JP2004086538A - Method and device for classifying feature quantity of image - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To precisely classify the feature quantities of image data for each image. <P>SOLUTION: A first classification parameter is generated by using a parameter indicating the distribution of feature vectors in a vector region, and the feature vectors are divided into a plurality of first vector regions by using the first classification parameter. Then, a second classification parameter is generated by calculating the representative point of each of the first vector areas, and the first vector region is divided into a plurality of second vector regions by using the second classification parameter. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an image feature quantity classification method and apparatus for classifying feature quantities extracted from image data for each similar feature quantity, and more specifically, for example, extracting object region information from image data such as a photographic image or a medical image. The present invention relates to an image feature amount classification method and apparatus for classifying feature amounts extracted from image data in order to divide image data into regions using feature amounts.
[0002]
[Prior art]
An image captured by a digital camera or the like includes, for example, a person, the sea, the sky, a building, and other backgrounds. If image data such as the person and background is extracted from the image as one object, different image processing can be performed for each object, and the meaning of each object can be determined and presented to the user. it can. For this purpose, it is necessary to divide the object area for each object included in the image.
[0003]
In the case of automatically extracting an object region from image data, it has been proposed to classify feature amounts and then divide image information using this classification result to automatically divide objects. That is, image data constituting one object has similar feature quantities such as color information, texture information, and wavelet coefficients. If this property is used, the image data can be divided into object areas.
[0004]
When classifying feature quantities, a vector quantization technique used in image compression or the like is used. Vector quantization is a region dividing method in a feature vector space with feature quantities as axes, and is an effective means for classifying pixel vectors having similar feature quantities. Specifically, a plurality of feature amounts for each pixel of the image are extracted, and a plurality of feature vectors having each feature amount as an element are generated. Thereafter, the plurality of feature vectors are classified into similar feature vectors (clustering). Based on the clustered result, the image data is classified and divided into a plurality of clustering regions.
[0005]
Here, each feature vector is classified into classification parameters having the closest distance among the plurality of classification parameters, whereby each feature vector is classified for each classification parameter. Therefore, when the plurality of feature vectors described above are classified, a plurality of classification parameters (representative vectors) that are the basis of classification are required.
[0006]
[Problems to be solved by the invention]
By the way, the distribution of the feature amount of the image greatly fluctuates due to the difference in the subject of the image. For example, when a person is reflected on the coast with the sea and sky as a background, and when a person is reflected on a small hill with the street as a background, the image has features such as color, brightness, and texture. Is different. For this reason, the distribution of the feature amount in the feature vector space differs for each image.
[0007]
On the other hand, in vector quantization used in image compression, an average classification parameter is created using various images as learning data to reduce information redundancy, and vector quantization is performed using the classification parameter. Therefore, when the vector quantization used in image compression is used as it is for the vector quantization for classifying the feature amount, there is a problem that the accuracy of the feature amount classification is low.
[0008]
In other words, the classification parameter used in image compression or the like is an average classification parameter using various images as learning data, and therefore the optimum classification parameter is not necessarily used for each image. For this reason, there is a problem in that the accuracy in dividing the region for each pixel based on the feature amount classification is lowered, and when the image data region is divided on the basis of the feature amount classification, the region cannot be correctly divided.
[0009]
SUMMARY OF THE INVENTION An object of the present invention is to provide an image feature quantity classification method and apparatus capable of accurately classifying feature quantities of image data for each image.
[0010]
[Means for Solving the Problems]
The image feature quantity classification method of the present invention includes a step of extracting a plurality of feature quantities from a plurality of image data, a step of generating a feature vector for each of the image data using the extracted feature quantities, Generating a classification parameter for classifying the plurality of generated feature vectors for each similar feature vector, and classifying the plurality of feature vectors for each similar feature vector using the generated classification parameter An image feature quantity classification method comprising: generating a plurality of vector regions, wherein a first classification parameter is generated using a parameter indicating a distribution of the feature vectors in the vector region; Dividing the feature vector into a plurality of first vector regions using a first classification parameter; Calculating a representative point for each of the first vector regions and generating a second classification parameter; and dividing the plurality of feature vectors into a plurality of second vector regions using the generated second classification parameter It is characterized by having.
[0011]
Here, the “vector region generation” is performed by classifying the feature vector for each classification parameter by determining which classification parameter is most similar among the plurality of classification parameters.
[0012]
Also, the step of generating the first classification parameter, the step of dividing into the first vector region using the first classification parameter, the step of generating the second classification parameter, and dividing into the second vector region using the second classification parameter The classification process having steps may be performed only once or may be performed repeatedly a plurality of times.
