JPH09330407A - Picture pattern recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To speedily and precisely rocognize an object within a picture of a normal form and a picture compressed based on an orthogonal extending method. SOLUTION: A small area extracting circuit 6 extracts plural pertinent small areas from the picture of the normal form and the picture of a compressed form. A mask processing circuit 7 prepares a mask processing small area. An orthogonal extending circuit 8 extracts a feature vector from the mask processing small area by using the orthogonal extending method. A sorting feature vector preparing circuit 11 decides a component to use for sorting judgement. A sorting feature extracting circuit 9 prepares a sorting feature vector. A sorting judging circuit 10 calculates the correlation value of the sorting feature vector and a standard pattern, decides whether the feature vector is similar to a recognition object class based on the correlation value and outputs only the sorting feature vector sorted to be similar to the recognition object class to a picture recognition circuit 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition apparatus for recognizing a binary or multi-valued image pattern using an orthogonal expansion method.

[0002]

2. Description of the Related Art Conventionally, in an image pattern recognition apparatus having means for performing recognition processing accompanied by signal / noise separation processing, signal / noise separation processing and detailed recognition are performed from a large number of sample images collected in advance according to the recognition target. A procedure for creating a plurality of templates used in the processing by using the orthogonal expansion method and a procedure for configuring a classification circuit are required. A method using a principal component analysis method for the orthogonal expansion method used here is known (for example, Arimura, Hagita: construction method of image selection space in statistical image recognition, advanced technology, Institute of Electronics, Information and Communication Engineers, PRU95- 147, 1995).

On the other hand, with the recent development of computer network technology and multimedia technology, images compressed by the image compression technology using the orthogonal expansion method have been widely distributed and spread. Along with this, a method and apparatus have been developed for performing orthogonal expansion at high speed and improving the speed of image compression and decompression processing. For example, JPEG
A method is known in which discrete cosine transform (DCT) for 8 × 8 pixels used in the method is realized at high speed by matrix calculation of a transform matrix of 8 × 8 matrix (for example, edited by Yasuda, International standard for multimedia coding). , Maruzen, 1993).
In addition, there is an increasing demand for image recognition for these compressed images. For example, in an image database search, a DCT is used to target an image compressed using DCT.
A method of content search in which an object to be searched is searched by using a part of the coefficient is known (for example, SW Smo).
liar, H .; Ahang, Content-Base
dVideo Indexing and Retri
eval, IEEE Multimedia, pp. 6
2-72, Summer, 1994).

[0004]

In the conventional image pattern recognition apparatus which creates a template by using the principal component analysis in the orthogonal expansion method, when the size and the number of sample images increase, it is necessary to calculate the eigenvectors of a large matrix. However, there is a drawback that it takes a lot of calculation time. For example, the size of the sample image is 30 × 30 pixels, the number is 100,
When set to 000, a 900 × 900 covariance matrix is created from the addition of 100,000 samples, and the eigenvector is calculated from the covariance matrix. Also, with this method, when the recognition target handled by each user is different, it takes a lot of calculation time each time, and it is necessary to newly calculate the eigenvector and recreate the template, which reduces the processing time. There was a problem with the conversion. Further, there is a problem in downsizing the image pattern recognition device because a means for storing the template is required.

When a compressed image that has been orthogonally expanded by DCT is processed using a conventional image pattern recognition device like the JPEG system, there is a problem in processing efficiency. Since the orthogonal expansion method used in the conventional image pattern recognition apparatus is a method different from that used in image compression, the compressed image is once restored to the original image, and then the orthogonal expansion method used in the pattern recognition apparatus is newly created. , The restored image will be processed again by the principal component analysis method. In such a case, since the type of orthogonal expansion method used is different, the processing amount increases without omitting the restoration processing and the orthogonal expansion processing.

As described above, in the conventional image pattern recognition apparatus having the means for performing the detailed recognition processing in combination with the signal / noise separation processing, the sample image having the size of m × n pixels is (m × n) dimensions. The template used in the signal / noise separation processing and the detailed processing and the classification / detail recognition circuit are recognized because the principal component analysis method is used for the signal / noise separation processing and the detailed recognition processing by regarding them as the one-dimensional vector data of It was not possible to efficiently configure the device in a short time according to the change of the target, and the data storage capacity of the device was large, so that the device could not be downsized.

The object of the present invention is to realize high performance from sample images collected in advance in an image pattern recognition apparatus with signal / noise separation processing in view of the advantages and trend of spread of the image compression / decompression method using DCT. In addition to quickly creating templates and classification / detail recognition circuits according to the recognition target, speed up recognition processing of the entire recognition processing system, improve recognition processing performance, and storage capacity of data used by the device An object of the present invention is to provide an image pattern recognition device which has a wide range of application and which has realized downsizing.

[0008]

An image pattern recognition apparatus according to the present invention inputs an image pattern storing means for inputting and storing a binary or multi-valued image pattern and a compressed image pattern obtained by compressing the image pattern. Then, a compressed image pattern storage means for storing, a small area extracting means for extracting a plurality of small areas in the image from the image pattern or the compressed image pattern, and whether or not the small area is similar to the recognition target Selection feature extraction means for creating a selection feature vector for classifying into two categories, and the selection feature vector obtained by the selection feature extraction means is compared with a selection standard pattern prepared in advance for correlation. A selection determination unit that determines a value and classifies the small region into two categories based on the correlation value, that is, whether the small region is similar to the recognition target, and the selection determination unit. Image recognition means for classifying only the small area determined to be similar to the recognition target into a plurality of recognition target categories, and the classification result of the recognition determination means, or each of the means appropriately displays It has image input / output means for inputting adjustment values set for operation.

