JPH05174149A - Picture recognition device - Google Patents

Abstract

PURPOSE:To enable the recognition after an object picture is easily found and the position and size are obtained by making a recognition object picture and an unknown picture in a mosaic state and searching/recognizing them. CONSTITUTION:A recognition object picture input part 1 fetches an object to be a recognition object, for instance, a typical human face picture if the object is human by using a TV camera or a scanner, etc., and inputs it in a rough search part 100. The object picture to be a recognition object is made in a rough mosaic state and rough position and size are found by scanning an unknown picture by using this picture as the feature of the object picture at the time of a search. Next, the object picture to be the recognition object is made in a fine mosaic state and accurate position and size are found by scanning around the already obtained rough position by using this picture. Further, recognition object picture is made in a fine mosaic and the object picture is recognized by performing the matching of the mosaic of the area obtaining the position and size in the already obtained unknown picture and the data which the previous recognition object picture is made in a finer mosaic.

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 accurately searching for a target image from input natural images and specifying a target.

[0002]

2. Description of the Related Art A typical natural image is a grayscale or color image. In order to detect and recognize a target image from the natural image, conventionally, attention has been paid to the shape of the image, and the shape information is used as a feature. It was to find out and recognize candidates with matching features from the input image. However, it is very difficult to extract the correct shape from the natural image. Conventionally, in order to separate the target object and the background, it is necessary to pre-learn the background or limit the background to a uniform background. However, various restrictions have been imposed, such as limiting the target to an artificial rigid object in order to facilitate the extraction of line segments.

[0003]

Even the detection of the presence or absence of the target image is accompanied by such difficulty, and it is almost impossible to grasp the exact position and size of the object in the natural image.

That is, it is extremely difficult to accurately extract an object from an arbitrary natural image, and therefore, there are many practical restrictions and difficulty in usefulness. An object of the present invention is to relax the above-mentioned various constraints, accurately find a target image on the screen, and recognize an object based on the result.

[0005]

SUMMARY OF THE INVENTION In the present invention, the characteristics of the object are not the shape but the grayscale or color information of coarse resolution. That is, it is based on the fact that even if the image has a significantly reduced resolution so that Lincoln can be identified from the Lincoln mosaic image, the target feature can be expressed.

Specifically, the object image to be recognized is roughly mosaiced, and this is used as a feature of the target image at the time of searching to scan the unknown image to find its approximate position and size. Next, the object image to be recognized is finely mosaiced, and using this, the vicinity of the roughly obtained position is scanned to find the accurate position and size. further,
The recognition target object image is finely mosaiced, and the target image is recognized by matching the mosaic of the region in which the position and the size have already been obtained in the unknown image with the data obtained by finely mosaicing the above recognition target object image.

[0007]

In this way, by mosaicing the image in several stages, it is possible to easily find the object from any screen and further recognize it.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing an embodiment for performing image recognition according to the present invention, and FIG. 2 is a configuration of a rough search unit shown in FIG.
FIG. 3 shows the configuration of the detail search unit shown in FIG. 1, and FIG. 4 shows the configuration of the recognition unit shown in FIG. FIG. 5 and FIG. 6 together represent one figure and are diagrams for explaining a process of mosaicing a search target image or an unknown image and searching / recognizing a target target from the unknown image.

In FIG. 1, reference numeral 1 denotes a recognition target image input unit, which uses a TV camera, a scanner or the like to capture an object to be recognized, for example, a face image of a typical person in the case of a person, and to a rough search unit 100. Specifically, it is input to the rough search target image buffer 101 in FIG. Therefore,
The content of the buffer 101 is a set of pixels represented by shading or color.

This image to be recognized is represented by a matrix F = [f]. As shown in FIGS. 5 and 6, this image is divided into M × N blocks each having a size of Wa pixels × Wa pixels, and the mosaic image FC is divided into blocks. Each block is represented by a representative value in the block, for example, an average value in the block if it is a shade, and a color having the maximum frequency in the block if it is a color.

The coarse search target image mosaic unit 102 mosaics the image in the buffer 101 and stores it in the coarse search target image dictionary 103. If necessary, a plurality of search targets, in this example, an average value of a plurality of face images may be taken as representative face mosaic data, or a plurality of rough search target images may be prepared.

On the other hand, the unknown image U is taken into the unknown image buffer 104 shown in FIG. 2 in the rough search unit 100 via the unknown image input unit 2 of FIG. Therefore, for the unknown image U, as shown in FIG.
In 5, a block size of Wa pixels × Wa pixels is divided into P × Q, a representative value is calculated for each block, and a mosaic image UC is calculated.
To get The unknown image input unit 2 may be the same as the recognition target image input unit 1 described above.