[0013]
An image feature amount classification apparatus according to the present invention includes a feature amount extraction unit that extracts a plurality of feature amounts from image data, and a feature vector that generates a feature vector for each image data using the extracted feature amounts. Generating means, classification parameter generating means for generating classification parameters for classifying the generated feature vectors for each similar feature vector, and the plurality of feature vectors using the generated classification parameters. A feature quantity classification device for an image having a feature vector classification unit that generates a plurality of vector regions classified for each similar feature vector, wherein the classification parameter generation unit includes a parameter indicating a distribution of the feature vector in the vector region A first classification parameter generating means for generating a first classification parameter using the first class parameter, and the generated first class parameter And having a second classification parameter generating means for generating a second classification parameters calculated respectively representative points of the first vector regions divided by the feature vector classifying means using a kind parameters.
[0014]
Here, the “first classification parameter generation means” may generate the first classification parameter for all of the plurality of vector regions, or the vector having the largest parameter indicating the distribution of the feature vector among the plurality of vector regions. The first classification parameter may be generated only for the region.
[0015]
Further, the “first classification parameter generation unit” may generate the first classification parameter by using it as a parameter indicating the distribution of the vector region, and for example, the variance of the feature vector can be used. Furthermore, the “second classification parameter generating unit” may generate the second classification parameter using the calculated representative point, and for example, the center of gravity of the vector region can be used.
[0016]
【The invention's effect】
According to the image feature quantity classification method and apparatus of the present invention, the first classification parameter is generated using the parameter indicating the distribution of the feature vector, is divided into the first vector region using the first classification parameter, The second classification parameter is generated using the representative points of the vector area, and is divided into the second vector area using the second classification parameter, thereby dividing the first vector area in accordance with the feature amount distribution for each image. As a result, it is possible to generate an optimum classification parameter suitable for the feature quantity distribution of the image for each image, so that the feature quantity can be classified with high accuracy.
[0017]
Further, not only the division of the first vector area by the first classification parameter, but also the generation of the representative point of the first vector area as the second classification parameter, not only the generation of the rough classification parameter by the variance, but also after that By generating accurate classification parameters, it is possible to classify feature quantities with high accuracy.
[0018]
It should be noted that the step of generating the first classification parameter, the step of dividing into the first vector region using the first classification parameter, the step of generating the second codebook, and dividing into the second vector region using the second classification parameter By repeating the division process including steps a plurality of times, the number of classification parameters can be increased with high accuracy, so that the feature quantities can be classified more precisely.
[0019]
In addition, the second classification parameter can be efficiently generated by generating the first classification parameter of the vector area having the largest parameter indicating the distribution of the feature vector among the plurality of vector areas.
[0020]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 1 is a block diagram showing a preferred embodiment of an image feature quantity classification apparatus according to the present invention. The image feature quantity classification apparatus will be described with reference to FIG. The image feature quantity classification device 1 classifies the feature quantity CQ of the image data PD for each feature vector CB in order to divide the clustering area of the image data PD for each object, and the like. Means 10, feature vector generation means 20, classification parameter generation means 30, feature vector classification means 40, and the like.
[0021]
The feature quantity extraction unit 10 has a function of extracting the feature quantity CQ from the image data PD and outputting it to the feature vector generation unit 20. The image data PD is M × N pixel data x constituting one image. _ij (I, j: 0 ≦ i ≦ N, 0 ≦ j ≦ M), and each pixel data includes various image information such as color information and luminance information. The feature amount extraction unit 10 has a function of extracting a set feature amount CQ from the image information of the image data PD. For example, the feature quantity extraction unit 10 uses the n feature quantities CQ = (color information such as RGB, YCC, Lab, and the like, texture information such as wavelet coefficients, DCT coefficients, etc. as the feature quantity CQ from the image data PD. ⁰ x _ij , ¹ x _ij , ² x _ij , ³ x _ij ... ^n-1 x _ij , ⁿ x _ij ) To extract. Note that the number and type of feature amounts CQ extracted by the feature amount extraction means 10 can be set by the user.
[0022]
The feature vector generation unit 20 has a function of generating a feature vector CB for each image data PD using the plurality of feature amounts CQ extracted by the feature amount extraction unit 10. Then, the feature vector generation means 20 stores the generated feature vector CB in the vector database 25. For example, when the feature quantity extraction unit 10 extracts n feature quantities CQ from each image data PD, the feature vector generation unit 20 uses the feature vector CB = ( ⁰ x _ij , ¹ x _ij , ² x _ij , ³ x _ij ... ^n-1 x _ij , ⁿ x _ij ). That is, the feature vector generation means 20 maps from the image space to the feature vector space.
[0023]
The classification parameter generation means 30 has a function of generating a classification parameter for classifying a plurality of feature vectors CB in the vector database 25 for each similar feature vector CB. The classification parameter generation unit 30 stores the generated classification parameters in the code book 35.