According to the embodiment of the present invention, the gray level of the small area extracted by the small area extracting means is corrected so as to be within the range of the predetermined gray level, and the predetermined image is obtained. It has a mask processing means for performing mask processing by a filter.

According to another embodiment of the present invention, the orthogonal expansion means is provided for applying the orthogonal expansion to the masked small area by using the DCT by the matrix operation of the transformation matrix to create the feature vector.

According to another embodiment of the present invention, the feature vector created by the orthogonal expansion means is input, and whether or not the masked small area is similar to the recognition target is determined from the components of the feature vector. The selection feature vector creating means for selecting a component effective in the selection process for roughly classifying the or, and recognizing the method of constructing the selection feature vector, and the selection feature extraction means includes the orthogonal expansion means for the feature vector. The selection feature vector is input from the selection feature vector creating means, and the selection feature vector is created from the feature vector according to the configuration method.

According to another embodiment of the present invention, a sorting standard pattern creating means for inputting the sorting feature vector created by the sorting feature extracting means and creating a sorting standard pattern from the sorting feature vector And a selection standard pattern storage means for storing the selection standard pattern.

The image pattern is input by means for inputting and storing the image pattern and means for inputting and storing the compressed image pattern obtained by compressing the image pattern by a known image compression means. Since the means for extracting the small area from the image according to the type of the image is provided, both the processing of the image pattern or the compressed image pattern obtained by compressing the image pattern by the known image compression means can be performed.

In the selection feature vector creating means, a component effective for the selection process is selected from the feature vector for each recognition target, so that the selection feature vector can be newly created for each recognition target and added or updated. In addition, it is possible to reduce the work load associated with the change of the recognition target.

In the orthogonal expansion means applied to the mask processing small area, when the DCT by the matrix operation of the transformation matrix is used, the components of the transformation matrix are defined by the same mathematical expression regardless of the change of the recognition task. In other well-known orthogonal expansion methods such as the principal component analysis method, pre-processing of creating a transformation matrix from learning samples is required every time the recognition task is changed. That is, by using the DCT based on the matrix calculation of the transformation matrix, it is possible to omit the preliminary work performed each time the recognition task is changed, and the work time can be shortened.

Further, when the DCT by the matrix operation of the transformation matrix is used in the orthogonal expansion means applied to the mask processing small area, the components of the transformation matrix are defined by the same mathematical expression regardless of the change of the recognition task, and thus the created transformation is performed. When storing a matrix, the mathematical formula or a transformation matrix created from the mathematical formula may be stored, whereas in other known orthogonal expansion methods such as the principal component analysis method, each time the recognition task is changed. All the elements of the transformation matrix need to be stored. That is,
By using the DCT by the matrix calculation of the conversion matrix, the storage capacity of the conversion matrix in the image pattern recognition processing device can be reduced.

In the orthogonal expansion method for the mask processing small area, the DCT by the matrix operation of the transformation matrix is used.
For example, since a part of known compression means such as JPEG can be used, both the image pattern and the compressed image pattern obtained by compressing the image pattern by the known image compression means can be processed.

When an image compressed by a known image compression means is input, the characteristic vector can be obtained only by partially restoring the compressed image, so that the characteristic vector is created from the compressed image. In this case, the feature vector can be created faster than when the compressed image is completely reconstructed and the orthogonal expansion method is applied to create the feature vector.

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of an image pattern recognition apparatus according to an embodiment of the present invention, FIG. 2 is a block diagram of the image selection circuit 3 of FIG. 1, and FIGS. 3, 4, 5 and 6 are image patterns of FIG. It is a flowchart of a process of a recognition device.

The image pattern recognition device comprises an image pattern storage circuit 1, a compressed image pattern storage circuit 2, an image selection circuit 3, an image recognition circuit 4 and an image input / output device 5.

The image pattern storage circuit 1 receives the input M ×
An image pattern consisting of binary or multi-valued N mesh size is stored. The compressed image pattern storage circuit 2 stores a compressed image pattern obtained by compressing an image pattern consisting of binary or multi-valued M × N mesh size by a known image compression means such as JPEG using DCT. . The image selection circuit 3 inputs the image pattern or the compressed image pattern, performs image selection processing, and detects a small area similar to the recognition target. The image recognition circuit 4
The input image is recognized by classifying the small areas into a plurality of recognition target categories. The image input / output device 5 inputs the information such as the recognition class number of the small area output from the image recognition circuit 4 and the classification result, displays the information, and also externally displays the image information such as a keyboard and a mouse. Input and display the set value, and the image selection circuit 3
The image recognition circuit 4 is caused to input the information regarding the selection processing, and the information regarding the recognition processing is input as appropriate.