Next, as shown in FIG. 5, the mosaic of the unknown image is scanned to search for a location that matches the coarse-search target image mosaic. That is, it is based on the following principle. When the arbitrary M × N mosaic data among the P × Q mosaic data of the unknown image is set to Ui (indicated by a shaded area), Ui
The position of Ui when the distance Di between the target mosaic data FC and the rough search target mosaic data FC is the minimum value is the rough search result. As the distance Di, there are various ways to output, but a typical one is the Euclidean distance.

Specifically, it is as follows. In FIG. 2, reference numeral 106 denotes a candidate selection unit that selects candidate data Ui (i = 1, 2 ...) That falls within the frame of M × N blocks from the unknown image mosaic unit 105. Generally, the upper left M × N block of the unknown image is generally selected as U1. The mosaic data U1 and the rough search target mosaic data FC from the rough search target image dictionary 103 are input to the distance calculation unit 107, and the distance is calculated. The result is stored in the position / size / distance storage unit 108 together with the M × N position data in the unknown image mosaic data UC at this time, that is, the U1 position data and the mosaic block size Wa.

Further, the candidate selection unit 106 selects the next candidate U2 from the unknown image mosaic unit 105, and the distance calculation unit 107 calculates the distance at this time, and the position and block size of the candidate U2 as well as the position / size.・ Distance storage unit 108
Store in. Similarly, M × N candidate Ui are sequentially selected for all regions of the unknown image mosaic, and the selected position, block size, and calculated distance are all position / size /
It is stored in the distance storage unit 108.

By the way, in the above, in the mosaicing of the unknown image, the same block size as the search target image, Wa pixels × Wa pixels, is used. If the search target image and the search target image in the unknown image have the same size, this size may be left unchanged, but generally, the size of the search target image in the unknown image is also unknown. Therefore, the unknown image mosaic unit 10
5 indicates the number of pixels of the block size, for example, 1 / Wa
Change to Wa 'such as 5 to 10/5. Therefore, the above procedure is repeated for each block size Wa ′, and the obtained distance, the position at that time, and the block size Wa ′ are stored in the position / size / distance storage unit 108.

Also, when a plurality of search target images are stored in the dictionary, similarly, the above procedure is repeated for each search target image, and the obtained distance and the position and block size at that time are determined.・ Size / distance storage unit 108
Store in.

In this way, when the distances in all cases are stored in the position / size / distance storage unit 108, the minimum value of the minimum distance detection unit 109 and the distance data in the position / size / distance storage unit 108 is calculated. Detect what you take. This value is sent to the recognition target presence / absence determination unit 110, and when this value is less than or equal to a threshold value, it is determined that the target image exists in the unknown image, and the position and block size thereof are extracted from the position / size / distance accumulation unit 108. And outputs the result to the rough search result output unit 111. The rough search result output unit 111 calculates the size of the search target image in the unknown image from the block size Wa by [M × W
It is calculated from a] × [N × Wa].

On the other hand, when the minimum distance value exceeds a certain threshold value, it is determined that the target image does not exist in the unknown image, and "no" is output to the rough search result output unit 111. The coarse search result output unit 111 immediately notifies the detailed search unit 200 of these results.

In this way, when the rough position and size of the recognition target are found by the rough search, the block of the mosaic is made small and the position is searched in detail. In the detailed search, the area of the target image may be narrowed down to a part as compared with the coarse search, but here, the same area is used.

First, as shown in FIG. 6, the recognition target is set to W in advance.
A block having a size of b pixels × Wb pixels (Wb <Wa) is divided into M ′ × N ′, and each block is represented by a representative value in the block.

Specifically, the detailed search target image buffer 2
Reference numeral 01 captures recognition target image data from the recognition target image input unit 1 and stores the recognition target image data in the detailed search target image mosaic unit 20.
Send to 2. The detailed search target image mosaicking unit 202 mosaics (FD) this image and stores it in the detailed search target image dictionary 203.

If necessary, the average value of a plurality of detailed search target images may be taken or a plurality of detailed search target images may be prepared, as in the case of the rough search. On the other hand, when the above-described rough search result is notified to the rough search result notification unit 220, the unknown image buffer 204 knows this, and the unknown image input unit 2 of FIG. Capture the relevant part. Therefore, the unknown image mosaic unit 2
On the other hand, as shown in FIG. 6, 05 is divided into P ′ × Q ′ with a block size of Wb pixels × Wb pixels, a representative value is calculated for each block, and a mosaic UD is obtained.

