KR100323364B1 - A method for extracting characterized vectors in color image and a retrieval method using the characterized vectors - Google Patents

Abstract

The present invention provides a method of extracting color and shape features of a color image using feature vectors that simultaneously consider color and shape information in a feature extraction process, which is a preprocessing process of a content-based image retrieval system, and using these feature vectors. An image retrieval method for performing inter-image registration. The color feature extraction method of color image is achieved through non-uniform quantization of color image and generation of Neighborhoodhood of Color Histogram (NICO) matrix and Global Information of Color Histogram (GICO) matrix. Improved Invariants Moments (IMI) considering only the entire outline of the object and the invariant moments of each group of the dominant color are achieved by calculating and using the extraction method, images and query images stored in the database The image retrieval method that compares and matches the shape information by comparing the color information matching and the coefficients of the constant moments of each group of the improved constant moment and the dominant color by comparing the component values of the NICO matrix and the GICO matrix. Through the comparison values, the similarity of the query image is determined. Therefore, the present invention can achieve more improved search efficiency even when the color histogram itself is modified by the addition of noise in the image and the change of the gamma correction value.

Description

A method for extracting feature vector from color images and image retrieval method using the same {A METHOD FOR EXTRACTING CHARACTERIZED VECTORS IN COLOR IMAGE AND A RETRIEVAL METHOD USING THE CHARACTERIZED VECTORS}

The present invention relates to an image retrieval method using a feature vector that simultaneously considers color and shape information in a preprocessing process of a content-based image retrieval system, and more particularly, to a feature vector that simultaneously considers color and shape information from color images And an image retrieval method for matching color and shape features between an image stored in an image database and a query image to be retrieved using a feature vector.

Due to the recent development of multimedia technology, a large number of images are acquired and stored due to the spread of a scanner and a large storage device, which are data input devices. As a result of this, many image databases are generated and used for various purposes. Typically, these databases are composed of a large number of images, which leads to considerable difficulty in finding a desired image. Therefore, there is a need for an efficient and automatic retrieval algorithm for retrieving images from a database.

In the early retrieval system, image retrieval based on text using the input keyword has been mainstream, but this method has some problems. First, when setting a keyword to be used as a search query for an image in a database, a unique keyword may not be set because user subjectivity is included. In addition, the search has a disadvantage that the user must remember a unique keyword for a given image.

Therefore, in a more general case, an image retrieval method based on contents is required, and various methods related thereto have been studied.

A general example of a content based retrieval system is illustrated in FIG. First, in the image storage process, the scanner digitizes the input image, undergoes a preprocessing process to extract the features of the digitized image, and then indexes the features of each image to store all the images in the image database. In the image retrieval process, when a query image is input to a scanner, the image is processed through a preprocessing process, which is a feature extraction process, and compared with the features of the image stored in the image database. As a result, the final search result image is displayed. Here, the extraction and query processing of the image feature is performed through a program on a PC, for example.

On the other hand, the most important issue in the content-based retrieval method is how to select a set of features used to describe each image or object. The characteristics of the image studied so far include color, shape, and texture, among which the color characteristics are the translation, rotation, and scale of objects in the image. Because of the invariant advantage of scale, image retrieval methods based on color features have been extensively studied.

A conventional content-based image retrieval method using a method of extracting color features from a color image is as follows.

Swain et al. (MJSwain and DHBallad, 'Color Indexing', International Journal of Computer Vision, Vol. 7, no.1, page 11-32, 1991) proposed a method of color histogram intersection. In this method, 2048 groups are generated by dividing each axis into 16 groups (bins) in a color space consisting of RGB (Red Green Blue) axes, and then stored in the database by comparing the histograms corresponding to each group. An attempt was made to match an image with a query image to be retrieved.

However, this method has the advantage of simplicity of comparison, while it is sensitive to noise due to too careful grouping and also has a large amount of computation when matching.