[0024]
The feature vector classification means 40 has a function of dividing a plurality of feature vectors CB into a plurality of vector regions BR (clusters) using the classification parameters of the code book 35 (clustering). Specifically, the feature vector classification unit 40 determines which classification parameter is most similar to each feature vector CB among the plurality of classification parameters. As a result, the feature vector classifying unit 40 classifies the plurality of feature vectors CB for each classification parameter, and generates a vector region BR including a plurality of similar feature vectors CB. The most similar classification parameter is selected, for example, by selecting a classification parameter RB that is closest to the feature vector CB (such as Euclidean distance and Mahalanobis distance).
[0025]
For example, when there are L classification parameters RB, the feature vector classification means 40 divides the feature vector space into L vector regions BR. Then, the feature vector classification unit 40 associates the feature vector CB with the classification parameter to which the feature vector CB belongs and stores it in the vector database 25. Therefore, each feature vector in the vector database 25 is stored for each feature vector region BR.
[0026]
Here, the classification parameter generation unit 30 includes a first classification parameter generation unit 31 that generates the first classification parameter HRB and a second classification parameter generation unit 32 that generates the second classification parameter RB. The first classification parameter generation unit 31 generates a first classification parameter HRB using a parameter indicating the distribution of the feature vector CB in each feature vector region BR. Specifically, the first classification parameter generating unit 31 uses the variance σ = (() of the feature vector CB as a parameter indicating the distribution of the feature vector CB. ⁰ σ _ij , ¹ σ _ij , ² σ _ij , ³ σ _ij ... ^n-1 σ _ij , ⁿ σ _ij ) Is generated. And the 1st classification parameter production | generation means 31 produces | generates the 1st classification parameter HRB using dispersion | distribution (sigma). Then, the first classification parameter generation means 31 stores the generated first classification parameter HRB in the code book 35.
[0027]
On the other hand, the second classification parameter generation unit 32 calculates the representative points of the vector regions BR divided by the feature vector classification unit 40 using the generated first classification parameter HRB, and generates the second classification parameter RB. . Here, the representative point is the center of gravity (average value) of each feature vector CB in the Euclidean metric space, and the second classification parameter RB = ( ⁰ RB _ij , ¹ RB _ij , ² RB _ij , ³ RB _ij ... ^n-1 RB _ij , ⁿ RB _ij ). When there are K feature vectors in the vector area, each component of the average value is ⁿ RB _ij = Σ ⁿ X _ij It is generated by calculating / K. Then, the second classification parameter generation unit 32 sets the generated representative point as the second classification parameter RB. The generated second classification parameter RB is used as a classification parameter for classifying the final feature quantity CQ.
[0028]
The second classification parameter generation unit 32 calculates the center of gravity (average value) at the Euclidean distance as a representative point, but it is difficult to calculate the center of gravity in a general metric space that does not use the Euclidean distance. A representative point other than the center of gravity may be calculated.
[0029]
The control unit 50 has a function of controlling operations of the classification parameter generation unit 30 and the feature vector classification unit 40. Specifically, the first classification parameter generation means 31, the second classification parameter generation means 32, and the feature vector classification means 40 are controlled so as to realize the feature quantity classification method as described below.
[0030]
FIG. 2 is a flowchart showing a preferred embodiment of the image feature quantity classification method of the present invention. The image feature quantity classification method will be described with reference to FIG. First, the feature quantity CQ of the image data PD is extracted by the feature quantity extraction means 10 (step ST1). A feature vector CB is generated by the feature vector generation means 20 using the extracted feature quantity CQ and stored in the vector database 25 (step ST2). Then, the center of the feature vector region BR composed of all the feature vectors CB is calculated by the second classification parameter generation means 32 (step ST3). Thereafter, the calculated center of gravity is stored in the code book 35 as an initial classification parameter RBref (step ST4).
[0031]
Next, the first classification parameter generation means 31 calculates the variance σ of the initial vector region BRref composed of all the feature vectors CB (step ST5). Then, two first classification parameters HRB1 = (RB + σ) and HRB2 = (RB−σ) obtained by adding / subtracting the variance σ to the calculated initial classification parameter RBref are generated and stored in the code book 35 (step ST6). Note that the first classification parameter generation means 31 adds the variance σ multiplied by an appropriate weight α to generate the first classification parameters HRB1 = (RBref + ασ), HRB2 = (RBref−ασ). Good.
[0032]
At this time, the first classification parameters HRB1 and RGB2 are stored in the code book 35 so as to overwrite the classification parameter that is the basis for generating the first classification parameters HRB1 and HRB2, that is, the initial classification parameter RBref. Accordingly, only the first classification parameters HRB1 and RGB2 exist in the code book 35. Thereafter, the feature vector region BR is divided into two first vector regions using the two first classification parameters HRB1 and HRB2 (step ST7).
