KR101539058B1 - Image classfication and detection apparatus and method thereof - Google Patents

Abstract

Disclosed are an image classification and search apparatus and a method thereof in the present specification and, more specifically, an image classification and search apparatus capable of obtaining a high search rate and a method thereof. The disclosed image classification apparatus in the present specification comprises: an input unit for receiving an image; an edge analysis unit for acquiring an edge histogram by extracting only an edge area from a rotation-invariant local binary pattern (LBP_ROT) with respect to the image; a color analysis unit for extracting a color histogram by converting the image into a hue-saturation-lightness (HLS) color space; and a storage unit for storing the edge histogram and the color histogram.

Description

[0001] IMAGE CLASSIFICATION AND DETECTION APPARATUS AND METHOD THEREOF [0002]

The present invention discloses an image classification and search apparatus and a method thereof, and more particularly, to an image classification and search apparatus and a method thereof capable of obtaining a high search rate.

With the rapid development of smart mobile environment, mobile shopping industry is growing day by day. In particular, there is an increasing need to search for visual searches that meet the needs of various users, especially those that are not stereotyped, but are not stereotyped, such as clothing. Therefore, the development of descriptors that improve the recognition of non - standardized objects by developing core technologies and application technologies related to them, reducing image recognition errors caused by existing image - visual retrieval techniques, and reducing the computation amount thereof are attracting attention.

When searching for clothes images, consumers mostly rely only on text-based keyword searches when searching for clothes, and techniques for searching clothing by inputting actual images themselves have not been developed yet. In addition, the currently developed apparel search technology is not yet applied to apparel search because its performance is insufficient yet.

SUMMARY OF THE INVENTION An object of the present invention is to provide an image classification and retrieval apparatus and a method thereof that can obtain a high retrieval rate. In particular, it is an object of the present invention to provide an image classification and searching apparatus and method which can obtain a high search rate even in an irregularized image.

The image classifying apparatus disclosed in this specification includes an input unit for receiving an image, an edge analyzing unit for extracting an edge histogram by a rotation binary invariant (LBP_ROT) method with respect to an edge region of the image, To a HLS (Hue, Saturation, Lightness) color space to extract a color histogram, and a storage unit for storing an edge histogram and a color histogram.

The image classification method disclosed in this specification includes the steps of receiving an image, extracting an edge histogram by a rotation invariant local binary pattern (LBP_ROT) method for an edge region of an image, converting an image into an HLS color space, Extracting the edge histogram, and storing the edge histogram and the color histogram.

The image search apparatus disclosed in this specification includes an input unit for receiving a test image, an edge analysis unit for extracting an edge histogram by a rotation invariant local binary pattern (LBP_ROT) method for the edge region of the test image, And a determination unit for comparing the edge histogram and the color histogram with the edge histogram and the color histogram for the pre-stored training image to determine the similarity.

The image retrieving method disclosed in this specification includes the steps of receiving a test image, extracting an edge histogram by a rotation invariant local binary pattern (LBP_ROT) method for the edge area of the test image, converting the test image into an HLS color space Extracting a color histogram, and comparing the edge histogram and the color histogram with the edge histogram and the color histogram for the pre-stored training image to determine the similarity.

According to the inventions disclosed in this specification, a high search rate can be obtained in image search. In particular, it is possible to achieve a high retrieval rate even in the non-standardized image, and it is possible to obtain a high retrieval rate in spite of the scale change or the rotation change of the image, but also the similarity of the important color in image retrieval is not lost.

1 is a diagram for explaining an image classification apparatus disclosed in the present specification.
2 is a diagram for explaining a local binary pattern.
3 is a diagram for explaining LBP_ROT.
FIG. 4 is a histogram showing the result of extracting LBP_ROT for one image and a left image of the same pattern rotated by 90 DEG.
FIG. 5 is a diagram for explaining a problem about limitations of image enlargement and reduction. FIG.
FIG. 6 is a histogram showing the feature vector of LBPROT_35 after excluding the flat region from the analysis object through edge detection.
FIG. 7 shows an example of a color histogram extracted through the color analysis unit disclosed in this specification.
8 is a diagram for explaining an image analysis method disclosed in this specification.
9 is a diagram for explaining an image search apparatus disclosed in the present specification.
10 is a diagram for explaining an image analysis method disclosed in this specification.
11 is a view for explaining an embodiment of the image classification and search method of the present specification.
12 and 13 show the results of comparing the image classification and search method of the present invention with the conventional technology.