Therefore, as in the case of the rough search, as shown in FIG. 6, the mosaic of the corresponding partial image of the unknown image is scanned to search for a location that matches the detailed search target mosaic FD. That is, when an arbitrary M ′ × N ′ mosaic of the P ′ × Q ′ mosaic UD of the partial image is Uj (indicated by a shaded area), the distance Dj between Uj and the search target mosaic FD The position of Uj when is minimized is the detailed search result.

Specifically, in FIG. 3, the candidate selection unit 2
06 represents M ′ × out of the unknown image mosaicking unit 205.
A candidate Uj that falls within the N ′ frame is taken out. Distance calculator 20
7 calculates the distance Dj of this Uj and the detailed search target mosaic FD from the detailed search target image dictionary 203, and stores the position data of Uj at this time and the block size Wb of the mosaic in the position / size / distance accumulating unit 208. send.

Here, if necessary, the block size Wb
Is changed (Wb ′) or a plurality of detailed search target images are used, as in the case of the rough search. Thus, when the minimum distance detection unit 209 detects the minimum value, the position data of Uj and the block size Wb at this time are sent to the detailed search result output unit 210. The detailed search result output unit 210 calculates an accurate size of the search target image from the block size Wb, that is, [M ′ × Wb] × [N ′ × Wb], and sends it to the recognition unit 300 together with the position data.

In this way, when the precise position and size of the recognition target are found by the detailed search, the mosaic block is reduced to recognize what the target is. First, as shown in FIG. 6, the recognition target is divided into M ″ × N ″ in advance by a block having a size of Wc pixels × Wc pixels (Wc <Wb). Each block of this mosaic is represented by a representative value in the block as in the above case. If the features to be recognized are concentrated in the partial area, that portion may be divided into M ″ × N ″. That is, in the process of searching, objects need universal features, but in the process of recognition, features that distinguish them are required, and finer mosaics and partial regions are effective.

Here, the case of using this partial area will be described. Specifically, the recognition target image buffer 301 takes in the data of the partial area of the recognition target image from the recognition target image input unit 1 and stores it in the recognition target image mosaicking unit 302.
Send to. The recognition target image mosaicking unit 302 mosaics (FDD) this image and stores it in the recognition target image dictionary 303.

Note that, usually, many recognition target images are mosaiced and stored in the recognition target image dictionary 303. On the other hand, when the detailed search result is notified from the detailed search unit 200 to the detailed search result notification unit 320, the unknown image buffer 304 knows this and the unknown image input unit 2 in FIG.
An image having an accurate position and size of the recognition target image in the unknown image U is captured. Image clipping / mosaicing unit 305
Is a block size of Wc pixels × Wc pixels as shown in FIG.
And the representative value is calculated for each block to obtain mosaic data UDD.

Therefore, the distance between the mosaic data UDD and the recognition target image mosaic data FDD is calculated, and if this is less than a certain threshold value, it is determined that the unknown image is the recognition target image. Alternatively, if there are a plurality of recognition target images, F
The object with the smallest distance from DD is determined as the recognition target. Here, a general case of the latter case will be described as an example.

Specifically, in FIG. 4, the distance calculation unit 3
Reference numeral 06 denotes the unknown image mosaic data UDD output from the image clipping / mosaicing unit 305 and the recognition target image dictionary 3
The distance of the recognition target image mosaic data FDD from 03 is calculated, and the distance at this time and the code of the recognition target image are sent to the distance storage unit 307. Similarly, the distance calculation unit 306 sequentially extracts the recognition target image mosaic data from the recognition target image dictionary 303, calculates the distance to the unknown image mosaic data UDD, and accumulates this value and the code of the recognition target image. Send to section 307.

Next, when the minimum distance detection unit 308 detects the minimum value in the distance storage unit 307, the minimum distance detection unit 308 sends the code of the recognition target image at this time to the recognition result output unit 309. The result is further sent to the output unit 3.

On the other hand, as a result of the rough search, the target image is "none".
In the case of, this is notified to the output unit 3. In this way, the approximate position of the recognition target image F is first found in the unknown image U, and the accurate position and size are found by using this as a clue, and the target is further specified.

[0034]

According to the present invention, by performing mosaic search and recognition on the recognition target image and the unknown image, no special condition such as uniformity with respect to the background is given, and when line segments are used. It is possible to easily find the target image, obtain its position and size, and recognize the target image without causing an error due to noise. In particular, it is effective for objects having similar shapes and colors, such as human faces and butterflies.