The image selection circuit 3 is a main part of the present invention and is shown in FIG.
As shown in FIG. 3, the small area extraction circuit 6, the mask processing circuit 7, the orthogonal expansion circuit 8, the feature extraction circuit 9, the selection determination circuit 10, the selection feature vector creation circuit 11, the selection standard pattern creation circuit 12, and the selection standard. It is composed of a pattern storage circuit 13. Here, the small area extraction circuit 6 inputs the image pattern stored in the image pattern storage circuit 1 or the compressed image pattern stored in the compressed image pattern storage circuit 2 in order to extract a plurality of small areas from the image. , A plurality of small areas are extracted from the image pattern using a small area cutout method according to the type of the image according to a predetermined setting method such as a small area size, a setting interval, and a setting position. The area is output to the mask processing circuit 7. The mask processing circuit 7 corrects the gray level of the small area so as to be within a predetermined gray level range, and
The small area is masked by a predetermined image filter to create a masked small area, and the masked small area is output to the orthogonal expansion circuit 8. The orthogonal expansion circuit 8 applies orthogonal expansion to the mask processing small area to create a feature vector. The orthogonal expansion circuit 8 is provided so that a part of a known image compression means such as JPEG can be used.
For the orthogonal expansion used in, the DCT by the matrix operation of the transformation matrix
(For example, Data Compression Handbook, M. Nelson, Hagiwara, Prentice Hall, Toppan).
The feature vector circuit for selection 11 inputs the feature vector created by orthogonally expanding the mask processing small area by the orthogonal expansion circuit 8 and roughly determines whether the mask processing small area is similar to the recognition target. In order to create a selection feature vector used in the selection process for classification from the feature vector, a component effective in the selection process is selected from the components of the feature vector,
The method of constructing the selection feature vector is determined, and the result is saved. The selection feature extraction circuit 9 inputs the feature vector from the orthogonal expansion circuit 8 and the selection feature vector constructing method determined and stored in the selection feature vector creation circuit 11 and accordingly selects the selection feature from the feature vector. Create a vector. Standard pattern creation circuit 12 for selection
Inputs the selection feature vector created by the selection feature extraction circuit 9, creates a selection standard pattern from the selection feature vector, outputs the selection standard pattern to the selection standard pattern storage circuit 13, and selects the selection standard pattern. The standard pattern storage circuit 13 stores the standard pattern for selection. The selection determination circuit 10 inputs the selection feature vector created by the selection feature extraction circuit 9, inputs the selection standard pattern stored in the selection standard pattern storage circuit 13, and collates with the selection feature vector. Then, the correlation value is obtained, and based on the correlation value, it is roughly classified into two categories, that is, whether the masking small area is similar to the recognition target, and it is determined that it is similar to the recognition target. Only the masked small image is output to the image recognition circuit 4.

Next, the processing of the image pattern recognition processing apparatus will be described with reference to FIGS.

FIG. 3 is a flow chart of processing from the image pattern storage circuit 1 to the small area extraction circuit 6, the mask processing circuit 7, and the orthogonal expansion circuit 8. FIG. 4 is a flowchart of the compressed image pattern storage circuit 2 to the small area extraction circuit 6 and the mask. FIG. 5 is a flow chart of the processing up to the processing circuit 7 and the orthogonal expansion circuit 8. FIG. 5 shows a selection feature vector creation circuit 11, a selection feature extraction circuit 9, a selection standard pattern creation circuit 12, based on the output of the orthogonal expansion circuit 8.
In the flow chart of the processing up to the selection standard pattern storage circuit 13 and the image recognition circuit 4, determination of constituent components of the selection feature vector used in the selection feature extraction circuit 9, and selection standard pattern used in the selection determination circuit 10, And a flow chart relating to the creation and storage of a standard pattern used in the image recognition circuit 4, FIG.
FIG. 3 is a flowchart of the processing from the selection determination circuit 10 to the image recognition circuit 4, showing the processing related to the signal / noise separation processing and the image recognition processing using the standard selection pattern stored in the standard selection pattern storage circuit 13. There is.

First, as a first operation example, when the image pattern storage circuit 1, the small area extraction circuit 6, the mask processing circuit 7, and the orthogonal expansion circuit 8 operate, that is, M × N.
An example of inputting an image pattern of a normal format consisting of binary or multi-valued mesh sizes and creating a feature vector will be described with reference to FIG.

In this image pattern recognition processing device, the secondary
Process a normal format image with pixels originally arranged
In this case, first, in the image pattern storage circuit 1, the processing is performed.
Image pattern I of target M × N mesh size
(X, y), (x = 1, 2, ..., M; y = 1,
2,. . . , N) is input (step 101). Soshi
To correct the gray level of this image pattern I (x, y)
Then, the gray level correction image I _F Create (x, y)
(Step 102). Transformation matrix performed by the orthogonal expansion circuit 8
In the DCT calculation using, the gray level is -128
In order to target the mesh in the range of 127,
Bell corrected image I_F The value of each mesh of (x, y) is -12
Adjust the gray level so that it is in the range of 8 to 127.
Image for DCT I_D Create (x, y) (step
103).

The small area extraction circuit 6 executes processing according to the type of the input image pattern. First, a small region W of m × n mesh from the DCT image I _D (x, y)
(U, v), (u = 1, 2, ..., m; v = 1,
2,. . . , N), the size, number and position of observation windows are set (step 104). Then, the parameters for cutting out the observation window region are determined according to this setting (step 105), and the DCT image I _D
A small area W (u, v) is cut out from (x, y) (step 106). Small area W cut out according to the setting
The actual cutout number of (u, v) is compared with the planned cutout number, and the cutout process is performed until the scheduled cutout is completed (step 107). When it is finished, it is determined whether or not all the cutout processing is finished, and when there is an unprocessed image pattern I (x, y), the above cutout processing is performed (step 108). All image patterns I (x,
When the process of cutting out the small area W (u, v) from y) is completed, the operation of the mask processing circuit 7 is started.

