JP2008152611A - Image recognition device, electronic equipment, image recognition method, and image recognition program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To quickly and precisely recognize a complicated object in an image. <P>SOLUTION: This image recognition device/method is provided with an image input part 10 for inputting the image photographed with a face as an example of an object, a binarization-normalizing part 13 for binarizing each pixel of the image, and for simplifying the image so as to calculate a local directional contribution feature amount, a directional contribution feature amount calculating part 14 for calculating the local directional contribution feature amount showing feature amounts of the binarized image, and a recognition part 15 for recognizing the face photographed in the image, based on the local directional contribution feature amount. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an image recognition technique for recognizing an object shown in an image.

Conventionally, face recognition technology for facial images with a face is a technology that performs face recognition using face-specific color information such as skin color and information about facial parts (such as eyes, nose, mouth, and eyebrows) such as contour lines as feature quantities. (For example, refer to Patent Documents 1 and 2) Various techniques such as a technique for performing face analysis by performing frequency analysis on a face image and extracting a feature amount (for example, refer to Patent Document 3) have been proposed. .
In recent years, in order to further improve the recognition accuracy based on facial features extracted from facial images, advanced learning methods such as neural networks such as support vector machines and boosting techniques are added to the recognition process. The recognition rate is increased, which enables practical use in application fields such as security systems that require high recognition accuracy.

On the other hand, character pattern recognition technology for recognizing character patterns has been studied for a long time.
In the 1970s, it was already applied to postal code reading technology. At that time, it was necessary to recognize character patterns using hardware that lacked computing power, so methods that are simple and that can achieve a high recognition rate are being studied. As this technique, a technique using a character line direction contribution feature amount has been proposed, which enables highly accurate recognition even with handwritten characters, and is used for recognition performance evaluation and various technique references (for example, (See Patent Document 4 and Non-Patent Document 1).
JP 2005-141437 A JP 2001-283224 A JP 2004-272326 A JP-A-57-164376 Norihiro Hamada, Seiichiro Naito, Isao Masuda, “A Handwritten Kanji Recognition Method Based on Global and Local Directional Contribution Density Features”, IEICE Transactions, IEICE, June 1983, Vol. J66-D, no. 6, pp. 722-729

By the way, although the above face recognition technology enables high-accuracy face recognition, the calculation algorithm is complicated and it is difficult for a small electronic device such as a digital camera or PDA whose processing capability is limited to process. A dedicated device is required. On the other hand, character pattern recognition technology
Although it is possible to provide a calculation algorithm that can be processed using hardware with low processing capability, it is difficult to recognize a face containing many complex line segments using this recognition technique.
The present invention has been made in view of the above-described circumstances, and provides an image recognition apparatus, an electronic device, an image recognition method, and an image recognition program capable of recognizing a complex object captured in an image with high speed and high accuracy. For the purpose.

[Mode 1] In order to achieve the above object, an image recognition apparatus according to mode 1 comprises an image input unit for inputting object image data including an object, and the object image data input by the image input unit. A binarization unit that binarizes each pixel using a predetermined luminance threshold value, and a direction contribution feature amount of a contour line included in a predetermined region of the object image data binarized by the binarization unit Based on the local amount contribution feature amount calculated by the feature amount calculation unit that calculates the local direction contribution degree feature amount and the feature amount calculation unit,
And a recognition unit for recognizing an object included in the object image data.

According to such a configuration, since the object image data is simplified by binarization, it is possible to use an algorithm using the direction contribution feature amount for recognizing an object image including many complex line segments. Thus, it is possible to recognize a complex object in an image at high speed and with high accuracy by simple processing.
Here, the object means an object for image recognition, and any object may be used for image recognition, such as a face, an animal, an object such as a car or a vase.
Further, the direction contribution feature amount is a feature obtained by extracting the direction information of the line segment. For example, a certain pixel has four directions of a horizontal direction, a right oblique 45 degree direction, a vertical direction, and a left oblique 45 degree direction. ,
This is a feature amount indicating how long the target pixel is continuous. Thereby, the geometric shape such as the direction of the line segment and the connection relation can be quantitatively obtained.
The image input unit may have any configuration as long as it can input object image data. For example, the image input unit may input object image data from an input device, an external device, or the like. The object image data may be acquired (acquired) or received from the image data, or the object image data may be read from a storage device or a storage medium. Therefore,
Input includes at least acquisition, reception and reading.