[Brief description of drawings]

FIG. 1 is a block diagram illustrating the entire invention.

2 is a diagram showing the contents of a coarse search unit 100. FIG.

FIG. 3 is a diagram showing the contents of a detail search unit 200.

FIG. 4 is a diagram showing the contents of a recognition unit 300.

FIG. 5 is a diagram illustrating a process in which a recognition target image and an unknown image are mosaiced, and a recognition target is searched and recognized in stages from the unknown image.

FIG. 6 is a diagram illustrating a process of forming a single figure together with FIG. 5, mosaicizing a recognition target image and an unknown image, and gradually searching for and recognizing the recognition target from the unknown image. ..

[Explanation of symbols]

 1 Recognition Target Image Input Section 2 Unknown Image Input Section 3 Output Section 100 Coarse Search Section 200 Detailed Search Section 300 Recognition Section

Claims (8)

[Claims] 1. A first means for dividing a natural image to be recognized into M × N in advance with a large block size, calculating a representative value of the shade or color of each block, and storing the representative value as dictionary data, , A second means for previously dividing the natural image to be recognized into M ′ × N ′ with a small block size, calculating a representative value of the shade or color of each block, and storing it as dictionary data. A third means for dividing the natural image to be recognized into M ″ × N ″ with a smaller block size, calculating a representative value of the grayscale or color of each block and storing the dictionary data, and an unknown natural image When given as an input, the input image is divided into P × Q with a large block size to calculate a representative value of each block, and an arbitrary M × N area is taken out from this to obtain a first means. The fourth means for calculating and accumulating the distance formed by the representative value group obtained in step (4) and the fourth means over the entire area of P × Q are repeatedly applied to obtain the minimum value from the obtained distances. When this minimum value is smaller than the threshold value given in advance, it is determined that there is an image corresponding to the recognition target image at the position where the minimum value is derived in the input image. On the other hand, when the minimum value is larger than the threshold value, The fifth means for determining that there is no image corresponding to the recognition target image, and when the fifth means determines that the recognition target is present at a certain position, M × in the vicinity of the corresponding position in the input image The N area is divided into P ′ × Q ′ by small blocks, the representative value of each block is calculated, and an arbitrary M ′ × N ′ area is taken out from this and a representative value group obtained by the second means is obtained. A sixth means for calculating the distance formed by and accumulating it, and for the entire area of P'x The sixth means is repeatedly applied, the minimum value is obtained from the obtained distances, and the position and size of the M ′ × N ′ image at this time are set as the position and size of the recognition target image. And a region of the corresponding position and size in the input image obtained by the seventh means is divided into M ″ × N ″ with a smaller block size to calculate the representative value of each block, and the third means An image recognition apparatus comprising: an eighth means for calculating a distance formed by the representative value group obtained in step (4), and if the distance is equal to or less than a certain threshold value, the image is recognized as a recognition target image. 2. In the first means, a target image is previously set as a plurality of similar image groups, which are not unique, and each image is divided into M × N to calculate a representative value of each block. The image recognition apparatus according to claim 1, wherein an average value is calculated for each corresponding block among a plurality of images, and the average value is stored as a dictionary. 3. In the second means, a target image is previously set as a plurality of not only unique but similar image groups, each image is divided into M ′ × N ′, and a representative value of each block is represented. The image recognition apparatus according to claim 1, wherein the average value is calculated for each corresponding block among a plurality of images, and the average value is stored as a dictionary. 4. In the third means, a target image is not a unique image but a plurality of image groups, and each image is divided into M ″ × N ″ to calculate a representative value of each block. The image recognition apparatus according to claim 1, wherein the image recognition apparatus stores the dictionary as a dictionary. 5. The second means, wherein a partial image of the recognition target image is M ′ × N ′ in advance with a small block size.
And the representative value of each block is calculated and stored as dictionary data. In the sixth means and the seventh means, the partial area of the M × N area obtained by the fifth means is P ′. 2. The distance to the relevant dictionary data is calculated for each of M ′ × N ′ among them divided into × Q ′.
The image recognition device described. 6. The third means, wherein the partial image of the recognition target image is M ″ with a smaller block size in advance.
× N ″, and the representative value of each block is calculated and stored as dictionary data.
M × N 'area with a smaller block size
The image recognition apparatus according to claim 1, wherein the image recognition apparatus is divided into 7. The fourth means is characterized in that the block size used for division is variable.
The image recognition device described. 8. The sixth means, wherein the block size used for division is variable.
The image recognition device described.