The mask processing circuit 7 also executes processing according to the type of the input image pattern. First, small area W
(U, v) is input, and a Gaussian function g (u, v) is used as a mask pattern and is multiplied for each mesh (u, v). The Gaussian function g (u, v) is expressed by the following equation when the size of the small area W (u, v) is w × w mesh.

[0030]

[Equation 1] The resulting masked small area W _M (u, v) is expressed as follows.

W _M (u, v) = W (u, v) × g (u, v) Next, the mask processing small area W _M (u, v) is set to the image block B (i, j), (i = 1, 2, ...
8; j = 1, 2 ,. . . The process of dividing into 8) is performed for each mask processing small area W _M (u, v) (step 110). For example, the mask processing small area W _M (u, v) having a size of 30 × 30 mesh is divided into image blocks of 8 × 8 mesh arranged in 4 rows and 4 columns.

A block including an end portion generated by division is detected (step 111), and a blank mesh portion in the block is processed (step 112). That is, 30 × 3
Masking small area W _M (u, v) with a size of 0 mesh
Generated when a block is divided into blocks of 8 × 8 mesh
The odd part of 8 mesh or 8 × 6 mesh is processed in step 112. The mesh value of the blank portion of the 8 × 8 mesh block is set to, for example, the average value of the mesh values of the half edges or 0.

Next, for block B (i, j), 8
Perform DCT using the transformation matrix C (i, j) with 8 rows,
An 8 × 8 DCT coefficient matrix D _R (i, j) is created from an 8 × 8 mesh block B (i, j) (step 1
13). Each element of the transformation matrix C (i, j) having 8 rows and 8 columns is represented by the following known formula (for example, Data Compression Handbook, M. Nelson, Hagiwara Translation, Prentice Hall Toppan).

[0034]

[Equation 2] This is performed by matrix multiplication of the block B (i, j) and the block B (i, j). The following known formulas are used for the calculation of DCT (eg, Data Compression Handbook, M. Nelson Hagiwara, Prentice Hall, Toppan).

[0035]

(Equation 3) Here, T represents the transpose of the matrix, and * represents the multiplication of the matrix. Of course, DCT is an example of these calculation methods, and it goes without saying that other calculation methods can be applied to the DCT of the block B (i, j).

8 × 8 DCT coefficient matrix D _R (i, j)
Each component of is a 8 × 8 mesh quantization table Q (i,
Quantize using j) (step 114). The quantization table Q (i, j) used for the DCT coefficient matrix D _R (i, j) of 8 rows and 8 columns has a size of 8 × 8 mesh in advance so that the quantization interval can be adjusted for each component. Create it. For example, a known 8 × 8 mesh quantization table Q (i, j) having the following values is used in consideration of the bias of the spatial frequency distribution in the image pattern (for example,
International standard for multimedia coding, edited by Yasuda, Maruzen).

[0037]

(Equation 4) Of course, this 8 × 8 matrix is the quantization table Q
It is an example of (i, j), and it goes without saying that another matrix of 8 rows and 8 columns having different components can be applied. The i-th row and j-th column component D of the DCT coefficient matrix based on the quantization table Q (i, j)
The quantization process of (i, j) is calculated as follows.

[0038]

(Equation 5) Here, D _Q (i, j) represents the value of the component in the i-th row and the j-th column after the quantization processing, and Round [·] represents the fractional part being rounded down.

Next, the value D _Q (i, j) after the quantization process is inversely quantized using the quantization table Q (i, j), and the DCT integer coefficient matrix D _I (i, j) Is created (step 115). The process ends when the DCT integer coefficient matrix D _I (i, j) can be created for all the blocks B (i, j). The i-th row and j-th column component D _I (i, j) of the DCT integer coefficient matrix is calculated from D _Q (i, j) and Q (i, j) as follows.

D _I (i, j) = D _Q (i, j) × Q (i, j) The DCT integer coefficient matrix D _I (i, j) is the selection feature vector creation circuit 11 or the selection feature extraction circuit 9 Sent to.

Next, as a second operation example, when the compressed image pattern storage circuit 2, the small area extraction circuit 6, the mask processing circuit 7, and the orthogonal expansion circuit 8 operate, that is, in step 115 of FIG. Instead of generating the coefficient matrix D _I (i, j), a compressed image pattern Z _C compressed by a known image compression means such as JPEG is used.
An example will be described in which (k) is input to the image pattern recognition apparatus and the DCT integer coefficient matrix D _I (i, j) is directly extracted from the compressed image Z _C (k).

First, a known JPEG compression method will be roughly described. In JPEG, the input image pattern is divided into 8 × 8 mesh sizes. Then, DCT is performed on the block of 8 × 8 mesh to generate 64 DCT coefficients arranged in a two-dimensional array of 8 rows and 8 columns from one block. Then, the quantization processing is performed on the 64 DCT coefficients. In addition, 64 quantized DC
Zigzag scanning processing is performed on the T coefficient, and the coefficients are arranged in a line. Then, run-length coding and entropy coding processing are performed on the coefficient string arranged in one line. To restore a compressed image, the compression process is performed in reverse order. That is,
Decoding is performed in the order of entropy decoding, run length decoding, inverse zigzag scan processing, inverse quantization processing, and inverse DCT (for example, latest MPEG textbook, Fujiwara Supervisor, ASCII Publishing).