In order to fully understand the present invention, operational advantages of the present invention, and objects achieved by the practice of the present invention, reference should be made to the accompanying drawings and the accompanying drawings which illustrate preferred embodiments of the present invention.

Hereinafter, the present invention will be described in detail with reference to the preferred embodiments of the present invention with reference to the accompanying drawings. However, the present invention can be implemented in various different forms, and is not limited to the embodiments described. In order to clearly describe the present invention, parts that are not related to the description are omitted, and the same reference numerals in the drawings denote the same members.

Throughout the specification, when an element is referred to as "including" an element, it does not exclude other elements unless specifically stated to the contrary. The terms "part", "unit", "module", "block", and the like described in the specification mean units for processing at least one function or operation, And a combination of software.

In this specification, an image classification and search apparatus and a method thereof are disclosed. The image classifying and searching apparatus and method disclosed in the present specification can be applied to various kinds of images and have a greater effect on images having irregular characteristics such as clothes. Particularly, it is possible to obtain a high search rate even when clothes are rotated or change in size when searching for non-standard clothes, and by implementing a color classification using a method of quantizing color spaces, .

Hereinafter, a detailed description will be given with reference to the drawings.

1 is a diagram for explaining an image classification apparatus disclosed in the present specification.

Referring to FIG. 1, an image classification apparatus includes an input unit 101 for receiving an image, an edge region extracting unit 102 for extracting only an edge region from a local Binary Pattern with ROTation-invariant (LBP_ROT) An edge analyzing unit 102, a color analyzing unit 103 for converting an image into a HLS (Hue, Saturation, Lightness) color space and extracting a color histogram, and a storage unit 104 for storing an edge histogram and a color histogram . Here, the image includes the training image, and the histogram data generated by the image classification apparatus can be used for image retrieval in the image retrieval apparatus. On the other hand, an image may include various kinds of images such as clothing images.

First, the edge analyzing unit 102 will be described. Local Binary Pattern (LBP) is a method of using binary values extracted through comparison of size between one pixel of an image and neighboring neighboring pixels.

2 is a diagram for explaining a local binary pattern.

Referring to FIG. 2, LBP represents a value of '1' for a pixel corresponding to a value equal to or greater than a value corresponding to the value of the center pixel in FIG. 2 (a) (Not shown in FIG. 2 (b)), and a value of '0' is given to the pixels which are not in the predetermined direction (refer to FIG. 2 b) in the clockwise direction). The LBP value in FIG.

3 is a diagram for explaining LBP_ROT.

The LBP values are divided into 256 types ranging from 0 to 255. Here, however, a pattern that is duplicated by rotation can occur.

Referring to FIG. 3, the pattern shown on the left and the pattern shown on the right are LBPs rotated by 45 degrees with respect to each other.

For example, as shown in Figure 2, LBP = 53 and LBP = 106 correspond to a 45 ° rotation of each other. In this way, it can be treated as the same object for LBP values constituting the same structural form through rotation.

Considering the rotation of 256 kinds of LBP values extracted in this manner, it is possible to reduce to 36 types, which is called LBP_ROT.

LBP_ROT is a method that considers rotation conversion in the characteristics of LBP, and bundles objects that have common patterns in existing binary patterns.

FIG. 4 is a histogram showing the result of extracting LBP_ROT for one image and a left image of the same pattern rotated by 90 DEG.

Referring to FIG. 4, it is confirmed that the LBP_ROT histogram of each of the rotated image and the non-rotated image is extracted in the same or similar manner. However, there is a limitation in LBP_ROT which can not solve the problem of size change due to image enlargement / reduction. Particularly, the problem of the limitation of the enlargement and reduction is a very important factor in relation to the non-standard garment search, which must be solved in the non-standard garment search field.