Methri et al. (BMMehtre, MSKankanhail, ADNarsimhalu and GCMan, 'Color Matching for Image Retrieval', Pattern Recognition Lett., Vol. 16, pages 325-331, 1995) refer to the reference color table method. In this method, the color of the image was divided into 27 groups, and each group tried to match each other as the Euclidean distance. Streeter et al. (M.Stricker and M.Orengo, 'Similarity of Color Image', Proceeding of SPIE 2420, page 381-392, 1995) describe cumulative histograms to account for visual similarities not considered in the histogram intersection method. cumulative histogram) method.

However, the color features used in these methods have a relatively stable response to geometric deformation of the object in the image, but for the factors that change the color histogram itself, such as the addition of noise or the change in the gamma correction value. It causes a very sensitive reaction. In other words, when a digitalized image image of a scene or an object is obtained by a camera sensor or a scanner, it is actually obtained in different viewing conditions, so there are minute noises. It becomes difficult to play back images correctly. In particular, since the gamma value is different according to the device used to acquire the color image, the gamma value can be corrected only by correcting the gamma value. In general, image acquisition is obtained through a scanner, in which case the gamma correction value is determined by software used in the scanner. However, the gamma value correction is practically impossible because the information on the gamma correction value cannot be obtained for the already acquired image.

In addition, if a preprocessed image is stored in a database for other purposes, the gamma correction value used in this case is unknown. Therefore, it is not possible to know which gamma value the image input to the scanner is corrected. Therefore, there is a need for an image retrieval algorithm capable of operating stably to any gamma correction value. For reference, a change in the histogram of an image obtained under different gamma correction values for the same image is shown in FIG. 2.

In addition to the above factors that change the color characteristics themselves, the color information also has a disadvantage in that spatial information such as the position of an object or the configuration of an image cannot be represented.

Therefore, one of the characteristics of color information that can overcome the disadvantages of using only such a color feature is a shape feature, and by using this shape feature, the color histogram information changed by a factor that changes the color feature is somewhat compensated for. Can be. In particular, when a query target is a trademark, since a combination of simple colors and objects having various shapes is required, a combination of color information and shape information is required.

As an image retrieval method using color and shape information, Jane et al. (AKJain and A.Vailaya, 'Image Retrieval Using Color and Shape', Pattern Recognition, Vol. 29, no. 8, page 1233-1244, 1996) The color histogram intersection method which slightly modified the image feature extraction method such as phosphorus and the directional histogram intersection method for shape information were used. Jane's image retrieval method will be described with reference to FIG.

Color histograms are generated by calculating histograms for 48 groups divided by equal quantization, dividing the input image into 16 groups for each axis in the color space consisting of the R, G, and B axes. Color histories are extracted by comparing the histograms with each other, and the edges of the input image are extracted using the Canny Edge Detector, and boundary histograms are generated by calculating boundary line histograms for 36 groups. The shape features are extracted by comparing the histograms with each other, and the final search is performed through the query processing step of performing image matching by calculating the distance between the color and shape features of the images stored in the image database and the color and shape features of the query images. An image retrieval method for outputting a resultant image is disclosed. .

However, the method adds shape information to the color information of the image to increase the search efficiency. However, similar to the conventional method using the color information, a simple histogram comparison method between groups generated by a uniform quantizer is used. Because of this, there is a disadvantage in that a sensitive response is given to the factors that change the color histogram itself. In addition, since the direction histogram of the boundary line is used as a method of extracting the shape information, the movement of the object in the image does not change, but it has a disadvantage that it is very sensitive to the rotation of the object.

Meanwhile, among the most widely used invariant moments, chain codes, Zernike moments, Fourier Descriptors, etc. When using a single feature with invariant advantages in rotation, translation, scale, etc. of an object in the image, it has a comparative advantage over other methods. Since arithmetic operations are performed on all pixels corresponding to a contour, there is a problem in that a large amount of computation is performed.