[Mode 2] Furthermore, the image recognition apparatus according to mode 2 is the image recognition apparatus according to mode 1, and for each pixel constituting the binarized object image data, the effective pixels are integrated for each column and / or row. The brightness threshold value is changed in accordance with the projection distribution.
According to this configuration, the complexity of the line segment of the object image is obtained as a projection distribution obtained by accumulating effective pixels for each column or / and row for each pixel constituting the object image data. Then, by changing the binarization luminance threshold according to the obtained projection distribution, a luminance threshold suitable for calculating the local direction contribution feature amount is used for various object image data. Binarization can be performed.
Here, the effective pixel means a pixel in which the pixel is determined to be effective by binarization.
For example, when black is valid and white is invalid, it means a black pixel. Here, it goes without saying that the concept of effective pixels can also be applied to binarization using two different colors other than black and white.

[Mode 3] Further, in the image recognition device according to mode 3, in the image recognition device according to mode 2, the binarization unit calculates an entropy for quantifying the state of the distribution of the projection, and according to the calculated entropy. And changing the brightness threshold. According to this configuration, the complexity of the projection distribution can be quantified as entropy, and the luminance threshold value in binarization can be appropriately set according to the entropy value.

[Mode 4] Furthermore, the image recognition device according to mode 4 is the image recognition device according to any one of modes 1 to 3, wherein the binarization unit is prior to binarization of the object image data, or After binarization, a filtering process for smoothing the line segment is performed, and the feature amount calculation unit performs the local direction contribution feature on the object image data binarized and filtered by the binarization unit. An amount is calculated.
According to this configuration, since the line segment of the object image data is smoothed by the filtering process, the photographing environment is bad and the binarization alone cannot be sufficiently simplified, and it is difficult to calculate the local direction contribution feature quantity. It is also possible to stably calculate the local direction contribution feature amount for the object image.

[Mode 5] Furthermore, the image recognition device according to mode 5 is the image recognition device according to any one of modes 1 to 4, wherein the object is a face, and the local direction contribution feature quantity for each of a plurality of different faces. And a storage unit that stores dictionary data for face recognition including the local direction contribution feature amount calculated by the feature amount calculation unit for the face included in the object image data, and the storage A face included in the object image data is recognized based on a local direction contribution feature amount included in the face recognition dictionary data stored in the section. According to this configuration, it is possible to perform face recognition with high speed and high accuracy with a simple process on a face image including many complex line segments.
Here, the face means an object corresponding to a face such as a person, an animal, or a robot imitating them.

[Mode 6] Further, an electronic device according to mode 6 includes the image recognition device according to any one of modes 1 to 5.
According to this configuration, an electronic device having an image recognition function can be provided.
Here, the electronic device may be any electronic device such as a camera, a scanner, a projector, a television, or a printer.

[Mode 7] Furthermore, in the image recognition method according to mode 7, an image input step for inputting object image data including an object and each pixel constituting the object image data input by the image input step A binarization step that binarizes using a luminance threshold of the image, and a local direction contribution that is a direction contribution degree feature amount of a contour line included in a predetermined region of the object image data binarized by the binarization step A feature amount calculation step for calculating a degree feature amount, and a recognition step for recognizing an object included in the object image data based on the local direction contribution feature amount calculated by the feature amount calculation step. It is characterized by.
Thereby, an effect equivalent to that of the image recognition apparatus according to mode 1 is obtained.