The problem of the limits of such enlargement and reduction is due to the difference in characteristics of the edge and flat areas of the image. The edge region exists in a linear form in the image region and has a one-dimensional characteristic. Conversely, the flat region exists in a plane form and has a two-dimensional characteristic. Therefore, assuming that the image has been enlarged twice, the edge area is increased by 2 times in proportion thereto, but the flat area is increased by 4 times in proportion to the square of the enlargement ratio.

FIG. 5 is a diagram for explaining a problem about limitations of image enlargement and reduction. FIG.

5, if the histogram is extracted through the LBP_ROT analysis for the original image on the left side and the image obtained by enlarging the original image twice, the feature vectors mainly extracted in the edge region are doubled, but the feature vector extracted mainly in the flat region Is increased by 4 times. As a result, a large change is made in the histogram composed of the ratio of extracted feature vectors, and LBP_ROT is sensitive to the magnitude change due to the enlargement and reduction of the original image.

 Therefore, in order to solve this problem, only the edge region corresponding to the elements necessary for the actual pattern analysis is considered as the extraction target of the LBP_ROT region. Extracting the edge region from the image can use various methods. This method also includes a Canny Edge Detection algorithm.

Only the edge region remains in the image extracted by the kenny edge detection. This removes one LBP_ROT value in the flat region from the histogram of the feature vectors of LBP_ROT, thus eliminating one of the 36 LBP_ROT feature vectors to extract 35 feature vectors. The method of extracting 35 feature vectors by removing feature vectors corresponding to the flat regions in 36 feature vectors of LBP_ROT is called LBPROT_35 in the present invention. This makes it possible to acquire excellent feature information and to minimize distortion due to the scale change of the original image by concentrating attention on the edge portion where the feature of the signal is compressed.

FIG. 6 is a histogram showing the feature vector of LBPROT_35 after excluding the flat region from the analysis object through edge detection.

Referring to FIG. 6, the influence of the size change is minimized, thereby improving the retrieval performance. In particular, the irregular retrieval performance can be maximized.

Next, the color analysis unit 103 will be described. Color recognition for images is a necessary process in constructing clothing matching algorithms as well as pattern recognition. It is necessary not only to analyze the color characteristic of the monochromatic image but also to comprehensively include all the various color distributions generated by the pattern. Considering the diversity of color, users should be able to accurately present the image of the desired color.

In addition, in order to increase the accuracy of color recognition, consideration should be given to changes in external light or shading of the image. In other words, it should be designed so that the change in brightness and shadows caused by light or shadows in everyday life rather than extreme contrast changes will not adversely affect the retrieval ability.

Accordingly, the color analyzer 103 described in the present specification extracts the color histogram by a method of dividing the converted HLS color space into a predetermined number of colors.

For this purpose, the image represented by the existing RGB color space is converted into the HLS color space in consideration of the contrast. The HLS color space is a method of representing colors as RGB, and expresses colors using coordinates of hue, saturation, and lightness.

The color (Hue, H) is the distribution of the relative arrangement angle when 0 ° is the longest wavelength in the color circle in which the visible ray spectrum is arranged in a ring shape. Therefore, the H value has a range of 0 ° to 360 °, and 360 ° and 0 ° indicate the same color red. Saturation (S) indicates the degree of saturation when the darkest state of a particular color is 100%, and 0% indicates achromatic regardless of color. Lightness (L) is the degree of lightness and darkness of a color, which is directly related to the change in lightness and darkness of the actual image.

Images converted from the RGB color space to the HLS color space can be expressed by quantifying the degree of brightness. Therefore, if the lightness or shadow of the image causes a change of light or shade, the change of the brightness (L) value occurs largely and the color (H) and the saturation (S) hardly change.

In order to utilize this point, the color analyzer 103 may divide the converted HLS color space into a predetermined number of colors, and the division may be performed within a predetermined brightness (L) range. To do this, quantize the HLS space into several colors and set a threshold value for the brightness (L), so that a color change that becomes extremely bright or dark beyond a natural brightness change is not due to illumination change, And black.