When extracting the DCT integer coefficient matrix D _I (i, j) from the compressed image pattern Z _C (k) instead of creating the DCT integer coefficient matrix D _I (i, j) in step 115 of FIG. An example of the above operation will be described with reference to FIG.
In the compressed image pattern storage circuit 2, first, the compressed image Z _C (k) to be processed is input (step 20).
1). Then, the DCT integer coefficient image D _{Z is applied} from the compressed image Z _C (k) using known entropy decoding, run-length decoding, inverse zigzag scanning process, and inverse quantization process.
(X, y) is created (step 202). This DCT
The mesh value of the integer coefficient image D _Z (x, y) is a DCT coefficient and its size is equal to the size of the restored image. This DCT
If the integer coefficient image D _Z (i, j) is divided into blocks of 8 × 8 mesh, inverse DCT is applied to each block, and the two-dimensional array generated by the conversion is rearranged at the original block position, The original restored image can be generated from the DCT integer coefficient image D _Z (i, j).

The small area extraction circuit 6 executes means corresponding to the type of the input image pattern. First, this DC
To extract a DCT integer coefficient block B _I (i, j) of 8 rows and 8 columns from the T integer coefficient image D _Z (x, y), DC
The size, number, and position of observation windows are set for the T integer coefficient image D _Z (i, j) (step 203). this is,
This corresponds to the process of step 104 in the first operation example shown in FIG. Next, DC of 8 rows and 8 columns included in the observation window
The T integer coefficient block B _I (i, j) is extracted from the DCT integer coefficient image D _Z (x, y) (step 204). Then, the mask processing circuit 7 also executes the means corresponding to the type of the input image pattern. First, the DCT integer coefficient block B _I (i, j) is input and mask processing is performed using the DCT mask pattern g _D (i, j).
This corresponds to the process of step 109 in the first operation example.

The DCT mask pattern g _D (i, j) corresponding to the Gaussian function g (u, v) is created in advance in the following manner.

First, the Gaussian function g (u, v) is set to 8
Divide into blocks of × 8 mesh. Then, DCT is performed using the conversion matrix C (i, j). Next, after performing a quantization process using the quantization table Q (i, j), inverse quantization is performed using the quantization table Q (i, j). The matrix of 8 rows and 8 columns created in this way is used for the DCT mask pattern g _D (i, j).

[0047] In the mask processing of DCT mask pattern g _D where (i, j) by the DCT integer coefficient block B _I (i, j) is, DCT mask pattern g _D (i, j) and DCT integer coefficients block B _I (i, j) for each mesh (i, j)
Take each one. A known DC is used for multiplication for each mesh.
Utilize multiplication operation between components of T matrix (for example,
B. C. Smith and L.M. A. Rowe, Al
gorithms for Manipulating
Compressed images, IEEE C
computer Graphics and Appl
ications, pp. 34-42, Vol. 13,
No. 5, Sep. , 1993). The result of multiplication, DCT
Obtain the integer coefficient matrix D _I (i, j) (step 20)
5). DCT integer coefficient matrix D extracted according to the setting
The actual extraction number of _I (i, j) is compared with the planned extraction number,
The extraction process is performed until the scheduled extraction is completed (step 206). When the extraction is completed, it is determined whether or not all the extraction processing is completed, and if there is an unprocessed compressed image Z _C (k), the above extraction processing is executed (step 207). DCT integer coefficient matrix D _I from all compressed images Z _C (k)
When the extraction process of (i, j) is completed, all the processes are completed. The DCT integer coefficient matrix D _I (i, j)
Is output to the selection feature vector creation circuit 11 or the selection feature extraction circuit 9.

Of course, entropy decoding, run-length decoding, inverse zigzag scan processing, and inverse quantization processing are performed on the entire compressed image Z _C (k) to obtain a DCT integer coefficient image DCT integer coefficient image D _Z (x , Y) to extract the DCT integer coefficient matrix D _I (i, j) is an example, and entropy decoding, run-length decoding, inverse zigzag scan processing, and inverse quantization processing are performed on compressed images. It goes without saying that other methods such as performing locally on Z _C (k) and extracting the DCT integer coefficient matrix D _I (i, j) can be applied.

Next, as a third operation example, a selection feature vector creation circuit 11, a selection feature extraction circuit 9, a selection standard pattern creation circuit 12, and a selection standard pattern storage circuit 1.
3. When the image recognition circuit 4 operates, that is, when the constituent components of the selection feature vector used in the selection feature extraction circuit 9 are determined, and the selection standard pattern used in the selection determination circuit 10 and the image recognition An example of creating and storing a standard pattern used in the circuit 4 will be described with reference to FIG.

First, it depends on whether or not it belongs to the recognition category.
The image pattern that is already known to exist
The small area extraction circuit 6 using the compressed image pattern.
An area is created, and then the orthogonal expansion circuit 8 is created from the small area.
Generated DCT integer coefficient matrix D _I Features for sorting (i, j)
Input to the vector creating circuit 11 (step 301).
In the selection feature vector creation circuit 11, a DCT integer
Coefficient matrix D_I The variance ratio of the (i, j) component of (i, j)
The calculation is performed like a step (step 302).

[0051]

(Equation 6) Here, F (i, j) is the dispersion ratio of the (i, j) component,

[0052]

[Outside 1] Is the mean and variance for the (i, j) component of the DCT integer coefficient matrix D _I (i, j) known to belong to the recognition category,

[0053]

[Outside 2] Represents the mean value and variance for the (i, j) component of the DCT integer coefficient matrix D _I (i, j) known not to belong to the recognition category.