[0033]
Further, the center of gravity of each divided first vector region HBR is calculated (step ST8), and the second classification parameters RB1 and RB2 are generated and stored in the codebook 35 (step ST9). At this time, the second classification parameters RB1 and RB2 are stored in the code book 35 so as to overwrite the classification parameters that are the basis for generating the second classification parameters RB1 and RB2, that is, the first classification parameters HRB1 and HRB2. Accordingly, only the first classification parameters RB1 and RB2 exist in the code book 35. Thereafter, the feature vector classifying unit 40 generates two second vector regions BR using the second classification parameters RB1 and RB2 (step ST10).
[0034]
This division process (step ST5 to step ST10) is repeated until the counter S reaches the set number of times (S = P) (step ST11). Therefore, P + 1 second classification parameters RB are stored in the code book 35. In other words, the feature vector space is divided into P + 1 feature vector regions BR, and the feature quantity CQ is classified in the form of the feature vector CB.
[0035]
FIGS. 3 to 6 are diagrams illustrating how regions are divided in the feature vector space. A specific example of the image feature amount classification method will be described with reference to FIGS. 1 to 3. In FIGS. 3 to 6, three feature amounts CQ = ( ⁰ x _ij , ¹ x _ij , ² x _ij ) Is extracted from the image data PD and the feature vector space is divided into five vector regions BR by repeating the division process (step ST5 to step ST10) four times (P = 4).
[0036]
First, using the extracted three feature quantities, a feature vector CB = ( ⁰ x _ij , ¹ x _ij , ² x _ij ) Is generated for each image data PD. Next, the center of gravity (x ₀ , Y ₀ , Z ₀ ) Is calculated, and this center of gravity is stored in the code book 35 as the initial classification parameter RBref. Thereafter, the variance σ = (σ for the initial classification parameter RBref _x , Σ _y , Σ _z ) Is calculated. Then, the variance σ is added to or subtracted from the calculated first classification parameter HRB, so that the first classification parameter HRB1 = (RB + σ) = (x ₀ + Σ _x , Y ₀ + Σ _y , Z ₀ + Σ _z ), HRB2 = (RB + σ) = (x ₀ −σ _x , Y ₀ −σ _y , Z ₀ −σ _z ) Is generated.
[0037]
Thereafter, the feature vector classifying unit 40 determines which of the two second classification parameters HRB1 and HRB2 for each feature vector is closest to the first classification parameter HRB1 and HRB2. Thus, as shown in FIG. 3B, the feature vector space is divided into two first vector regions HBR1 and HBR2.
[0038]
Next, the second classification parameter generation means 32 calculates the center of gravity for each of the two first vector regions HBR1 and HBR2. The calculated centroids are stored in the code book 35 as second classification parameters RB1 and RB2, respectively. Thereafter, the region is divided by the feature vector classification means 40 using the second classification parameters RB1 and RB2. Then, as shown in FIG. 3C, the feature vector space is divided into two second vector regions BR1 and BR2.
[0039]
Next, P = 2 second region division processing (step ST5 to step ST10) is performed. First, the variance σ of each feature vector with respect to the second classification parameters RB1, RB2 _BR1 , Σ _BR2 Is calculated for each of the second vector regions BR1 and BR2. Then, in the second classification parameter generation means 32, the calculated variance σ _BR1 , Σ _BR2 It is determined which variance σ is larger. In FIG. 3C, the variance σ of the feature vector region BR1 _BR2 Is bigger. Then, in the second classification parameter generation means 32, the second vector region BR2 is set as a vector region divided into two.
[0040]
Then, as shown in FIG. _BR2 The two first classification parameters HRB11 = (RB2 + σ _BR2 ), HRB12 = (RB2-σ _BR2 ) Is calculated. Thereafter, the second classification parameter RB2 in the code book 35 is replaced with the calculated first classification parameters HRB1 and HRB2. Therefore, the code book 35 is in a state where the first classification parameters HRB1 and HRB2 and the second classification parameter RB1 are stored.
[0041]
Then, the feature vector classification means 40 performs region division using the code book 35, thereby generating three first vector regions. Then, the second classification parameter generation unit 32 calculates the centroid for each of the three feature vectors to generate the three second classification parameters RB11, RB12, and RB13, and then stores them in the codebook 35. Further, the feature vector classification means 40 divides the feature vector space into three second vector regions BR11, BR12, BR13 as shown in FIG.
[0042]
Next, S = third division process (step ST5 to step ST10) is performed. Variance σ of each feature vector CB for each second vector region BR11, BR12, BR13 _BR11 , Σ _BR12 , Σ _BR13 Is calculated and the largest variance σ _BR12 Is selected. Then, as shown in FIG. 5A, in the second classification parameter generation means 32, the second vector area BR12 is set as a vector area to be classified into two.