For example, black and white threshold values are independent of color (H) and saturation (S). Threshold values are set so that they are black when the brightness (L) is 15% or less and white when the brightness is 80% . The remaining brightness (L) is divided into red, orange, yellow, green, light blue, blue, violet, pink and gray for colors having a value between 15% and 80%. On the other hand, a color having a saturation (S) of 10% or less within a limited brightness (L) can be considered as a gray color, and thereafter, a color with a limited brightness (15% <L <80%) and a limited saturation (H) is between 344 ° and 360 ° and 0 ° and 9 ° is red, between 10 ° and 37 ° is yellow, between 38 ° and 65 ° is yellow, between 66 ° and 145 ° is green and between 146 ° and 181 ° is light , Blue for 182 ° to 255 °, blue for 256 ° to 315 °, and pink for 316 ° to 344 °. Thus, if each of these colors is a dimension of one feature vector, the distribution of 11 colors can be analyzed for the subsequent image, and then the ratio of specific colors in the whole image can be represented by a histogram.

FIG. 7 shows an example of a color histogram extracted through the color analysis unit disclosed in this specification.

7 is a histogram of the results obtained by quantizing each of the left images with 11 colors. Finally, the distribution of colors in the histogram extracted from the individual images can be sequentially read from the left color direction to the right color direction, so that a color descriptor, which is a 11-dimensional feature vector, can be created.

The histogram according to the edge analyzing unit 102 and the color analyzing unit 103 described above is stored in the storage unit 104. [ The stored histogram can be used as training image data in searching the actual image. For example, when performing a search on a test image, the edge histogram and the color histogram for the test image may be compared with the edge histogram and the color histogram of the training image data stored in the storage unit 104 to find a similar image .

The image classification apparatus disclosed in this specification may further include a Dense-SIFT analysis unit 105 for extracting a Dense-SIFT (Scale Invariant Feature Transform) histogram for an image. In this case, the storage unit 104 stores the Dense- You can save the SIFT histogram. SIFT means to extract feature vectors for a local patch centered on each feature point after selecting feature points that are easy to identify, such as corner points in the image, and the Dense-SIFT analysis unit 105 calculates 128 Dimensional SIFT vectors, classify them by performing K-means clustering, build a KD tree, which is a data structure that can index multi-dimensional point data, and classify bin) is extracted from the histogram.

The histogram according to the Dense-SIFT analysis unit 105 described above may be stored in the storage unit 104 together with the histogram according to the edge analysis unit 102 and the color analysis unit 103, Can be used as training image data.

8 is a diagram for explaining an image analysis method disclosed in this specification.

Referring to FIG. 8, the image analyzing method includes receiving an image (S801), extracting only an edge region from an LBP_ROT for the image (S802), converting the image into an HLS color space, Extracting an edge histogram (S803), and storing an edge histogram and a color histogram (S804).

The image classification method may further include a step of extracting a Dense-SIFT histogram for the image (S805), and a Dense-SIFT histogram may also be stored in the storing step (S804).

The edge region can be extracted by the kenny edge detection algorithm.

The color histogram can be extracted by a method of dividing the converted HLS color space into a predetermined number of colors, wherein the division can be performed within a predetermined brightness (L) range.

Here, the image may include a garment image.

Details of other image analysis methods are the same as those of the above-described image analysis apparatus, and will not be described here.

9 is a diagram for explaining an image search apparatus disclosed in the present specification.

Referring to FIG. 9, the image search apparatus includes an input unit 901 for receiving a test image, an edge analyzing unit 902 for extracting only edge regions from LBP_ROT for a test image and obtaining an edge histogram, And a determination unit 904 for comparing the edge histogram and the color histogram with the edge histogram and the color histogram of the pre-stored training image to determine the similarity. Here, the determination unit 904 may output an image corresponding to the test image based on the similarity of the edge histogram and the color histogram. The edge histogram and the color histogram for the training image may be stored in the storage unit 905.