[Embodiment 8] Furthermore, the image recognition program according to Embodiment 8 is a computer that includes an image input unit that inputs object image data including an object, and each pixel that constitutes the object image data input by the image input unit. And binarizing means for binarizing using a predetermined luminance threshold, and a local direction contribution feature amount of a contour line included in a predetermined area of the object image data binarized by the binarizing means A feature amount calculating means for calculating a target direction contribution feature quantity, and a recognition means for recognizing an object included in the object image data based on the local direction contribution feature quantity calculated by the feature quantity calculation means. It is made to function.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operations and effects as those of the image recognition apparatus according to mode 1 are obtained.

[Mode 9] Furthermore, the recording medium of mode 9 includes a computer that includes an image input unit that inputs object image data including an object, and each pixel that constitutes the object image data input by the image input unit. , Binarization means for binarization using a predetermined luminance threshold value, and local direction contribution feature amounts of contour lines included in a predetermined area of the object image data binarized by the binarization means A function amount calculation unit that calculates a direction contribution feature amount, and a recognition unit that recognizes an object included in the object image data based on the local direction contribution feature amount calculated by the feature amount calculation unit It is a computer-readable recording medium on which an image recognition program to be recorded is recorded.
With such a configuration, when the program is read from the storage medium by the computer and the computer executes processing in accordance with the read program, the same operation and effect as those of the image recognition apparatus according to mode 1 can be obtained.
Here, the storage medium is a semiconductor storage medium such as RAM or ROM, a magnetic storage type storage medium such as FD or HD, an optical reading type storage medium such as CD, CDV, LD, or DVD, or a magnetic storage type such as MO. / Optical reading type storage medium, and any storage medium can be used as long as it can be read by a computer regardless of electronic, magnetic, optical, etc. .

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a diagram showing a configuration of an image recognition system 1 according to the present embodiment.
The image recognition system 1 recognizes the person 9 in the photographed image by performing image recognition on the camera 2 that shoots the face of the person 9 as an example to generate image data and the image data of the camera 2. A computer 3 serving as an image recognition device, and a display device 4 connected to the computer 3. The camera 2 is a digital camera that captures a still image, and image data of the captured still image is input to the computer 3. Note that a video camera that captures a moving image may be used as the camera 2, and in this case, individual frames (image data) of the captured data are input to the computer 3. The display device 4 displays image data to be recognized, image recognition results, and the like.

FIG. 2 is a block diagram showing a functional configuration of the computer 3.
In this figure, an image input unit 10 is a connection interface with the camera 2, and image data of the camera 2 is input from the image input unit 10. The face area extraction unit (object area extraction unit) 11 applies the image area to the face area 40 from the image area based on the skin color that is the determination criterion of the face area 40 (see FIG. 4) that is the image area where the face is captured. To extract. The storage unit 12 includes various data such as skin color reference color data 16 and face recognition dictionary data 17 that are skin color information serving as a determination reference for the face area 40, and an image recognition program that causes the computer 3 to function as an image recognition device. Various programs such as 18 are stored.
The computer 3 includes a CPU as calculation execution means, a ROM and a CPU as storage means.
The image recognition program 18 is stored in the ROM. When the CPU executes the image recognition program 18, the functions of the respective units are realized from the image input unit 10 described above.

The binarization / normalization unit 13 binarizes each pixel of the face area 40 into a black pixel and a white pixel,
The center of the mesh area 25 (see FIG. 3) used when calculating the local direction contribution feature amount described later is aligned with the center of the face area 40, and the face area 40 is within the mesh area 25. Normalization for enlarging or reducing the face area 40 is performed. The direction contribution feature amount calculation unit 14
A local direction contribution feature amount, which will be described later, indicating the feature amount of the face region 40 is calculated. The face recognition unit 15
The face in the face area 40 is recognized based on the local direction contribution feature quantity of the face area 40 and the face recognition dictionary data 17. It should be noted that each of these units is processed by the computer 3 using the image recognition program 18
Of course, each unit may be configured by a dedicated hardware circuit.