A method of extracting feature vectors that simultaneously considers color and shape information from color images according to the present invention is to perform non-uniform quantization of pixels of color images with respect to each axis in HSI (Hue, Saturation and Intensity) color space. A color group of pixels is generated and a histogram for each of the generated groups is calculated, and a color information vector of a color image vector of a neighboring information of color histogram (NICO) matrix and a global information of color histogrm (GICO) matrix is extracted. Feature extraction step; Determining the number of superior colors from the NICO matrix generated in the color feature extraction step; Improved Invariant Moment (IMI) for the pixels corresponding to the contour of the object in the color image to extract the overall contour of the object, and calculate the constant moment for each dominant color based on the predetermined number of dominant colors. A shape feature extraction step of extracting individual contours of the superior color; Comparing the NICO and GICO matrix component values of the query image extracted in the color feature extraction step with the NICO and GICO matrix component values of the image stored in the image database; Comparing the IMI coefficient values and the invariant moment coefficient values for each dominant color of the query image generated in the shape feature extraction step; And determining the similarity between the query image and the image in the image database by combining the four color and shape feature comparison result values described above with each other to determine a final search image.

One object of the present invention is to provide a content-based image retrieval method that can extract and use a complex feature vector that simultaneously considers color and shape information in a preprocessing process for extracting features of an input image.

Another object of the present invention is to consider efficient color quantization using a non-uniform quantization method, Lloyd-Max Quantizer, and spatial distribution of quantized colors when extracting color features of an image. By using the NICO matrix and the GICO matrix, it is possible to flexibly cope with the factors that change the color histogram itself, such as noise addition or comma correction, which is a problem of the conventional search method. It is to provide a method for extracting color features of a color image, and an image retrieval method using the same.

Another object of the present invention is to use an improved invariant moment (IMI) that takes into account the entire contour of an object in the image, and an invariant moment comparison method that considers the shape of each group of dominant colors when extracting shape features of an image. In addition, the shape feature extraction method can be used to stabilize the rotation of an object in an image, which is a problem of the conventional retrieval method, and the overall contour can be represented with a small amount of computation. It is to provide a video search method.

1 is a block diagram illustrating a general content based retrieval system.

2 is a graph showing a change in the histogram of an image obtained under different gamma correction values in the same image.

3 is a block diagram illustrating a conventional image retrieval method using color and shape information of an image.

4 is a block diagram showing a color and shape feature extraction method of a color image according to the present invention;

5 is a block diagram of an image retrieval method utilizing the color and shape feature extraction method of FIG. 4 in accordance with the present invention.

6 illustrates an NICO matrix generation in accordance with the present invention.

7 is an exemplary view showing a GICO matrix generation in accordance with the present invention.

<Explanation of symbols for main parts of drawing>

10: video input step

20: color feature extraction step

30: Determining the Number of Dominant Collars

40: Shape Feature Extraction Step

50: GICO matrix component value comparison step

60: NICO matrix component value comparison step

70: IMI count value comparison step

80: invariant moment comparison step for each superior color

90: step of determining similarity of color and shape feature registration

100: result image

4 is a schematic block diagram illustrating a method of extracting color and shape features of a color image according to the present invention.

In the color and shape feature extraction method of the present invention, an image input step 10 in which a query image is input to a scanner and digitized, a color feature for generating and extracting color feature vector NICO and GICO matrices of an image from the input digital image signal Extraction step 20, determining the number of dominant colors from the NICO matrix generated during the color feature extraction step 30, and shape feature vectors of the image from the input digital image signal Enhanced Constant Moment (IMI) and dominant color Shape feature extraction step 40 for calculating and extracting the constant moment of the star.