[0043]
Then, as shown in FIG. 5 (a), the selected variance σ _BR13 The two first classification parameters HRB21 = (RB12 + σ _BR12 ), HRB22 = (RB12−σ _BR12 ) Is calculated. Then, the calculated first classification parameters HRB21 and HRB22 are stored in the code book 35. Then, the first classification parameters HRB21 and HRB22 and the second classification parameters RB11 and RB13 are stored in the code book 35. Thereafter, the feature vector classifying unit 40 divides the region using the code book 35, and divides the feature vector space into four first vector regions.
[0044]
Further, the center of gravity is calculated for each first vector region, and four second classification parameters RB11, RB13, RB21, and RB22 are generated and stored in the codebook 35. Then, as shown in FIG. 5B, the feature vector classifying unit 40 divides the feature vector space into four second vector regions BR21, BR22, BR23, and BR24.
[0045]
Similarly, S = P = the fourth division process (step ST5 to step ST10) is performed. Variance σ of feature vectors for each feature vector region BR21, BR22, BR23, BR24 _BR21 , Σ _BR22 , Σ _BR23 , Σ _BR24 Is calculated and the largest variance σ _BR21 Is selected. Then, as shown in FIG. 6A, in the second classification parameter generation means 32, the second vector region BR21 is set as a vector region divided into two.
[0046]
And the selected variance σ _BR21 The two first classification parameters HRB31 = (RB11 + σ _BR21 ), HRB32 = (RB11−σ _BR21 ) Is calculated. The calculated first classification parameters HRB31 and HRB32 are stored in the codebook 35, and the first classification parameters HRB31 and HRB32 and the second classification parameters RB21, RB22 and RB13 are stored in the codebook 35.
[0047]
Then, the feature vector classifying unit 40 divides the region using the code book 35, and divides the feature vector space into five first vector regions. Further, the center of gravity is calculated for each first vector region, and second classification parameters RB13, RB21, RB22, RB31, and RB32 as shown in FIG. 6B are generated and stored in the codebook 35. . Then, the feature vector classifying unit 40 divides the region using the code book 35, and divides the feature vector space into five second vector regions BR13, BR21, BR22, BR31, BR32.
[0048]
According to the above-described embodiment, the first classification parameter HRB is generated using the variance σ, and after the region is temporarily divided by the second codebook 352, the center of gravity in each feature vector region BR is calculated to calculate the second classification. By generating the parameter RB, the feature quantity can be classified efficiently and accurately. That is, when the code book 35 is generated only from the center of gravity, if the distribution of the feature vector is bipolar (see the feature vector region BR11 in FIG. 4C), the center of gravity indicates the middle of the bipolarized distribution. In other words, a long-distance classification parameter is generated from any distribution.
[0049]
On the other hand, when the code book 35 is generated using only the variance σ, the first classification parameter matching the distribution of the feature vector CB can be generated. However, since the region division using the variance σ is only coarsely classified, the accuracy of image feature amount classification is degraded. Therefore, as described above, after the region is roughly divided using the first classification parameter HRB using the variance σ, the region is precisely divided using the second classification parameter RB including the center of gravity, so that an efficient and accurate feature can be obtained. The quantity CQ can be classified. As described above, if the feature quantity can be classified accurately, the area of the image data PD can be accurately divided based on the classification of the feature quantity CQ, so that the accuracy of object extraction can be improved.
[0050]
The embodiment of the present invention is not limited to the above embodiment. For example, in FIG. 3 to FIG. 6, the first classification parameter generation means 31 generates the first classification parameter HRB for the feature vector region having the largest variance σ among the plurality of feature vector regions. The first classification parameter HRB may be generated for all regions. In this case, the first classification parameter HRB and the second classification parameter RB are P ² Only one will be generated.
[0051]
Further, the feature amount extraction unit 10 is exemplified for the case where the user sets the feature amount CQ to be extracted from the image data PD. However, the feature amount extraction unit 10 has categories such as “landscape image” and “medical image”. It has an image information database in which information and the type of feature quantity CQ to be extracted are linked and stored, and when the user selects category information, the feature quantity CQ extracted from the image data PD may be changed. . As a result, the feature quantity CQ to be extracted can be changed according to the type of the image, so that the feature quantity CQ that facilitates the division of the object area in the image data PD can be classified.
[0052]
Specifically, for example, it is assumed that there is image data PD having a region irradiated with direct X-rays such as a medical image and a region irradiated with X-rays through a subject. This image data PD is characterized in that the outside of the irradiation field stop has a low density and little texture, the direct X-ray part in the irradiation field stop has a high density and little texture, and the subject part has an intermediate density and a lot of texture. . Therefore, density information and texture information are extracted as feature amounts from each image data PD.