As described above, in this embodiment, the local direction contribution feature amount is used for face recognition. This local direction contribution feature amount is published in the above-mentioned Non-Patent Document 1, and will be briefly described here by taking character pattern recognition as an example.
The direction contribution feature amount is one of features obtained by extracting the direction information of the character line, and a certain pixel is subject to four directions of the horizontal direction, the right oblique 45 degree direction, the vertical direction, and the left oblique 45 degree direction. This indicates how long the pixels (for example, black pixels) are continuous, and this indicates the geometric shape such as the direction of the character lines and the connection relationship.
As shown in FIG. 3A, the direction contribution degree dm (m = 1, 2, 3, 4) is determined by extending the tentacles in the eight directions from the point P1 in the character line, and the black dot connection length li in each direction. (i = 1, 2, ...
8) is used to express the following equation.

In the above formula (1), d1 is a horizontal direction, d2 is a right oblique 45 degree direction, d3 is a vertical direction, and d4 is a direction contribution component in a left oblique 45 degree direction. When the direction contribution d1 is specifically calculated, the following equation is obtained from the above equation (1).

In the example shown in FIG. 3A, since the black pixels are connected in the vertical direction for a relatively long time at the point P1, the value of the direction contribution d3 in the vertical direction is large. Thus, the direction contribution degree dm can quantitatively evaluate the connection relationship together with the direction of the character line.
The direction contribution is a global direction contribution density (GDCD).
Two are known: y) and Local Direction Contributivity Density (LDCD). The global direction contribution characteristic is a characteristic that captures the difference between the character line direction and the connection relation in a global manner. On the other hand, the local direction contribution feature reflects the difference in the direction of the character line and the connection relationship in addition to the line density in the predetermined local region.

As shown in FIG. 3B, the local direction contribution feature is obtained by dividing the character pattern into m × m coarse mesh regions 25 and obtaining direction contributions for all black pixels in each mesh region 25. These are obtained as an average obtained by projecting them in four directions for each mesh region 25. Therefore, four feature values are obtained for each mesh region 25, and m × m × 4 feature values are obtained for the entire character.
As described above, the local direction contribution feature is obtained by dividing the local feature of the character pattern in the segmentation state obtained by dividing the global feature of the character pattern by the m × m mesh regions 25. It is obtained from the degree of contribution, and it can be said that the features of both global features and local features are fused. Therefore, it can be said that it reflects the external features of the character pattern, the connection of the character line segments, and the detailed shape.
In the present embodiment, this local direction contribution feature is used as a direction contribution feature, and the direction contribution feature amount calculation unit 14 shown in FIG. 2 uses the face region 40 as a local region, and this face region. Forty local direction contribution feature quantities are calculated.

However, in order to calculate the local direction contribution feature amount, it is necessary that the local region is a binary image and that a line segment is clearly extracted. In general, an image of a face is a multi-valued image such as a color image or gray scale, and further contains many unclear lines due to facial wrinkles, etc., so it is suitable for calculating the local direction contribution feature amount. It is not an image.
Therefore, the binarization / normalization unit 13 binarizes each pixel of the face region 40 extracted by the face region extraction unit 11 to such an extent that the local direction contribution feature quantity can be calculated.

More specifically, when the face area 40 extracted from the image data is a grayscale image,
When each pixel is binarized with a predetermined luminance threshold value (for example, half the maximum luminance value), FIG.
As shown, each pixel is binarized into a black pixel or a white pixel. However, in this state, many unclear lines and fine lines usually remain due to facial wrinkles and cheeks.
Here, it is assumed that the face area 40 is configured by arranging Q × R pixels in a matrix, and when the horizontal direction is the X axis and the vertical direction is the Y axis, black pixels are provided for each vertical column. X-axis projection dx (q) (where 1 ≦ q ≦ Q), and the cumulative number obtained by integrating the number of black pixels for each row in the Y-axis direction. The Y-axis projection dy (r) (where 1 ≦ r ≦
R).