6 and 7 illustrate examples of generating the NICO matrix and the GICO matrix. Looking at the color feature extraction step 20 according to the present invention with reference to Figs. 4, 6 and 7, a non-uniform quantization scheme is adopted. This approach is similar to the human visual recognition method, and is based on the quantization of the image signal in the HSI (Hue, Saturation and Intensity) color space, which is similar to the human visual recognition method. We want to create groups through quantization for each axis in HSI color space.

The color feature extraction step 20 according to the present invention comprises a color quantization step 21, a histogram generation step 22, and a NICO and GICO matrix generation step.

The color quantization step 21 is a non-uniform quantization step of generating a group by considering a probability distribution of pixels rather than a predetermined group of input image signals. As an example for non-equalization quantization, the Lloyd-Max quantizer (Lloyd) Max Quantizer). The quantizer is applied independently for each HSI axis to optimize the mean square error in each axis to enable efficient partitioning. In particular, the non-uniform quantization step was further divided because the H-axis component is more sensitive to the human visual system than the other two axes when the axis is divided in the HSI color space. As a result, all 16 groups are generated by dividing into four groups for the H axis and two groups for the S and I axes.

The histogram generation step 22 for the HSI axis is a step of calculating the histogram for the 16 groups. The ratio of pixels belonging to the divided groups for each HSI axis among all the pixels of the input image is calculated for all the pixels. It is to find the probability that the corresponding color belongs to.

In addition, the color feature extraction method of the present invention has two methods that can be distinguished even if they have similar histograms by considering the spatial distribution of unequalized quantized groups. 23) and generating a global information of color histogram (GICO) matrix (24).

First, the NICO matrix generation step 23 for calculating the ratio of pixels on the boundary of the quantized group to neighboring pixels is subdivided into the following steps.

① After a color quantization process, a labeling step for a divided group of a given image,

(2) checking the group numbers of neighboring pixels for the pixels corresponding to the boundary of each group, and calculating the number of the group numbers;

(3) calculating a ratio of pixels corresponding to a group of all pixels and substituting them into diagonal component values of the matrix;

④ calculating ratios for all pixels of neighboring groups for each group and substituting the non-diagonal component values of the matrix, respectively, and

Generating a square matrix of diagonal and non-diagonal component values of the matrix by a given number of groups.

6 shows an example of generating a NICO matrix. An example of an image after being labeled is shown in FIG. 6 (a), and the NICO matrix generated based on this is enlarged a part of the image area shown in FIGS. 6 (c) and 6 (a) to show the relationship between neighboring pixels. It shows in (b), respectively.

Referring to the calculation of the underlined component values in Fig. 6 (c), first, a diagonal component value is obtained by counting each labeled group number, calculating a ratio of the number and the number of group numbers of the entire area, and making a non-diagonal angle. The component value counts the group number of neighboring pixels, i.e., 4 and 2, 5, 6 neighboring pixels of the pixel corresponding to the border of one labeled region shown in FIG. 6 (b), and It is obtained by obtaining the ratio of the number and the number of group numbers for the whole area.

The NICO matrix generated as described above represents a histogram between groups in which the diagonal component is quantized, and when the groups of pixels in columns and rows in the image are adjacent after the non-diagonal components of the matrix are labeled, the histograms of the groups. Accordingly, the NICO matrix generated according to the present invention may represent a local spatial distribution of colors in consideration of the spatial relationship between each color group in the space in the image, that is, only the neighboring pixels.

As a result, it is possible to consider the association between each group in space while utilizing the existing histogram method through the NICO matrix as described above.

As a second matrix generation step, the GICO matrix generation step 24 using the difference between the variance value and the average value of pixel coordinates corresponding to each quantized group is subdivided into the following steps.

(1) after color quantization, a labeling step for a segmented group of a given image,

② calculating the average value and the variance value of the coordinate values of each pixel through the following equations 1 and 2 for the coordinate values of the pixels corresponding to each divided group;

Here, M _i is an average value of the coordinate values of the pixels corresponding to the i-th group, and σ ² is a variance value of the coordinate values of the pixels corresponding to the i-th group. R _i represents a set of pixels belonging to the i-th group, N _i is the number of pixels belonging to R _i , and f (k, l) is an x, y coordinate value of the corresponding pixels.