[0053]
Then, when a plurality of feature vector regions are generated by the image feature amount classification apparatus 1 of FIG. 1 and mapped to the image data PD, it is the outermost vector region BR or the average density in the vector region BR is predetermined. The irradiation field region can be determined based on at least one index that is equal to or less than the value or whether the average texture amount in the vector region BR is equal to or less than the predetermined value. Based on the result of this region division, it is possible to prevent deterioration in diagnostic ability due to illusion by adding image processing for increasing the density of the area that has become low density due to the irradiation field stop of the X-ray image.
[0054]
Or, from the obtained feature vector area, the relative position relationship between the clustering areas such as the subject at the center of the image and the X-ray part directly at the corner of the image, the average density in the feature vector area The background portion (direct X-ray portion) and the subject portion are discriminated based on at least one index of the average texture amount in the feature vector region, and the subject portion and the background portion (direct X-ray portion) are respectively determined. Appropriate image processing parameters may be calculated and image processing may be performed for each region. As a result, the dynamic range of the subject and the dynamic range of the entire image can be controlled separately, so that an image with an optimal dynamic range can be obtained. For example, when dynamic range compression is performed over the entire image data PD, the dynamic range of the subject is appropriate, but the outside of the subject (particularly the direct X-ray part) may be low in density and may have an image that is not tightened. On the other hand, the overall image quality of the image data PD can be improved by changing the compression condition of the dynamic range for each vector area as described above.
[0055]
In this way, the feature amount classification that enables the region division of the image data to be accurately performed by extracting the feature amount suitable for the region division in the image data PD and classifying the feature amount. It can be performed.
[0056]
1 may be realized using hardware resources, or may be realized by cooperation of software and a computer. The image feature quantity classification program includes feature quantity extraction means for extracting a plurality of feature quantities from image data, and feature vector generation means for generating a feature vector for each of the image data using the extracted feature quantities. Classification parameter generation means for generating a classification parameter for classifying the generated plurality of feature vectors for each similar feature vector, and a plurality of the feature vectors using the generated classification parameter In the image feature quantity classification program for causing a computer to function so as to have an area dividing means for dividing into vector areas, the classification parameter generating means uses the parameter indicating the distribution of the feature vectors in the vector area to perform a first classification. First classification parameter generation means for generating parameters, and the generated first classification A computer is caused to function as having a second classification parameter generating unit that calculates a representative point of the vector region divided by the region dividing unit using a parameter and generates a second classification parameter. It can be said that it is an image feature amount classification program.
[0057]
By using the result of region division (class classification) in the feature vector space, the feature vector region is mapped to the image space, and then the object region OR is extracted from the image data. At this time, if region division is performed according to the vector region BR classified in the feature vector space, even if the feature vector space includes 10 vector regions, the real space region includes, for example, a large number of regions of 100 or more. This is because even image data PD in the same vector area BR may exist in different areas in the real space. Specifically, when the region division of the image data PD is performed based on the classification result of the feature vector CB, a plurality of clustering regions DP are formed as shown in FIG.
[0058]
Therefore, it is necessary to extract the object region OR as shown in FIG. 7B by integrating the clustering regions DP divided into regions based on the vector region BR. However, if the accuracy of region integration is low, there is a problem that it cannot be correctly classified for each object in the image. Therefore, as will be described below, an object extraction apparatus 100 capable of performing correct area integration is provided.
[0059]
Specifically, FIG. 8 is a block diagram illustrating an example of an object extraction apparatus, and the object extraction apparatus 100 will be described with reference to FIG. The object extraction device 100 in FIG. 8 includes an image feature amount classification device 1 that classifies a plurality of feature amounts CQ extracted from the image data PD described above for each similar feature amount CQ, and features by the image feature amount classification device 1. A region dividing unit 101 that classifies the image data PD based on the classification result of the quantity CQ to generate a plurality of clustering regions DP, and an object extraction unit 110 that extracts the object region OR by integrating the generated clustering regions DP. Have
[0060]
Here, when the region dividing unit 101 generates a clustering region DP, a label is attached to each clustering region DP (labeling process). Further, as shown in FIG. 9A, the region dividing unit 101, together with the label, the starting point coordinates (sx, sy), the region width wx, and the region height wy (see element 9 (b)) together with the label. Each clustering region DP is stored in the database 111.
[0061]
When the vector region BR in the feature vector space is mapped onto the image frame PF, there may be an isolated pixel that is not a region cluster (a group of pixels). Therefore, the area dividing unit 101 may perform the smoothing process so as to belong to a vector area (cluster) BR that is closest in terms of space or feature quantity of the image frame PF. Thereby, the isolated pixel when the area is divided can be removed as noise.