As shown in FIG. 4A, when the number of line segments included in the face area 40 is large, black pixels are distributed over a wide range of the face area 40, and thus the X-axis projection dx (q) and the Y-axis projection dy. The distribution of (r) is
The drop of the peak is small and the overall curve becomes smooth.
On the other hand, in order to calculate the local direction contribution feature quantity of the face area 40, as shown in FIG. 4B, the line segments included in the face area 40 include left and right eyebrows 41 and 42, and left and right eyes 43. , 44, nose 45, and mouth 46, the face area 40 needs to be simplified to the extent that only the contour lines of the facial parts are present. In the face area 40 thus simplified, the X-axis projection dx (q) and the Y-axis projection d
The distribution of y (r) is a curve having a large peak drop and a sharp plurality of peaks as a whole.

Accordingly, by quantifying the distribution state (complexity) of the X-axis projection dx (q) and the Y-axis projection dy (r), the line segment included in the face region 40 contributes to the local direction by binarization. It is possible to determine whether or not the degree feature amount has been simplified to a degree that can be calculated. In this embodiment, entropy is used for this quantification. The projection entropy Ex of the X-axis projection dx (q) and the projection entropy Ey of the Y-axis projection dy (r) are obtained by Expression (3) and Expression (4), respectively. In the following equations (3) and (4), Nx and Ny are X-axis projections d, respectively.
The total number of x (q) and Y-axis projection dy (r).

In the example shown in FIG. 4A, since the distribution of the X-axis projection dx (q) and the Y-axis projection dy (r) is gentle, the projection entropy obtained from the above equations (3) and (4). Ex and Ey become smaller, and in the example shown in FIG. 4B, the X-axis projection dx (q) and the Y-axis projection dy (r)
The projection entropy Ex and Ey are large because there are many peaks in the distribution and the shape is sharp.
That is, the larger the binarization luminance threshold for the image of the face area 40 is, the more the face area 40 becomes.
Since the number of black pixels contained in the image is reduced and the image (line segment) is simplified, the projection entropies Ex and E
y increases. Therefore, until the projection entropy Ex, Ey exceeds the projection entropy threshold Eth indicating that the image has been simplified to the extent that the local direction contribution feature quantity can be calculated,
By increasing the binarization luminance threshold and binarizing the face area 40, an image of the face area 40 capable of calculating the local direction contribution feature amount is obtained. Face region 40 capable of calculating local direction contribution feature quantity
As shown in FIG. 4B, the face area 40 includes line segments of outlines of facial parts such as left and right eyebrows 41 and 42, left and right eyes 43 and 44, nose 45 and mouth 46. This is an image that only contains approximately. If the face area 40 is further simplified, the outline of the face part is deformed or deleted, and information necessary for face recognition is lost.
Note that the projection entropy threshold Eth depends on the size of the mesh region 25, the resolution of the image, and the like, and is thus experimentally obtained in advance. The projection entropy threshold Eth is set for each of the X-axis projection entropy Ex and the Y-axis projection entropy Ey.

Next, the operation of this embodiment will be described.
5 and 6 are flowcharts of the image recognition program 18 executed by the computer 3. As shown in FIG. 5, when image data showing a person 9 is input from the camera 2 to the image input unit 10 (step S1), the face region extraction unit 11 extracts a face region 40 from the image region of the image data. (Step S2). The extraction of the face area 40 is performed using area separation that separates the areas of the objects in the image. More specifically, as shown in FIG. 7A, in general, the image 20 to be recognized includes a background together with the person 9, and in the region separation, the background is reflected in the object or background in the image 20. Each region of the object is separated.

For example, in the image 20 in which the signboard 21, the mountain 22 and the sky 23 are reflected in the background of the person 9, as shown in FIG. 7B by the area separation, the separation area 30 and the signboard separated from the person 9 area. Separation region 31 from which 21 regions are separated, and separation regions 32A to 32C from which mountain 22 regions are separated
The image area is separated into the separation area 33 obtained by separating the sky 23 area.
Then, the face area extraction unit 11 compares the color information of the face area with the skin color reference color data 16 from each of the separation areas 30 to 33, specifies an area containing a lot of skin colors, and FIG. As shown
The separation area 30 including the person 9 is specified. Further, the face area extracting unit 11 selects the face area from the separation area 30 based on the skin color reference color data 16 and the outline included in the separation area 30 as shown in FIG. 40 is extracted.