③ substituting the variance value obtained in Equation 2 into the diagonal component of the matrix,

④ By substituting the value obtained from Equation 1 for each group into Equation 3 as follows, the value normalized to the city-block distance for the average coordinate value between each group is converted to the non-diagonal component of the matrix. Assigning, and

Here, M _R (x, y) and M _C (x, y) represent the coordinate mean value of the group corresponding to the column of the GICO matrix and the coordinate mean value corresponding to the row, and N _C , N _R represent the row direction and column of the image. Indicates the magnitude of the direction.

Generating a square matrix consisting of diagonal and non-diagonal components of the matrix by the given number of groups.

7 shows an example of generating a GICO matrix. An example of a labeled image after the quantization process is shown in FIG. 7 (a), and a GICO matrix generated based on this is shown in FIG. 7 (b).

Therefore, the GICO matrix generated as described above represents a variance of the coordinate values of each group whose diagonal components are labeled so that the spatial distribution of a specific color in the image is represented. By normalizing the spatial distance between the groups of superheat to the overall image size to represent the spatial relationship between different groups, the spatial distribution of the global color can be represented because the distribution of the quantized color can be considered in the entire image. do.

Next, the shape feature extraction step 40 will be described in detail with reference to FIGS. 4, 6 and 7.

The present invention employs the following two methods of extracting shape features using invariant moments.

First, the shape feature extraction method using improved constant moment invariants (IMI), which is made by transforming existing constant moments, is largely used in the morphology calculation step (41), the object contour extraction step (42), and the IMI calculation step for the entire contour. (43).

The IMI calculation is to perform calculation on only pixels corresponding to the contour of the object. Therefore, in order to extract shape features of an object in an image using IMI, first, an outline of an object must be found. In the present invention, a morphology operation 41 is performed using a morphology operator to find the outline of an object. do. Then, the contour of the object is extracted (42) using the operation value, and then the contour of the entire object represented by seven coefficients is extracted through the IMI calculation step 43 for the extracted contour. do.

On the other hand, when IMI is used in the first shape feature extraction step, it is useful to look at the overall outline of an object in the image, but still has the disadvantage that it does not reflect the shape feature of each color. In order to examine the individual contours of colors, it is necessary to extract the constant moments of each color.

Therefore, the present invention further includes the step of calculating the invariant moment for each dominant color as the second shape feature extraction method.

An example of determining 30 the number of dominant colors and generating a comparison matrix for the dominant color is illustrated in Figs. 7 (c) and 7 (e). To determine the number of dominant colors, in the NICO matrix shown in FIG. 6 (c), if the group having the highest histogram is three arbitrarily determined by the user (FIG. 7 (e)), the GICO matrix is arranged in the order of the group having the highest histogram. After that (see Fig. 7 (C)), an example of the generated GICO matrix is shown in Fig. 7D when the number of superior colors is three arbitrarily determined by the user. By using only the superior color, it is possible to compare and analyze only the predetermined number of important color information without using all the color information in the image.

Accordingly, as shown in FIG. 4, the invariant moment for each superior color is determined 30 by the user in the color feature extraction step using the NICO matrix described above. It is generated through the invariant moment calculation 44 for each color, and the coefficient is generated for each number of the leading colors to become the number of the leading color numbers C _dom * 7. Therefore, as in the case of using the IMI, it is possible to extract a shape feature that is invariant to the rotation, movement, and scale of an object in the image by the number of colors.

Next, a color and shape feature matching method comprising a combination of two of the above-described color feature extraction method and two of the shape feature extraction method will be described with reference to FIGS. 4 and 5.

First, referring to the color feature matching method, the NICO and GICO matrices, which are the color features of the query image, are extracted from the query image, which is the searched image, through the above-described image input step and color feature extraction step. The component values of the NICO and GICO matrices of the video stored in the base are compared.