[0062]
The object extraction unit 110 determines a minimum region selection unit 112 that selects a minimum clustering region DPmin having the smallest number of pixels from a plurality of clustering regions DP, and an adjacent clustering region DPa that is integrated with the extracted minimum clustering region DPmin. The integrated determination unit 113 and the region integration unit 114 that integrates the determined adjacent clustering region DPa and the minimum clustering region DPmin are provided.
[0063]
Here, when the minimum clustering area DPmin is equal to or smaller than the minute pixel threshold, the integrated determination unit 113 minimizes the adjacent clustering area DPa having the largest number of pixels among the adjacent clustering areas DPa adjacent to the minimum clustering area DPmin. The clustering area is determined to be integrated with the clustering area DPmin.
[0064]
On the other hand, when the minimum clustering area DPmin is equal to or smaller than the small pixel threshold, the integrated determination unit 113 determines that the adjacent clustering area DPa having the closest feature among the adjacent clustering areas DPa is a clustering area to be integrated with the minimum clustering area DPmin. It is.
[0065]
The above-described minute pixel threshold value is set to, for example, 1/100 pixel number of the total number of pixels of the image data PD, and the small pixel threshold value is, for example, 1/10 pixel number of the total number of pixels of the image data PD. Is set to
[0066]
Further, when the integrated determination unit 113 selects an adjacent clustering region DPa having a close feature, for example, by referring to the distance in the feature vector space between the minimum clustering region DPmin and the adjacent clustering region DPa, the adjacent clustering region DPa having the shortest distance is used. Is supposed to be selected.
[0067]
Alternatively, the change amount of the minimum clustering region DPmin, the adjacent clustering region DPa, and the feature amount CQ (for example, if YCC is the feature amount CQ, the average value of the Y edge signal, the Cr edge signal, and the Cb edge signal, etc.) is small. The adjacent clustering area DPa may be selected.
[0068]
Further, when the clustering region DP after the integration of the plurality of clustering regions DP is the adjacent clustering region DPa, the average value of the feature amount CQ calculated from the adjacent clustering region DPa after the integration is used.
[0069]
Next, the region integration method will be described with reference to FIG. First, as shown in FIGS. 1 to 7, the vector region BR is divided in the feature vector space using the feature quantity CQ of the image data PD (step ST21). Thereafter, the area dividing unit 101 performs area division on the image data PD in accordance with the divided vector area BR. Then, the image data PD is divided into a plurality of clustering areas DP (step ST22). At this time, each clustering region DP is stored in the database 111 in a labeled state.
[0070]
Note that when the classification result of the feature vector space is mapped onto the image data PD, there may be an isolated pixel that is not an area cluster (a group of pixels). At this time, the region dividing unit 90 has a function of performing a smoothing process so as to belong to the closest cluster spatially and / or feature amount of the image data PD in order to remove isolated pixels as noise. It may be.
[0071]
Thereafter, the minimum clustering region DPmin having the smallest number of pixels is extracted from the plurality of clustering regions DP by the minimum region selection means 112 (step ST23). Also, the adjacent clustering region DPa adjacent to the extracted minimum clustering region DPmin is selected by the integrated judgment unit 113 (step ST24).
[0072]
Then, the integrated determination unit 113 determines whether the number of pixels in the minimum clustering region DP is equal to or smaller than the minute pixel threshold (step ST25). When the minimum clustering area DPmin is equal to or smaller than the minute pixel threshold, the number of boundary pixels between the minimum clustering area DPmin and the adjacent clustering area DPa is measured for each adjacent clustering area DPa. Then, the minimum clustering area DPmin is integrated with the adjacent clustering area DPa having the largest number of boundary pixels (step ST26).
[0073]
Specifically, in FIG. 11A, it is assumed that the clustering region A is the minimum clustering region DPmin that is equal to or smaller than the minute pixel threshold. Since the clustering region A is adjacent to the clustering regions C and D, the clustering regions C and D become the adjacent clustering region DPa. The number of adjacent pixels with which the minimum clustering area A and the clustering areas C and D are in contact is measured. Then, in FIG. 11A, the number of boundary pixels with the adjacent clustering region D is larger than the number of boundary pixels with the adjacent clustering region C. For this reason, the clustering area A is integrated with the clustering area D as shown in FIG.
[0074]
Thus, when the minimum clustering region DPmin is a minute clustering region, an optimal object region OR can be generated by combining with the adjacent clustering region DPa having the largest number of adjacent pixels.
[0075]
On the other hand, when the number of pixels in the minimum clustering region DPmin is equal to or smaller than the small pixel threshold value (step ST27), the distance in the feature vector space from the adjacent clustering region DPa is calculated. Then, the minimum clustering area DPmin is integrated with the adjacent clustering area DPa having the shortest distance (step 28).