Among the components of the DCT integer coefficient matrix D _I (i, j), the component having a large variance ratio is selected and used as the constituent component of the selection feature vector F. The selection feature vector F
It is possible to newly create, update, or change the constituent components of each of the recognition tasks. Further, the image input / output device 5 sets the selected number. It goes without saying that this setting can be made in advance (step 303).

DC generated by the small area extraction circuit 6 from the image pattern or the compressed image pattern by the selection feature circuit 9.
The T integer coefficient matrix D _I (i, j) is input, and the components selected from the DCT integer coefficient matrix D _I (i, j) are arranged according to the determination by the selection feature vector creation circuit 11, and the selection is performed. Feature vector F = (f ₁ , f ₂ , ..., F
_s ,. . . , F _L ). When the number of selected components is set to L, the number of dimensions of the selection feature vector is L (step 304). For example, in the first operation example, for the image pattern W (u, v) whose size is 30 × 30 mesh and is known to belong to the recognition category,
Divide into 16 blocks of 8 × 8 mesh size,
DCT integer coefficient matrix D _{I of} 8 rows and 8 columns from each block
Create (i, j). Then, for the 64 components of the 16 DCT integer coefficient matrix D _I (i, j), the average value and the variance for samples belonging to the recognition category and samples not belonging to the recognition category are calculated in each case, and (i,
j) Calculate the variance ratio for each component (step 302).
This large variance ratio L is selected from the 64 × 16 variance ratios (step 303), and the selection feature vector F is created by the component corresponding to the variance ratio (step 30).
4). Of course, this method of constructing the selection feature vector F is an example, and it goes without saying that another method can be applied to the creation of the selection feature vector F.

The standard pattern for selection circuit 12 creates a standard pattern G _F used in the selection circuit 10 from the feature vector F. A known clustering method such as the k-means method is applied to the plurality of feature vectors F created by the operation of the selection feature extraction circuit 9 to calculate Q clusters (step 305). The number Q of clusters is set in advance via the image input / output device 5. The k-means method here is an example of a clustering method, and it goes without saying that other clustering methods can be applied to the generation of clusters.

The standard pattern used in the selection determination circuit 10 is created based on the result of this clustering processing. For example, first, in the L-dimensional feature space consisting of the selection feature vector F, Q ^L box-shaped ranges that cover Q clusters are set. That is, the selection range for the cluster q

[0058]

[Outside 3] Is set for each component so that the component f _s of the feature vector F satisfies the following equation (step 306).

[0059]

(Equation 7) Here, Pr {•} is the probability, T _B is the image input / output device 5
Is a threshold value (0 <T _B <1) set via the, and α is a constant set via the image input / output device 5,

[0060]

[Outside 4] Represents the average value and standard deviation of the cluster q with respect to the f _{s -th} component.

Next, in the standard pattern forming circuit 12 for selection,
Is a component f of the selection feature vector F _s (S =
1 ,. . . , L), f_s Is the selection range

[0062]

[Outside 5] Boolean variable g _s = 1 if any of the above is included. Otherwise, assign the Boolean variable g _s = 0. As a result, an L-bit binary number G is created from the L-dimensional selection feature vector F.

A DCT integer coefficient matrix D _I (D _I ) created by operating the orthogonal expansion circuit 8 on the small area created by the small area extraction circuit 6 from the image pattern or compressed image pattern belonging to the class to be recognized in the learning data. i, j) steps 301, 302, 303, 304, 305, 3
By operating 06, an L-bit binary number G is created, and a set G _F = {G | L-bit binary number created from a learning image pattern belonging to the class to be recognized} is created, and a standard pattern is created. (Step 306).

The operation of step 301 described here is as follows.
This is an application of a known standard pattern creation method, and it is an example of the standard pattern creation method, and it goes without saying that other methods can be applied to the creation of the standard pattern (for example, Arimura, Hagita: Statistical image recognition). Method for Constructing Image Selection Space, PRU95-147, Communication Technique, 199
5). Selection range

[0065]

[Outside 6] The standard pattern G _F and the standard pattern G _F are input to and stored in the selection standard pattern storage circuit 13 (step 307). Selection feature vector creation circuit 11 due to change in recognition task
When the constituent component of the sorting feature vector F is newly created / updated / changed due to the operation of, the sorting standard pattern creating circuit 12 and the sorting standard pattern storage circuit 13 are operated to select the sorting range.

[0066]

[Outside 7] It goes without saying that the standard pattern G _F is newly created, updated, and changed.

The selection determining circuit 10 uses the L-dimensional selection feature vector F created from the small area created by the small area extracting circuit 6 from the image pattern or the compressed image pattern to make the small area similar to the recognition target. It is roughly classified whether or not there is. Classification is selection range

[0068]

[Outside 8] L created from the L-dimensional sorting feature vector F using
The binary number G of bits and the standard pattern G _F are compared in the following manner (step 308). That is, L created from the L-dimensional screening feature vector F created by operating the screening feature extraction circuit 9 in the small area created by the small area extraction circuit 6 from the image pattern or the compressed image pattern.
For the binary number G of bits, if GεG _F , the small area is determined to be a candidate image, and otherwise, the small area is determined to be a rejected image. Only the L-dimensional selection feature vector F for the small area determined as the candidate image is input to the image recognition circuit 4. That is, the L-dimensional sorting feature vector F corresponding to the selected small area is set as the sorting feature vector F _S (step 309).