Given the weights W ₁ and W ₂ for the diagonal and non-diagonal component values of each matrix, and comparing the matrix component values of the query image with the matrix component values in the image database through Equations 4 and 5 as follows. A Euclidean distance calculation is performed to perform.

(If i = j)

(If i ≠ j)

Here, W ₁ and W ₂ are weights, Q is a query image, I is an image in an image database, and λ _ij is a component value of a NICO matrix and a GICO matrix. Is the result of the color information generated from the comparison between the component values of the NICO matrix, Is the result of the color information generated from the comparison between the component values of the GICO matrix. The weights W ₁ and W ₂ for the diagonal and non-diagonal component values of the matrix can be assigned by the user.

In the case where the weight W ₁ is greater than the weight W ₂ , the weight is given to the color composition ratio of the entire image in the image, and vice versa, the space between groups of colors of each color quantized than the composition ratio of each color in the image. More emphasis is placed on relationships. Accordingly, the efficiency of the search can be increased by appropriately giving the respective weights according to the object and the object to be searched.

In the case where the weight W ₁ is larger than the weight W ₂ , the weight distribution is given to the spatial distribution between the groups in the image, and vice versa, the weight distribution W ₁ is more important to the spatial distance between the color groups.

As a result, the color feature matching method of the present invention comprises the steps of comparing the NICO matrix and the GICO matrix component values (50 and 60) for calculating the distances of the color features of the query image and the stored image, and by the user input. It consists of weighting steps 51 and 61 where the weights W ₁ and W ₂ are multiplied.

Next, referring to the shape feature matching method, the invariant for the IMI and the superior color, which are the shape features of the query image, is passed through the above-described color feature extraction step 40 and the superior color determination step 30 from the query image as the image to be searched. When the moment is extracted, these values are compared with the constant moment values of the IMI and the dominant color of the image stored in the image database.

Comparing the IMI coefficient values of the query image and the image in the image database by Equation 6 below, and comparing the constant moment coefficient values of the superior color of the superior image of the query image and the image in the image database by Equation 7 below. Euclidean distance calculation of each coefficient value is performed.

here, , Are IMI coefficients corresponding to each query image and the image in the image database, Is the result of the shape information from the comparison between the IMI coefficients.

here, , Are Invariant coefficients corresponding to the query image and the image in the image database, respectively, and C _dom is the number of dominant colors. Is the result of the shape information from the comparison between the constant moment coefficients for each dominant color.

Accordingly, the shape feature matching method according to the present invention includes an IMI coefficient value comparison step 70 for comparing invariant moment coefficient values for shape features of a query image and a stored image, a comparison of invariant moment coefficient values for each superior color 80, and the And weighting steps 71 and 81 in which the coefficient values are multiplied by the weights w ₃ and w ₄ by the user input.

In addition, the present invention further includes a color and shape feature matching similarity determining step 90 as a step for outputting a final search result image by combining the four result values that are the color and shape feature matching results with each other.

The determining step 90 combines the similarity between the query image Q and the image H in the database. Calculated by Equation 8 below, the final search image is output based on the result.

here, , Is the result of color information, , Is computed from the result of the shape information, Represents the combined similarity of the sum of the color and shape information. In addition, w ₁ , w ₂ , w _3, and w ₄ represent weights for the NICO matrix, the GICO matrix, the IMI, and the invariant moment calculation results according to the dominant colors, respectively.

Accordingly, the color and shape feature matching method of the present invention results in a combination of the four matching methods described above, resulting in an image with high overall similarity.

In the color image retrieval method according to the present invention, a complex feature vector considering both color and shape information is used, and an efficient non-uniform quantization method, the Lloyd-Max quantizer, is used for the feature vector. In order to consider the spatial distribution between color groups generated after color quantization and quantization, the NICO matrix and the GICO matrix are applied to increase the search efficiency based on color information. In addition, by using the improved invariant moment (IMI) and the invariant moment for each dominant color to obtain shape information, the computational efficiency is greatly reduced and the search efficiency is increased.

