio image retrieval apparatus 100 according to the present invention includes: an object orientation determination unit 110 for selecting the largest edge among the five edges of the global edge histogram as a representative edge of the corresponding bio image; An average calculator 120 that calculates an average of the entire bio-image block for the selected representative edge, and compares the size of the edges of the same type as the representative edge by traversing each divided block of the bio-image; A first selector 130 that selects the comparison block as a characteristic edge block when an edge size is larger than a size of the representative edge average, a feature block number determiner 140 that determines the number of selected characteristic edge blocks, and A local partial calculation unit 150 for calculating variance considering the characteristic edge blocks and neighboring blocks when the number of selected characteristic edge blocks is greater than a threshold; The second selector 160 which reselects the characteristic edge blocks by the threshold value in the order of the higher blocks having the largest variance of the variances, combines the edge histograms of the selected characteristic edge blocks with the local and global edge histograms. A feature vector generator 170 to generate and an image database 180 to store the generated feature vector are provided.

Here, the object orientation determining unit 110 obtains a local edge histogram from the bio image, calculates an average of five edges of the local edge histogram, and generates a global edge histogram representing statistical direction characteristics of the entire image. The local and global edge histograms are input, and the edge having the largest size among the five edges of the global edge histogram is selected as the representative edge of the bio image.

FIG. 2 is a flowchart illustrating a process of generating a feature vector using characteristic edge blocks in the bio image retrieval apparatus according to the present invention shown in FIG.

Referring to FIG. 2, for example, a corresponding bio image is divided into blocks having 4 × 4 = 16 sub-images, and two vertical, horizontal and diagonal directions (45 degrees and 135 degrees) in the divided sub images. Create a local edge histogram that extracts the non-directional edges. In addition, by calculating the average of the five edges of the local edge histogram, a global edge histogram representing the statistical direction characteristics of the 16 divided sub-images are generated, and the generated local edge histogram and global edge histogram are input to the bio image retrieval apparatus. (Step 201). In this case, selecting the representative edge to find the representative direction of the main object in the bio image is performed as in Equation 1, and selects the edge having the largest size among the five edges of the input global edge histogram as the representative edge. (Step 202).

Where G _ver , G _hor , G ₄₅ , G ₁₃₅ , and G _non are the vertical, horizontal, 45, 135, and omnidirectional global edge histograms, respectively.

When the representative edge of the bio image is selected as described above, a representative edge average of all blocks of the 16 sub-images is calculated (step 203). Each block of the sub-image is traversed to compare the size of the edge with the same type as the representative edge in each block (step 204) (step 205). As a result of the determination, if the edge value of the compared block is larger than the representative edge average value, the block is selected as the characteristic edge block (step 206), and if the value is small, the next block is compared.

Each block is traversed and compared, and then the total number of selected characteristic edge blocks is obtained and compared with a threshold value (step 207). When the total number of selected characteristic edge blocks is less than or equal to the threshold value, local and global images of the bio image are obtained. The edge histogram and the edge histogram of the selected characteristic edge block are combined to generate a feature vector and stored in the database (step 208).

However, when the total number of the selected characteristic edge blocks exceeds the threshold, the local partial value is calculated considering the neighboring blocks of the characteristic edge block, and the threshold value is set in the order of higher blocks in which the calculated local partial size is larger. Reselect as much as possible (steps 209 and 210). Then, a feature vector is generated and stored in the database in the same manner as above (combining local and global edge histograms of the input bio-image with edge histograms of the reselected characteristic edge blocks). Step 208).

In this case, the process of re-selecting the characteristic edge block using the local partial considering the neighboring block will be described in detail with reference to FIG. 3.

If the edge size of each block of the sub-image has a similar distribution, that is, if the main object does not have a prominent representative edge, the selected representative edge and the same type of edge and the total edge mean of each block have similar values. do. Therefore, the number of blocks to be selected increases, which increases the error in performing the similarity comparison, thereby degrading the search performance. In order to solve this problem, in the present invention, as shown in FIG. 3, local variance is used. The local partial acid is a variance capable of detecting a change degree (activity) of a local image, and may be used to determine a local distribution when each block of the sub image has a distribution similar to the overall mean.

Referring to FIG. 3, when specifying a threshold for the number of characteristic edge blocks, the threshold is first designated as the number of characteristic edge blocks selected from reference images stored in the database (step 301).