The image search apparatus may further include a Dense-SIFT analysis unit 906 for extracting a Dense-SIFT histogram for the image, and the determination unit 904 may include a Dense-SIFT histogram and a training image Dense- The similarity can be judged by comparing the SIFT histogram. Here, the determination unit 904 may output an image corresponding to the test image based on the similarity of the edge histogram, the color histogram, and the Dense-SIFT histogram.

The determination unit 904 determines whether the histograms of the test image and the training image are similar or not by calculating a score for each histogram using a histogram intersection method for each of the edge histogram, color histogram, and Dense-SIFT histogram. , Then multiplying the score of the three calculated histograms and selecting the candidate with the highest value as the result.

On the other hand, the Dense-SIFT histogram can be extracted using the K-D tree generated at the extraction of the training image Dense-SIFT histogram.

The edge region can be extracted by the kenny edge detection algorithm.

The color histogram may be extracted by dividing the converted HLS color space into a predetermined number of colors, and the division may be performed within a predetermined brightness (L) range.

The test image may include a garment image, and the image search apparatus may further include a garment patch extraction unit 907 that extracts the garment image from the input image and outputs the garment image to the input unit 901.

The description of the edge analyzing unit 902, the color analyzing unit 903, the dense-SIFT analyzing unit 906, and the like will be omitted here.

If the test image in the image analysis device is a garment image, the input image may be an image that includes objects other than garment. Therefore, in the case of a portrait photograph, after detecting a person by using face detection in order to extract a clothing patch by the clothing patch extracting unit 907, a patch window of a certain size of the clothing area is opened, You can use a method to automatically extract patches. In this process, Viola-Jones' face recognition algorithm, which is a well-known method, can be used, and a certain portion of the upper half of the body can be extracted using the face ratio and inputted as a test image.

10 is a diagram for explaining an image analysis method disclosed in this specification.

10, in the image analysis method, a test image is inputted (S1002), an edge histogram is obtained by extracting only an edge region from an LBP_ROT for a test image (S1003), a test image is converted into an HLS color space (S1004) of extracting the color histogram, and comparing the edge histogram and the color histogram with the edge histogram and color histogram of the pre-stored training image to determine the similarity (S1006).

The image search method may further include a step of extracting a Dense-SIFT histogram for the image, and the determining step S1006 may include a Dense-SIFT histogram of the Dense-SIFT histogram and a training image of the pre-stored training image. The degree of similarity can be determined. Here, in the determining step S1006, an image corresponding to the test image may be output based on the similarity of the edge histogram, the color histogram, and the Dense-SIFT histogram.

As described above, the histogram intersection method can be used to determine whether the histogram of the test image and the training image is similar or not.

On the other hand, the Dense-SIFT histogram can be extracted using the K-D tree generated at the extraction of the training image Dense-SIFT histogram.

The edge region can be extracted by the kenny edge detection algorithm.

The color histogram may be extracted by dividing the converted HLS color space into a predetermined number of colors, and the division may be performed within a predetermined brightness (L) range.

The test image may include a garment image, and the image retrieval method may further include a garment patch extraction step (S1001) for extracting and outputting a garment image from the input image.

Among the detailed descriptions of the image retrieval method, the description overlapped with the description of FIG. 9 will be omitted.

11 is a view for explaining an embodiment of the image classification and search method of the present specification. Figure 11 (a) illustrates one embodiment of a block diagram for image classification, and Figure 11 (b) illustrates one embodiment of a block diagram for image retrieval.

11A, when the training image (S1101: Training Image) is inputted, the image classification is performed in parallel with the Dense-SIFT (S1104), edge analysis (S1102) of LBP_ROT and color analysis (S1103) based on HLS And a histogram is created for each.

In the present specification, a total of 810 pieces of clothing patch images were uniformly set to a size of 100x100 pixels and used as a training image in one embodiment. Apparel patch images may include atypical apparel patch data obtained by arbitrarily giving a rotation and scale transformation of 10 pieces each from 36 specially produced clothes.