Therefore, according to the image retrieval method according to the present invention, compared to the conventional image retrieval method, it is possible to obtain an improved retrieval result not only for the original image but also for the image modified by rotation, movement, noise addition, and gamma correction values of objects in the image. .

Claims (15)

In the method for extracting the feature vector from the color image, For each axis in an HSI color space, a color group of the pixels is generated through non-uniform quantization of the pixels of the color image, and a histogram for each of the generated groups is calculated to produce a color feature vector of the color image. Color feature extraction step of extracting NICO and GICO matrices; Determining the number of leading colors from the NICO matrix generated in the color feature extraction step; And The entire contour of the object is extracted through an improved constant moment (IMI) calculation for the pixels corresponding to the contour of the object in the color image, and based on the predetermined number of the superior colors, the invariance for each superior color. Shape feature extraction step of extracting individual contours of the superior color by calculating a moment Characteristic vector extraction method comprising a. The method of claim 1, wherein the color feature extraction step, An uneven quantization step of unevenly allocating a plurality of color groups for all pixels of the color image in an HSI color space consisting of an HSI color axis, Generating a histogram for each HSI axis by calculating a probability that a pixel belonging to each of the color groups allocated for the HSI axis belongs to all pixels of a color corresponding thereto; and Generating a NICO matrix indicating a ratio of neighboring pixels on the boundary of the quantized color group and a GICO matrix indicating a ratio of pixels corresponding to each of the quantized color groups in a full color image; Characteristic vector extraction method comprising a. The method of claim 1, wherein the extracting the shape feature comprises: A morphology calculation step of performing a morphology operation using a morphology operator to find the contour of the object in the color image; Extracting an outline of the object in a color image from the morphology calculation result; Extracting an overall contour of the object by calculating an improved constant moment (IMI) using pixels corresponding to the contour of the object, and Calculating a predetermined invariant moment for each dominant color Characteristic vector extraction method comprising a. 3. The method of claim 2, wherein the non-uniform quantization step uses an Lloyd-Max Quantizer for non-uniform quantization. 3. The method of claim 2, wherein the histogram generation for each axis of the HSI is performed by assigning four color groups to each of the H-axis and two each to the S-axis and the I-axis. . The method of claim 2, wherein the generating of the NICO matrix comprises: labeling a divided group of color images after the color quantization; Identifying a group number of neighboring pixels for pixels corresponding to the boundary of each color group, and counting the number of the group numbers; Substituting a ratio of pixels corresponding to a group to which all pixels belong to a diagonal component value of the matrix; Substituting the non-diagonal component values of the matrix with respect to the total pixels of neighboring groups for each group, and Generating a square matrix of diagonal and non-diagonal component values of the matrix by a given number of groups. The method of claim 2, wherein the generating the GICO matrix, After the color quantization, labeling a segmented color group of a given image; Calculating an average value and a variance value of the coordinate values of the pixels based on the coordinate values of the pixels corresponding to the divided groups; Substituting the calculated variance into the diagonal component values of the matrix; Substituting the normalized value of the non-diagonal component value of the matrix with a city-block distance for an average coordinate value between groups; and And generating a square matrix composed of diagonal and non-diagonal component values of the matrix by the given number of groups. The method of claim 7, wherein the average value and the variance value are calculated by Equations 1 and 2, wherein Equations 1 and 2 are as follows. [Equation 1] [Equation 2] Here, M _i is an average value of coordinates of pixels corresponding to the i th group, σ ² is a variance value of coordinate values of pixels corresponding to the i th group, and R _i is a set of pixels belonging to the i th group N _i is the number of pixels belonging to R _i , and f (k, l) is the x, y coordinate value of the corresponding pixels. The method of claim 7, wherein the value of normalizing the city-block distance is obtained from Equation 3, wherein Equation 3 is as follows. [Equation 3] Here, M _R (x, y) and M _C (x, y) represent the coordinate mean value of the group corresponding to the column of the GICO matrix and the coordinate mean value corresponding to the row, and N _C , N _R represent the row direction and column of the image. Characteristic vector extraction method, characterized in that indicating the size of the direction. 