When the face area 40 is extracted in this way, the binarization / normalization unit 13 converts the face area 40 into a gray scale image when the image of the face area 40 is a color image, and the gray scale face area 4
Each pixel of 0 is binarized with a predetermined luminance threshold (for example, half the maximum luminance) (step S3). Thereby, for example, as shown in FIG. 8A, a binary image including a relatively large number of line segments in the face region 40 is obtained.
Then, the binarization / normalization unit 13 applies the above formulas (3) and (3) to the face area 40 after binarization.
4) is used to calculate the projection entropies Ex and Ey (step S4), and it is determined whether these projection entropies Ex and Ey are greater than or equal to the projection entropy threshold Eth set for each (step S5).

When the projection entropy Ex, Ey is smaller than the projection entropy threshold Eth (step S5)
: NO), since the face area 40 has not yet been simplified to such an extent that the local direction contribution feature quantity can be calculated, the binarization / normalization unit 13 performs blackening of the face area 40 by binarization processing. In order to reduce the number of pixels (number of line segments), a predetermined value is added to the current luminance threshold value to increase the luminance threshold value (step S6).
.

Here, when a face is photographed in a dark illumination environment and the shadow in the face area 40 is strong and many wrinkles and spots are captured, the luminance threshold value is sufficiently increased and binarization processing is performed. Even if the projection entropy Ex and Ey of the face area 40 are difficult to exceed the projection entropy threshold Eth, and the binarization process is performed with a luminance threshold that the projection entropy Ex and Ey exceed the projection entropy threshold Eth. There is a possibility that the line segment of the facial part is removed more than necessary from the face area 40 and the recognition cannot be performed correctly.

Therefore, the binarization / normalization unit 13 raises the luminance threshold value in the above step S6, and then determines whether or not the luminance threshold value exceeds the limit threshold value that simplifies the face area 40 to the extent that it cannot be recognized. Judgment is made (step S7). When the luminance threshold is smaller than the limit threshold (step S7:
NO), since the face area 40 is not oversimplified even if binarization is performed by this luminance threshold, the binarization / normalization unit 13 returns the processing procedure to step S3, and again the binarization Process. As a result, for example, as shown in FIG. 8B, the image of the face region 40 is simplified compared to FIG. 8A.

On the other hand, if the luminance threshold is equal to or greater than the limit threshold (step S7: YES), binarization cannot be performed using this luminance threshold.
Therefore, the binarization / normalization unit 13 performs filtering processing for smoothing the line segment on the entire face region 40 to simplify the image and obtain an image in which the local direction contribution feature quantity can be calculated ( Step S8). For this filtering process, for example, a Gaussian filter having a standard deviation with a diameter of about the mesh region 25 can be used.
In the present embodiment, when the luminance threshold reaches or exceeds the limit threshold, the filtering process for smoothing the line segment of the face area 40 is performed. However, before the binarization process is performed in step S3, the filtering process is performed. May be performed in advance. In this case, it is desirable to reduce the smoothing effect by reducing the standard deviation of the Gaussian filter so that the necessary information included in the face area 40 is not lost.

Now, when the projection entropy Ex, Ey is greater than or equal to the projection entropy threshold Eth (step S5: YES), or after executing the filter processing in step S8, the face area 40 is calculated to the extent that the local direction contribution feature quantity can be calculated. This shows that the image (line segment) has been simplified. For example, as shown in FIG. 8C, only face parts such as eyebrows, eyes, nose and mouth remain in the face area 40, and these face parts are shown with relatively few line segments. It becomes a state.
After the face region 40 that can calculate the local direction contribution degree feature quantity is obtained in this way, the binarization / normalization unit 13 determines that points that are not included in the face part, that is, the number of connection points is predetermined. After removing noise from the face area 40 by removing a few or less isolated points (step S9), normalization is performed to match the face area 40 to the m × m mesh areas 25 for obtaining the local direction contribution feature amount. I do.