In the case of the Dense-SIFT analysis (S1104), 128-dimensional SIFT vectors are obtained from the patches created from training images, classified by performing K-means clustering, and a KD tree is constructed based on the K- Create a histogram of bin values for class values.

In an actual application example of the Dense-SIFT analysis (S1104), 32x32 windows are moved in units of 4 pixels for each of a plurality of training images, 128 feature vectors are extracted per one feature point, 361 feature points can be extracted from a single patch with reference to 100x100 image. As a result, 361 128-dimensional SIFT vectors are extracted from one image, and 361x810 SIFT vectors are generated from 810 DBs in total, and K-means clustering processes can be performed by K words. Here, the word means the center of K-means clustering.

In the case of the edge analysis (S1102), a histogram is stored in which 35 binaries are binned by applying LBP_ROT to each image after leaving only the binarized edge by applying a cane edge detector to the patches obtained from the training images .

In the case of the color analysis (S1103), histograms storing 11 color values based on the HLS color space are stored for the patches for the training image.

Since the three analyzes (S1102, S1103, and S1104) are independent of each other, when the image search is performed, the result obtained by multiplying the similarity results of the individual histograms can be selected as the final result (S1131) .

Referring to FIG. 11B, when a test image to be tested is input similar to an image classification, a test patch is extracted through a face detection process (S1111) (S1112, S1113, and S1114) are performed in the three methods (S1102, S1103, and S1104) described in FIG. 11 (a). In the case of the Dense-SIFT analysis (S1114), the histogram of the word can be obtained using the K-D tree obtained in the image classification process (S1121).

After each histogram is obtained through each of the minutes (S1112, S1113, S1114), scores are obtained by the Histogram Intersection method (S1122), and the final score is obtained by multiplying the scores by the three scores, and a candidate for obtaining the highest value is output The search is completed (S1131).

12 and 13 show the results of comparing the image classification and search method of the present invention with the conventional technology.

Fig. 12 shows a comparison of the accuracy of extracting similar clothing with the conventional Dense-SIFT by performing a search for each piece of clothing image patch extracted from 36 clothes. FIG. 13 shows an example of ten similar clothing patches retrieved for each input garment patch in the conventional Dense-SIFT and the method proposed in this specification.

As shown in FIG. 12, in the case of extracting five similar clothing patches or extracting ten similar clothes patches, the method proposed in this specification is more similar to the case of using only the existing Dense-SIFT alone It can be seen that clothing patches can be accurately found from the database of the storage medium.

Also in FIG. 13, all but the 29th test patches were found to have garment patches obtained from more exact same clothes than Dense-SIFT. On the other hand, although not shown, Dense-SIFT achieved 23 successes (63.9%) in 36 experiments, while 34 suggestions (94.4%) were relatively successful As shown in Fig.

According to the method and apparatus disclosed in the present specification, a high search rate can be obtained even for a change in the rotation or size of the original image. Particularly, when the conventional method is used for the clothing search, there is a problem that a lot of distortion occurs in the training and searching process due to the irregular characteristics of the clothes, and the performance is very poor. However, according to the method and apparatus disclosed in the present specification, So that it is possible to exhibit high search performance even in non-standard clothes.

Especially, unlike the existing LBP_ROT, it focuses attention on the edge part where the signal characteristics are compressed, so that it can acquire excellent feature information and minimize the distortion due to the scale change of the original image. The HLS- The color space is quantized for color classification, which is of great importance, and is made into a color-based histogram, thereby exhibiting excellent performance in color search.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. It can be understood that the modification and application of the present invention are possible. For example, each component specifically shown in the embodiments can be modified and implemented. It is to be understood that the present invention may be embodied in many other specific forms without departing from the spirit or essential characteristics thereof.