In the image recognition circuit 4, the selection feature vector F
Enter the _S, for example, to create an identification function R (F _S) for recognition processing on the basis of the known Bayesian classifier rule (Step 3
10). That is,

[0070]

(Equation 8) here,

[0071]

[Outside 9] Represent the mean vector, the covariance matrix, and the occurrence probability of the selection feature vector F _S belonging to the recognition target class C, respectively. The generated discriminant function R (F _S ) is stored,
The process ends (step 311). Of course, it goes without saying that the recognition method using the identification function based on the Bayes identification rule is an example of recognition after the selection determination process, and other known recognition methods can be applied.

Finally, as a fourth operation example, when the selection feature extraction circuit 9, the selection feature vector creation circuit 11, the selection determination circuit 10, the selection standard pattern storage circuit 13, and the image recognition circuit 4 operate, that is, According to the first operation example, an image pattern of a normal format consisting of binary or multi-valued M × N mesh size, or the second operation example, the image pattern is a known image compression such as JPEG. An example will be described in which a compressed image pattern compressed by the means is input and the recognition result is displayed.

First, the selection range created in step 306

[0074]

[Outside 10] Further, the standard pattern G _F is inputted from the standard pattern storage circuit 13 for selection, and the discrimination function R (F _S ) is inputted from the image recognition circuit 4 (step 401). DC from orthogonal expansion circuit 8
Input the T integer coefficient matrix D _I (i, j) (step 4
02). The designation of the constituent components of the selection feature vector F determined by the selection feature vector creation circuit 11 is input from the selection feature vector creation circuit 11 (step 403). According to the designation of the component, the process of step 304 is performed,
A selection feature vector F is created from the DCT integer coefficient matrix D _I (i, j) (step 404). The process of step 308 is performed on the selection feature vector F, and the selection determination process is performed from the selection feature vector F (step 405). If the determination result is a rejected image (step 406), all DCT integer coefficient matrices D _I (i,
It is judged whether or not the process for j) is completed (step 407), and if all are processed, the process is completed. If there is an unprocessed DCT integer coefficients matrix D _I (i, j), the process returns to step 402 to input the raw DCT integer coefficients matrix D _I (i, j), the process is repeated. If the determination result is a candidate image (step 406), the corresponding feature vector F is recognized using the discriminant function R (F _S ) (step 408), and the recognition result is the image input / output device 5
Output to and displayed (step 409). All DC
It is determined whether the processing for the T integer coefficient matrix D _I (i, j) is completed (step 410), and if all are processed, the processing is completed. Otherwise, step 4 to input the raw DCT integer coefficient matrix D _I (i, j)
Returning to 02, the process is repeated.

In the above first, second, third and fourth operation examples, the setting values of the respective circuits can be changed as appropriate according to the result of the selection judgment process or the recognition process from the image input / output device 5. It is possible to adjust the operation of each circuit.

As described above, according to the first, second, third and fourth operation examples, the image pattern storage circuit 1, the compressed image pattern storage circuit 2, the small area extraction circuit 5, the mask processing circuit 7, Since the orthogonal expansion circuit 8 is provided, an image pattern of a normal format consisting of binary or multi-valued M × N mesh size, or a compressed image pattern obtained by compressing the image pattern by a known image compression means such as JPEG. Both recognition processes can be performed. In addition, since the DCT using the transform matrix capable of high-speed calculation is used as the orthogonal expansion method for creating the feature vector from the image pattern, high-speed processing becomes possible as compared with the case of using other orthogonal expansion methods. , The processing efficiency is improved. In particular, by using a method of performing a discrete cosine transform of a block divided into 8 × 8 meshes to create a feature vector, the process of creating a feature vector conforms to a known JPEG compression method, and a known JPEG decompression method. Since a part of the compressed image pattern can be used to create a feature vector, a part of the compressed image pattern restoration process can be omitted in the process for the compressed image pattern, which improves the processing efficiency. Remarkably improved.

Further, the selection judgment circuit 10 and the image recognition circuit 4
Since the image patterns are roughly classified and recognized in more detail, the image patterns are recognized by multi-step recognition including two-step processing of large classification and recognition. Therefore, performance in recognition speed and recognition accuracy is improved. A high image pattern recognition device can be realized.

[0078]