Since the threshold value is affected by the type and characteristics of the reference image of the database (the characteristics of the object and the background), the threshold value is considered in selecting the ear canal threshold, and in the present invention, the characteristics of the reference images are considered and selected. The average number of characteristic edge blocks is obtained and 8 is selected as the threshold.

Subsequently, when the selected characteristic edge blocks and the local edge histogram and the global edge histogram of the bio image are input (step 302), the selected characteristic edge block is selected using an average of the selected characteristic edge block and four neighboring blocks. Compute the local partial acid for the field (step 303).

The calculation method of the local partial acid will be described with reference to FIG. 4.

Referring to FIG. 4, the local partial acid does not calculate an edge average of the entire bio image, but, for example, four blocks (eg, adjacent up, down, left, and right directions of the selected characteristic edge block X (i, j)) are used. Find the edge average of X (i, j + 1), X (i, j-1), X (i-1, j), X (i + 1, j), obtain the edge mean and the selected characteristic edge block It can obtain | require by using the difference with the edge of (X (i, j)).

At this time, the method of obtaining the value of the local partial acid is performed through the equations (2) and (3).

Here, LV (i, j) refers to the average of the edges of the upper, lower, left and right blocks of the horizontal i-th, vertical j-th block, Lvar is the local partial, N is the number of each block in the horizontal, vertical direction Say.

Next, the selected characteristic edge blocks are sorted according to the largest local variance value obtained, and the upper block having the largest variance value is reselected as the characteristic edge blocks (steps 304 and 305).

At this time, the step of reselecting is repeated while comparing whether the number of characteristic edge blocks to be reselected reaches a threshold (step 306). Here, when the number of the reselected characteristic edge blocks is selected by the threshold value, the remaining selected characteristic edge blocks having the lower variance value are dropped. Then, the edge histogram of the reselected characteristic edge block and the input local and global edge histograms are combined to generate a feature vector and store it in a database (step 307).

5 is a flowchart illustrating a similarity comparison process for searching for a bio image according to the present invention.

Referring to FIG. 5, first, original bio-images are collected through animal cells, plant cells, electron microscope images, and the like, and bio-images generated through transformation, rotation, and transformation based on the collected original bio-images and the like. Build a database together (step 501).

Next, local and global edge histograms of reference images input into the database are stored in the database (step 502).

Subsequently, when the query image is input, a feature vector for the query image is generated by the method described with reference to FIGS. 2 through 4 (steps 503 and 504). In this case, in the present invention, 14 related images are selected and classified into a similar image (Ground Truth) for evaluating performance of a search for a query image.

Then, the same block of the reference image existing in the database is selected for the selected characteristic edge block of the query image to generate a feature vector of the reference image by the method described with reference to FIGS. 2 to 4, and the two features of the query image and the reference image The absolute error of the vector value is accumulated (steps 505 and 506).

At this time, the method of accumulating the absolute error of the feature vector value is performed using Equations 4 and 5.

Where N is the sum of the number of global histogram bins and the number of five histogram bins of the selected characteristic edge blocks.

here, QEHD (i) refers to the i th histogram bin extracted from the query image, and REHD (i) refers to the i th histogram bin extracted by the same process from the reference image.

Next, the reference images are arranged in the order of the smallest sum of the accumulated differences, and as a result, the upper reference images are provided to the user of the retrieval system (step 507).

6 is a graph comparing the search performance of the bio-image according to the present invention with the existing search performance on the basis of Average Normalized Modified Retrieval Rank (ANMRR).

Here, ANMRR is an index of search performance evaluation that can evaluate the ranking in consideration of error values between feature vectors. Normalized ANMRR always has a value from 0 to 1, and a value closer to 0 shows better performance. When all relevant images are searched within a specified rank, it has a value of 0. On the contrary, if there is no related image searched within the designated rank, it has a value of 1.

Referring to the results shown in FIG. 6, the biometric image search (Proposed Hist) according to the present invention shows improved search performance compared to the local Hist search results of the conventional similarity comparison using the feature vector of the local edge histogram. The results demonstrate the effectiveness of the additional location information comparison for the main objects present in the image.

In addition, we show improved ANMRR search performance in comparison with the search using the L (G / S Hist), which is generated by averaging the local edge histogram in four block units.

Therefore, through the experimental results of the search performed, according to the present invention, the search that compares the feature vectors using the characteristic edge block which reduces the comparison error of the featureless edge and provides the information about the main object is compared with the existing search scheme. In comparison, it can be seen that about 10.7% is improved.