That is, as shown in FIG. 6, the binarization / normalization unit 13 measures the center of gravity OG and the size (number of vertical and horizontal pixels) of the face region 40 (step S10), as shown in FIG. 9B. Further, the center of gravity OG of the face region 40 is aligned with the center OM of the m × m mesh regions 25, and the major axis of the face region 40 (the vertical direction in the example shown in FIG. 9A) is m × m meshes. The face area 40 is enlarged or reduced so as to fit in the area 25 (step S11), and thereby normalization is performed.
For enlargement or reduction of the face area 40, affine transformation is used in which the aspect ratio of the face area 40 is kept constant. Further, when the face included in the face area 40 is tilted or faces in a direction other than the front, the face is not tilted and the front is turned before the enlargement or reduction of the face area 40. It also corrects the image. The face orientation can be detected based on the position and balance of facial parts such as eyes, nose and mouth when a face area is extracted.

Note that the size of the mesh region 25 is set as appropriate according to the processing capability of the device that executes the image recognition program 18 and the target recognition accuracy.
The size is set to about 100 pixels or more. Increasing the number of pixels included in one mesh region 25 can increase the accuracy of the calculation amount. The mesh region 25 is preferably square but may be rectangular.
Further, the removal of isolated points in step S9 may be performed after normalization in step S11.

As described above, when the face area 40 is binarized and normalized, the direction contribution feature amount calculation unit 14
A local direction contribution feature quantity is calculated for each mesh region 25. That is, as shown in FIG. 6, the direction contribution feature quantity calculation unit 14 initializes the mesh number M (1 ≦ M ≦ m × m) specifying the mesh region 25 to “1” (step S12). , Mth mesh region 25
The local direction contribution feature quantity is calculated for (step S13). Next, the direction contribution feature quantity calculation unit 14 determines whether or not the current mesh number M is equal to m × m, that is, whether or not the last mesh region 25 has been reached (step S14). M is m × m
(Step S14: NO), the mesh number M is incremented by “1” (step S15), and the processing procedure is stepped to calculate the local direction contribution feature quantity for the next mesh region 25. Return to S13.
On the other hand, when the mesh number M is equal to m × m (step S14: YES), these local direction contributions are shown to indicate that the local direction contribution feature quantity has been calculated for all the mesh regions 25. Based on the feature amount and the dictionary data 17 for face recognition, the face recognition unit 15
Who is the face reflected in the face area 40, or the face area 40 is the face recognition dictionary data 17.
Face recognition is performed (step S16).
).

As described above, according to the present embodiment, the binarization / normalization unit 13 binarizes the face area 40 by changing the luminance threshold based on the projection entropies Ex and Ey of the face area 40, and locally converts the face area 40. Since the direction contribution feature amount is simplified to a degree that can be calculated, an algorithm using the direction contribution feature can be used for face recognition of the face region 40 including many complex line segments.
A person's face can be recognized at high speed and with high accuracy by simple processing.
Thereby, for example, highly accurate face recognition can be realized even in a small device having a relatively low processing capability such as a digital camera, a PDA, or a mobile phone.

In addition, according to the present embodiment, the binarization / normalization unit 13 is configured to perform a filter process for smoothing a line segment before or after binarization of the face region 40. Even in the face area 40 in which the photographing environment is bad and cannot be simplified simply by binarization, it is possible to stably calculate the local direction contribution feature.

The above-described embodiment is merely an aspect of the present invention, and can be arbitrarily modified and applied within the scope of the present invention.
For example, in the above-described embodiment, each of the projection entropies Ex and Ey, which are the entropies of projection in the vertical and horizontal directions of the face area 40, is compared with the respective threshold entropy Eth, thereby simplifying the face area 40. Although it has been determined whether or not it has been completed, a simple determination may be made using only one of the projection entropies Ex and Ey.
For example, in the above-described embodiment, the case where the computer 3 as the image recognition apparatus recognizes a face as an object has been illustrated. However, the present invention is not limited to this. The present invention is applicable.
In addition, for example, the image recognition apparatus according to the present invention can be implemented in any electronic device such as a camera, a scanner, a projector, a television, and a printer.