As described above, the present invention has the following effects. (1) According to the invention of claim 1, a binary or multi-valued image pattern and the image pattern are, for example, JPEG.
A means for inputting and storing a compressed image pattern compressed by a known image compression means such as a format, a means for extracting the image pattern and a plurality of small areas from the compressed image pattern, and a processing small area as a recognition target. Selection feature extraction means for creating a selection feature vector for roughly classifying into two categories of similarity and whether the processing small region is similar to the recognition target based on the selection feature vector. The classification determining means for roughly classifying into the two categories, and only the processing small area determined to be similar to the recognition target by the classification determining means are classified into the recognition target category based on the classification feature vector. By providing a means for
In the image recognition processing by the multi-step processing, an image pattern of a normal format in which the format of an input image is binary or multi-valued having a size of M × N mesh, or a known image compression means such as JPEG is used for the image pattern. By this, it becomes possible for the user to process the compressed image pattern without being aware of it. (2) The invention according to claim 3 uses the DCT by the matrix operation of the transformation matrix for the orthogonal expansion means applied to the mask processing small area, so that the components of the transformation matrix are the same regardless of the change of the recognition task. Since it can be defined by a mathematical formula, in other known orthogonal expansion methods such as the principal component analysis method, a transformation matrix is created based on the learning sample that is required every time when the recognition task is changed. The pre-processing for creating the matrix can be omitted, the creation time of the conversion matrix can be shortened, and moreover, when storing the created conversion matrix,
While it is sufficient to store the equation or the transformation matrix created from the equation, in other known orthogonal expansion methods such as the principal component analysis method, all the elements of the transformation matrix are stored every time the recognition task is changed. There is a need to. That is, by using the DCT by the matrix calculation of the conversion matrix, there is an effect that the storage capacity of the conversion matrix in the image pattern recognition processing device can be reduced. Further, since the DCT by the matrix operation of the transformation matrix can be used for the orthogonal expansion means, a part of the known image compression means such as JPEG can be used, so that the image is compressed by the known image compression means. When a compressed image is input, the feature vector can be obtained by only partially restoring the compressed image. Therefore, when creating the feature vector from the compressed image, the compressed image is completely orthogonal to the restored image. Compared with the case where the expansion method is applied to create the feature vector, there is an effect that the feature vector can be created faster. (3) According to the invention of claim 5, the sorting feature vector used in the sorting determination means is configured for each recognition target with a component effective in the sorting determination means, and the sorting feature vector creating means is associated with a change in the recognition task. By operating the selection feature vector creation means capable of newly creating / updating / changing the constituent components of the selection feature vector, and the selection determination based on the selection feature vector created according to the setting of the selection feature vector creation means. When a standard feature pattern for selection is newly created / updated / changed, the standard pattern selection means for creating a standard pattern for selection used in the means and the standard pattern storage means for selection that stores the standard pattern for selection are provided. Automatically creates new sorting feature vectors and sorting standard patterns for each recognition target by creating, updating, and changing new sorting standard patterns. To create, add, because you have to be able to update, there is an effect that can reduce the work load on the user due to the change of the recognition target.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an image pattern recognition device according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of an image selection circuit 3 in FIG.

3 is a diagram illustrating an image pattern recognition apparatus of FIG.
7 is a flowchart of a process of creating a feature vector by inputting a normal-format image pattern having binary or multi-valued mesh sizes.

FIG. 4 is a diagram showing an image pattern recognition apparatus of FIG. 1 in which an input image format is a normal format image pattern consisting of binary or multi-valued M × N mesh size by a known image compression means such as JPEG. It is a flow chart of processing which creates a compressed compressed image pattern feature vector.

FIG. 5 is a diagram showing an image pattern recognition device of FIG. 1 in which a feature vector created by the procedure shown in the flowchart of FIG. 3 or 4 is input, and a component effective in the selection determination process is selected from the components of the feature vector. Create a vector,
9 is a flowchart for creating a standard pattern and a dictionary table used in the selection determination process and the image recognition process.

6 is a flowchart of a process of displaying a recognition result based on the feature vector created by the procedure shown in the flowchart of FIG. 3 or 4 in the image pattern recognition apparatus of FIG.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 image pattern storage circuit 2 compressed image pattern storage circuit 3 image selection circuit 4 image recognition circuit 5 image input / output device 6 small area extraction circuit 7 mask processing circuit 8 orthogonal expansion circuit 9 selection feature extraction circuit 10 selection determination circuit 11 selection feature Vector creation circuit 12 Standard pattern creation circuit for selection 13 Standard pattern storage circuit 101 for selection 201-207, 301-311, 4
01-410 steps

Claims (5)

[Claims] 1. An image pattern recognition device for recognizing an image pattern by multi-step processing including two-step processing of large classification and recognition, and an image pattern storage means for inputting and storing a binary or multi-valued image pattern. A compressed image pattern storing means for inputting and storing a compressed image pattern obtained by compressing the image pattern; a small area extracting means for extracting the image pattern or a plurality of small areas in the image from the compressed image pattern; A selection feature extraction unit for creating a selection feature vector for classifying into two categories depending on whether the region is similar to the recognition target, and the selection feature vector obtained by the selection feature extraction unit are prepared in advance. A correlation value is obtained by collating with a selected standard pattern for selection, and whether or not the small area is similar to the recognition target is determined based on the correlation value. A selection determining unit that classifies the categories into one category; an image recognition unit that classifies only the small area that is determined to be similar to the recognition target by the selection determining unit into a plurality of recognition target categories; An image pattern recognition apparatus comprising image input / output means for displaying a classification result or for inputting an adjustment value set for the respective means to operate properly. 2. A mask that corrects a gray level of a small area extracted by the small area extraction unit so as to be within a range of a predetermined gray level, and performs mask processing by a predetermined image filter. The image pattern recognition device according to claim 1, further comprising processing means. 3. The image pattern recognition apparatus according to claim 2, further comprising an orthogonal expansion means for applying a orthogonal expansion to the masked small area by using a DCT by a matrix operation of a conversion matrix to create a feature vector. 4. A selection process for inputting a feature vector created by the orthogonal expansion means and roughly classifying whether or not the masked small region is similar to the recognition target from the components of the feature vector. A selection feature vector creating means for selecting a valid component and recognizing a method of constructing a selection feature vector, wherein the selection feature extraction means inputs the feature vector from the orthogonal expansion means, 4. The image pattern recognition apparatus according to claim 3, wherein said configuration method is input from said selection feature vector creating means, and a selection feature vector is created from said feature vector according to said configuration method. 5. A sorting standard pattern creating unit for inputting a sorting feature vector created by the sorting feature extracting unit and creating a sorting standard pattern from the sorting feature vector, and storing the sorting standard pattern. The image pattern recognition apparatus according to claim 4, further comprising a standard pattern storage means for selection.