4. The method according to claim 3, wherein the constant moment for each superior color has a coefficient generated for each superior color so that the number of superior colors C _dom * 7 is obtained. An image retrieval method for performing matching of a query image to be retrieved using a feature vector in a color image and an image stored in an image database, For each axis in an HSI color space, a color group of the pixels is generated through non-uniform quantization of the pixels of the color image, and a histogram for each of the generated groups is calculated to produce a color feature vector of the color image. Color feature extraction step of extracting NICO and GICO matrices; Determining the number of leading colors from the NICO matrix generated in the color feature extraction step; The entire contour of the object is extracted through an improved constant moment (IMI) calculation for only pixels corresponding to the contour of the object in the color image, and based on the predetermined number of the superior colors, for each superior color A shape feature extraction step of extracting individual contours of the superior color by calculating invariant moments; Comparing the NICO and GICO matrix component values of the query image extracted in the color feature extraction step with the NICO and GICO matrix component values of the image stored in the image database; Comparing the IMI coefficient value and the invariant moment coefficient value for each of the dominant colors with the IMI coefficient value and the invariant moment coefficient value for each dominant color of the image stored in an image database; And Color and shape feature matching similarity determination step of determining the final search result image by determining similarity between the query image and the image in the image database by combining the four color and shape feature comparison result values. Image search method comprising a. The method of claim 11, wherein the NICO and GICO matrix component values are compared by Euclidean distance calculation of the component values of the respective matrices by Equations 4 and 5. Is [Equation 4] (i = j) [Equation 5] (i ≠ j) Where W ₁ and W ₂ are weights for the diagonal and non-diagonal components of each matrix, Q is the query image, I is the image in the image database, and λ _ij is the component value of the NICOO matrix and the GICO matrix. Indicates Is the result of the color information generated from the comparison between the component values of the NICO matrix, Is a result of the color information generated from the comparison between the component values of the GICO matrix. 12. The method of claim 11, wherein the comparison of the IMI coefficients and the constant moment coefficient values for each dominant color is performed through Euclidean distance calculation for each of the coefficient values according to Equations 6 and 7. Equations 6 and 7 are [Equation 6] [Equation 7] here, , Are IMI coefficients corresponding to each query image and the image in the image database, Is the result of the shape information from the comparison between the IMI coefficient values, , Are constant moment coefficients corresponding to the query image and the image in the image database, respectively, where C _dom is the number of dominant colors. Is a result of shape information resulting from comparison between invariant moment coefficient values for each dominant color. 12. The method according to claim 11, wherein before the user determines a final search result image by combining the four color and shape feature comparison result values with each other, each weight w _{1 for the} user to arbitrarily determine the importance of the comparison result value. and giving w ₂ , w _3, and w ₄ . The method according to any one of claims 11 to 14, wherein the color and shape feature matching similarity determination for determining the final search result image by combining the four values that are the result of the color and shape feature comparison are determined by Equation (8). As an image retrieval method performed, Equation 8 is as follows. [Equation 8] here, , Is the result of color information generated from the comparison of NICO and GICO matrix component values, , Is the result of the shape information generated from the comparison between the IMI coefficient values and the invariant moment coefficient values by the dominant color, and w ₁ , w ₂ , w ₃ and w ₄ are the invariants by NICO matrix, GICO matrix, IMI and dominant color, respectively. Represents the weight for the moment calculation result, Is an integrated similarity sum of color information and shape information.