It is a figure which shows the structure of the image recognition system which concerns on embodiment of this invention. It is a block diagram which shows the functional structure of a computer. It is a figure for demonstrating a direction contribution. It is a figure for demonstrating simplification of a face area | region. It is a flowchart of an image recognition process. It is a flowchart of an image recognition process. It is a figure for demonstrating area separation. It is a figure which shows an example of the simplification of the face area by binarization. It is a figure which shows the normalization of a face area | region.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Image recognition system, 2 ... Camera, 3 ... Computer (image recognition apparatus), 10 ... Image input part, 11 ... Face area extraction part, 12 ... Memory | storage part, 13 ... Binarization and normalization part, 14 ... Direction Contribution feature amount calculation unit, 15 ... face recognition unit, 17 ... dictionary data for face recognition, 18 ... image recognition program, 25 ... mesh region, 40 ... face region, dx ... X-axis projection, dy ... Y-axis projection, Eth ... Projective entropy threshold, Ex, Ey ... Projective entropy.

Claims (8)

An image input unit for inputting object image data including the object;
A binarization unit that binarizes each pixel constituting the object image data input by the image input unit using a predetermined luminance threshold;
A feature amount calculation unit that calculates a local direction contribution feature amount that is a direction contribution feature amount of a contour line included in a predetermined region of the object image data binarized by the binarization unit;
An image recognition apparatus comprising: a recognition unit that recognizes an object included in the object image data based on the local direction contribution feature amount calculated by the feature amount calculation unit. The image recognition apparatus according to claim 1,
The binarization unit includes:
The image recognition apparatus characterized by changing the said brightness | luminance threshold value according to distribution of the projection formed by integrating | accumulating an effective pixel for every column or / and a row about each pixel which comprises the said binarized said object image data. The image recognition apparatus according to claim 2,
The binarization unit includes:
An image recognition apparatus, wherein entropy for quantifying the state of the projection distribution is calculated, and the luminance threshold value is changed in accordance with the calculated entropy. In the image recognition device according to any one of claims 1 to 3,
The binarization unit performs a filter process for smoothing a line segment before or after binarization of the object image data,
The feature amount calculation unit calculates the local direction contribution feature amount for object image data that has been binarized and filtered by the binarization unit. The image recognition device according to claim 1,
The object is a face;
A storage unit for storing face recognition dictionary data including the local direction contribution feature amount for each of a plurality of different faces;
The recognition unit includes a local direction contribution feature amount calculated by the feature amount calculation unit for a face included in the object image data, and a local direction included in face recognition dictionary data stored in the storage unit. An image recognition apparatus characterized by recognizing a face included in the object image data based on a contribution feature amount.   An electronic device comprising the image recognition device according to claim 1. An image input process for inputting object image data including the object;
A binarization step of binarizing each pixel constituting the object image data input by the image input step using a predetermined luminance threshold;
A feature amount calculating step of calculating a local direction contribution feature amount that is a direction contribution feature amount of a contour line included in a predetermined region of the object image data binarized by the binarization step;
A recognition step of recognizing an object included in the object image data based on the local direction contribution feature amount calculated by the feature amount calculation step. Computer
An image input means for inputting object image data including the object;
Binarization means for binarizing each pixel constituting the object image data input by the image input means using a predetermined luminance threshold;
Feature amount calculating means for calculating a local direction contribution feature amount that is a direction contribution feature amount of a contour line included in a predetermined region of the object image data binarized by the binarization means;
An image recognition program that functions as a recognition unit that recognizes an object included in the object image data based on the local direction contribution feature amount calculated by the feature amount calculation unit.

Images