[0076]
Specifically, as shown in FIG. 11B, it is assumed that the clustering region B is a minimum clustering region DPmin having a small pixel threshold value or less. Then, the adjacent clustering regions DPa of the clustering region B are the clustering regions C and D. Then, when the texture information is a feature quantity for measuring the distance, it is determined which of the clustering regions C and D is close to the texture of the clustering region B. Then, the clustering region B is integrated with the clustering region C as shown in FIG.
[0077]
Thus, when the minimum clustering region DPmin is a small clustering region, an optimal object region can be generated by integrating with the adjacent clustering region DPa having the most similar features.
[0078]
Further, the integration process (steps ST23 to ST28) described above is repeated until there is no small clustering area, that is, until the minimum clustering area DPmin becomes larger than the small pixel number threshold. Then, a plurality of clustering areas DP shown in FIG. 7A are integrated, and an object area OR (at least a macro area equal to or larger than the small pixel threshold) is generated as shown in FIG. 7B.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a preferred embodiment of an image feature amount classification apparatus according to the present invention.
FIG. 2 is a flowchart showing a preferred embodiment of the image feature amount classification method of the present invention.
FIG. 3 is a diagram showing a feature vector space in the image feature quantity classification method of the present invention;
FIG. 4 is a diagram showing a feature vector space in the image feature quantity classification method of the present invention;
FIG. 5 is a diagram showing a feature vector space in the image feature quantity classification method of the present invention;
FIG. 6 is a diagram showing a feature vector space in the image feature amount classification method of the present invention;
FIG. 7 is a diagram illustrating a state in which a clustering region is generated using a vector region divided using an image feature amount classification device;
FIG. 8 is a block diagram illustrating an example of an object extraction device using an image feature amount classification device;
9 is a diagram showing an example of clustering area information stored in a database in the object extraction device of FIG. 8;
FIG. 10 is a flowchart illustrating an example of an object extraction method.
FIG. 11 is a diagram showing a state of region integration in the object extraction method.
[Explanation of symbols]
1 Image feature classifier
10 Feature extraction means
20 Feature vector generation means
30 Classification parameter generation means
31 First classification parameter generation means
32 Second classification parameter generation means
35 Codebook
40 Feature vector classification means
BR vector region
CB feature vector
CQ feature
HRB classification parameters
DP clustering domain
PD image data
HRB first classification parameter
RB Second classification parameter

Claims (5)

Extracting a plurality of feature amounts from the image data; generating a feature vector for each of the image data using the extracted plurality of feature amounts; and for each feature vector similar to the generated feature vectors Generating a classification parameter for classifying the plurality of feature vectors, and generating a plurality of vector regions obtained by classifying the plurality of feature vectors for each similar feature vector using the generated classification parameter A quantity classification method,
Generating a first classification parameter using a parameter indicating a distribution of the feature vector in the vector region;
Dividing the feature vector into a plurality of first vector regions using the generated first classification parameter;
Calculating a representative point for each of the divided first vector regions to generate a second classification parameter;
And a step of dividing the plurality of feature vectors into a plurality of second vector regions using the generated second classification parameter. Generating the first classification parameter; dividing the first vector area using the first classification parameter; generating the second classification parameter; and second vector area using the second classification parameter. The image feature quantity classification method according to claim 1, wherein the division process including the step of dividing the image into a plurality of times is repeated a plurality of times. Feature amount extraction means for extracting a plurality of feature amounts from image data, feature vector generation means for generating a feature vector for each image data using the extracted feature amounts, and the generated features Classification parameter generation means for generating a classification parameter for classifying a vector for each similar feature vector, and a plurality of the plurality of feature vectors classified for each of the similar feature vectors using the generated classification parameter In an image feature quantity classification device having a feature vector classification means for generating a vector region,
The classification parameter generating means is
First classification parameter generation means for generating a first classification parameter using a parameter indicating a distribution of the feature vector in the vector region;
Second classification parameter generation means for generating a second classification parameter by calculating representative points of the first vector regions divided by the feature vector classification means using the generated first classification parameter. An apparatus for classifying feature quantities of images as features. The said 1st classification parameter production | generation means produces | generates the 1st classification parameter of the said vector area | region with the largest parameter which shows the distribution of the said feature vector among several said vector area | regions. An image feature classifying device. The first classification parameter generation means uses the variance of the feature vector area as a parameter indicating the distribution of the vector area, and the second classification parameter generation means uses the center of gravity of the first vector area as the representative point. 5. The image feature amount classification apparatus according to claim 3, wherein the image feature amount classification device is calculated as follows.

Images