As described above, the generation of the feature vector using the characteristic edge block of the bio image according to the present invention and the bio image retrieval apparatus and method comparing the same have the following effects.

First, by generating a feature vector using the characteristic edge block and the local partial of the image, and comparing the similarity using the feature vector, it is possible to provide the position information for the main object existing in the image, The search performance can be improved by reducing the error value of the similarity comparison.

Second, in the field of searching for bio images having different characteristics from natural images, the method of the present invention may enable the application of descriptors to match the characteristics of the background and the object of the analyzed bio image.

Third, the method of the present invention can be applied to a study for combining optimal descriptors that complexly describe the characteristics of a bio image, thereby improving search performance.

Those skilled in the art will appreciate that various changes and modifications can be made without departing from the spirit of the present invention.

Therefore, the technical scope of the present invention should not be limited to the contents described in the above embodiments, but should be defined by the following claims.

Claims (14)

Generating a feature vector of the query image; Generating a feature vector of the reference image; Accumulating the sum of absolute errors between the feature vector of the query image and the feature vector of the reference image, arranging the reference images in the order of the smallest difference of the accumulated difference, and providing the upper images as a result of the search to the user. Searching method for a bio image comprising a. The method of claim 1, And classifying and defining a set of ground truths related to the query image. The method of claim 1, And collecting an original bio-image and constructing a database of reference images by including an image generated through rotation, distortion, and transformation of the original bio-image. The method of claim 1, Generating the feature vector of the reference image or the query image, Inputting a local edge histogram and a global edge histogram of the image; Selecting an edge having the largest size among the five edges of the input global edge histogram as a representative edge, Calculating a representative edge average for all blocks of the image; Iterates over each block of the image and compares the edges of the same type as the representative edge and the size of the representative edge in each block, and if the edge value of the compared block is larger than the representative edge average value, the corresponding block as a characteristic edge block. Selecting step, And combining the local and global edge histograms with the edge histograms of the selected characteristic edge block to generate a feature vector. The method of claim 4, Obtaining the total number of the first selected characteristic edge blocks, If the total number of the first selected characteristic edge blocks is less than or equal to the threshold value, all of them are selected as characteristic edge blocks. If the total number of the first selected characteristic edge blocks exceeds the threshold, the local partials are calculated considering the neighboring blocks of the first selected characteristic edge blocks, and the order of the higher blocks having the larger size of the local partial mountains calculated. And reselecting the second characteristic edge block by the threshold value. The method of claim 5, Calculating a local partial acid using an average of four neighboring blocks of the first selected characteristic edge block and an error of the first selected characteristic edge block; Aligning the first predetermined characteristic edge block according to the calculated magnitude of variance; And re-selecting an upper block having a large variance by a threshold value in all of the aligned characteristic edge blocks into a secondary characteristic edge block. The method according to claim 5 or 6, The threshold value is determined in consideration of the characteristics of the object and background of the reference image and the number of characteristic edge blocks. The method of claim 4, wherein The global edge histogram is obtained by using an average of five edges of a local edge histogram in the input image. The method of claim 1, And selecting a block of a reference image that is the same as the selected characteristic edge block of the query image to generate a feature vector of the reference image. An object orientation determination unit for selecting the largest edge among the five edges of the global edge histogram of the input bio image as a representative edge; An average calculator configured to calculate an average of all blocks of the image with respect to the representative edge; Iterates over each block of the image and compares the edges of the same type as the representative edge and the size of the representative edge in each block. The first sorting part to select, A feature vector generator for generating a feature vector by combining the edge histogram of the selected characteristic edge block with a local edge histogram and a global edge histogram of the input bio image; And a video database for storing the generated feature vector. The method of claim 10, A feature block number discrimination unit for obtaining the total number of characteristic edge blocks selected by the first selector; A local partial calculation unit for calculating a local partial acid for the characteristic edge block by using an average of four blocks neighboring the characteristic edge block when the number of the characteristic edge blocks exceeds a threshold value; And a second sorting unit which reselects the characteristic edge blocks by the threshold value in the order of the higher blocks having the larger size of the local partial mountains. The method of claim 11, The threshold value is determined in consideration of the characteristics of the object and background of the reference image and the number of characteristic edge blocks. The method of claim 10, The global edge histogram is obtained by using an average of five edges of a local edge histogram in the input image. The method of claim 10, And a similarity comparison unit for comparing the absolute error between two feature vector values of the query image and the reference image with respect to the bioimage to determine similarity.

Images