101: Input unit
102: edge analysis unit
103: color analysis unit
104:
105: Dense-SIFT analysis unit

Claims (18)

An input unit for receiving at least one image;
An edge analyzer for extracting an edge region from a local Binary Pattern with ROTATION-invariant (LBP_ROT) for the image to obtain an edge histogram;
A color analyzer for converting each of the images into a HLS (Hue, Saturation, Lightness) color space and extracting a color histogram; And
And a storage unit for storing the edge histogram and the color histogram,
The LBP_ROT recognizes the patterns generated by rotating the image in 45-degree increments among the 256 patterns that can be distinguished by applying a Local Binary Pattern (LBP) technique to each pixel of the image, Wherein the number of patterns is reduced to 36. The method according to claim 1,
And a Dense-SIFT analyzing unit for extracting a Dense-SIFT (Scale Invariant Feature Transform) histogram for the image,
Wherein the storage unit stores the Dense-SIFT histogram. The method according to claim 1,
Wherein the edge region is extracted by a Canny Edge Detection algorithm. The method according to claim 1,
Wherein the color histogram is extracted by a method of dividing the converted HLS color space into a predetermined number of colors. 5. The method of claim 4,
Wherein the classification is performed within a predetermined brightness (L). The method according to claim 1,
Wherein the image is a garment image. Receiving an image;
Extracting only an edge region from a local Binary Pattern with ROTATION-invariant (LBP_ROT) for the image to obtain an edge histogram;
Converting the image into HLS (Hue, Saturation, Lightness) color space to extract a color histogram; And
Storing the edge histogram and the color histogram,
The LBP_ROT recognizes the patterns generated by rotating the image in 45-degree increments among the 256 patterns that can be distinguished by applying a Local Binary Pattern (LBP) technique to each pixel of the image, Wherein the number of patterns is reduced to 36. An input unit for receiving a test image;
An edge analyzer for extracting only an edge region from a local Binary Pattern with ROTation-invariant (LBP_ROT) to obtain an edge histogram;
A color analyzer for converting the test image into a HLS (Hue, Saturation, Lightness) color space and extracting a color histogram; And
And a determination unit for determining the similarity by comparing the edge histogram and the edge histogram and the color histogram of the color histogram and the pre-stored training image,
The LBP_ROT recognizes the patterns generated by rotating the image in 45-degree increments among the 256 patterns that can be distinguished by applying a Local Binary Pattern (LBP) technique to each pixel of the image, Wherein the number of patterns is reduced to 36. 9. The method of claim 8,
Wherein the determination unit outputs an image corresponding to the test image based on the edge histogram and the similarity of the color histogram. 9. The method of claim 8,
And a Dense-SIFT analyzing unit for extracting a Dense-SIFT (Scale Invariant Feature Transform) histogram for the image,
Wherein the determination unit determines the similarity by comparing the Dense-SIFT histogram and the training image Dense-SIFT histogram with respect to the pre-stored training image. 11. The method of claim 10,
Wherein the determination unit outputs an image corresponding to the test image based on the similarity of the edge histogram, the color histogram, and the Dense-SIFT histogram. 11. The method of claim 10,
Wherein the Dense-SIFT histogram is extracted using a KD tree generated upon extraction of the training image Dense-SIFT histogram. 9. The method of claim 8,
Wherein the edge region is extracted by a Canny Edge Detection algorithm. 9. The method of claim 8,
Wherein the color histogram is extracted by a method of dividing the converted HLS color space into a predetermined number of colors. 15. The method of claim 14,
Wherein the distinction is performed within a predetermined brightness (L). 9. The method of claim 8,
Wherein the test image is a garment image. 17. The method of claim 16,
And an apparel patch extracting unit for extracting the apparel image from the input image and outputting the apparel image to the input unit. Receiving a test image;
Obtaining an edge histogram by extracting only an edge region from a local Binary Pattern with ROTATION-invariant (LBP_ROT) for the test image;
Converting the test image into a HLS (Hue, Saturation, Lightness) color space to extract a color histogram; And
Comparing the edge histogram and the edge histogram and the color histogram of the color histogram with the pre-stored training image to determine the similarity,
The LBP_ROT recognizes the patterns generated by rotating the image in 45-degree increments among the 256 patterns that can be distinguished by applying a Local Binary Pattern (LBP) technique to each pixel of the image, Wherein the number of patterns is reduced to 36.

Images