JPH07262373A - Color image recognition method - Google Patents

Abstract

PURPOSE:To recognize an object at high speed by converting and compressing a data amount with table conversion while securing information quantity required for image recognition in the case of executing the recognition processing of a color image. CONSTITUTION:An inputted color image signal RGB 1 is divided into small areas (meshes) by a mesh dividing part 2, a chromaticity signal is extracted and compressed for each small area by a feature amount extracting part 3 while referring to a conversion and compression table 4, and a chromaticity histogram is prepared and stored in a feature amount memory 5. A code book 6 stores the chromaticity histogram of an identifying object. Then, a vector quantizing part 7 calculates a distance between the chromaticity histogram in the feature amount memory and the code book, and the code of the code book, for which the distance is minimum, is held in a vector quantized value memory 8. A recognition part 9 calculates a distance between the vector quantized value histogram of an input image and a vector quantized value histogram concerning the identifying object of a dictionary, and it is decided whether the input color image is the identifying object or not.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color image recognition method for recognizing a specific image in a color image.

[0002]

2. Description of the Related Art It is expected that products for processing color images, such as color copying machines, color printers, color scanners and color image communication equipment, will increase more and more in the future. Although color images have become easier to use than before due to advances in hardware, especially low cost and large memory capacity, and lower communication costs, the amount of color image data is enormous (for example, , A3 size is 96 Mbytes), so that the same processing as a binary image cannot be performed at present.

In particular, image recognition (specific image recognition, OCR
In a technique requiring complicated processing such as (), the processing amount becomes enormous, and image recognition in a color image is more difficult to realize.

[0004]

Conventionally, as a method of identifying a specific color image such as a bill, paying attention to the distribution in the color space peculiar to each pattern portion forming the image, Identify the distribution in the unique color space that appears,
There is a method of extracting an image portion having the same characteristic as the specified characteristic (see Japanese Patent Laid-Open No. 4-180348). However, with this method, images that have the same spread in the color space cannot be identified even if the color distributions inside are different, that is, if the spreads in the color space are the same, the spread Images with different color distributions may be erroneously detected as specific images.

An object of the present invention is to secure the amount of information necessary for image recognition when performing color image recognition processing, and to convert and compress the amount of data by table conversion.
Objects that are included in the input image (specific image) are efficiently compressed without performing arithmetic processing using a conversion formula.
An object of the present invention is to provide a color image recognition method for recognizing an image with high accuracy and high speed.

Another object of the present invention is to reduce the influence of a show-through of a color image (original) to be recognized and to perform vector quantization on only image information to be recognized, thereby improving the recognition rate. It is to provide a color image recognition method with improved processing speed.

Further, an object of the present invention is to provide a color image recognition method in which recognition accuracy and processing speed are improved by not assigning vector quantization values when the distance from the codebook is larger than a predetermined threshold value. Especially.

[0008]

In order to achieve the above object, according to the invention of claim 1, the input color image is divided into a plurality of regions, and the feature amount is calculated from the color images in the divided regions. To generate a histogram by compressing the extracted features, and vector quantizing by comparing the histogram with a codebook created in advance,
The color image is recognized by collating the vector quantized value of the color image with the dictionary of the identification object.

The invention according to claim 2 is characterized in that the comparison process with the codebook is changed according to the information of the histogram.

According to a third aspect of the invention, a predetermined vector quantization value is given when the frequency of the histogram exceeds a predetermined value.

According to a fourth aspect of the invention, when the frequency of the histogram exceeds a predetermined value, the comparison process with the codebook is not performed.

According to a fifth aspect of the present invention, a predetermined vector quantization value is given when the distribution width of the histogram exceeds a predetermined value.

According to a sixth aspect of the present invention, when the distribution width of the histogram exceeds a predetermined value, the comparison process with the codebook is not performed.

The invention according to claim 7 is characterized in that a predetermined vector quantization value is given when the comparison result between the histogram and the codebook is larger than a predetermined distance.

According to an eighth aspect of the invention, the output image is deteriorated when the color image recognition process is performed and it is determined that the input color image includes a specific image.

[0016]

The input color image signal RGB is divided into small areas (mesh) by the mesh dividing section. In the feature quantity extraction unit, the input color image signal RG is input for each of the divided small areas.
The chromaticity signal of B is extracted, the extracted chromaticity signal is compressed by referring to the conversion compression table, and a chromaticity histogram is created and stored in the feature amount memory. The chromaticity histogram of the identification object is created and stored in the codebook in advance. Then, the vector quantization unit calculates the distance between the chromaticity histogram of the feature amount memory and the codebook, and the code of the codebook having the smallest distance is vector-quantized as the vector quantization value of the small area. Hold in value memory. A vector quantized value histogram is obtained and stored in advance in the dictionary for the identification object. The recognition unit calculates the distance between the vector quantization value histogram of the input image and the vector quantization value histogram of the dictionary, and determines whether or not the input color image is the identification target.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be specifically described below with reference to the drawings. FIG. 1 shows the configuration of an embodiment of the present invention.
In FIG. 1, the input color image signal (RGB) 1
A mesh division unit 2 that divides the color image obtained from the above into small regions (mesh), a feature amount extraction unit 3 that extracts a feature amount from the color image data in the small region, and a conversion that transforms and compresses the extracted feature amount A compression table 4, a feature amount memory 5 that holds the compressed feature amount, a vector quantization unit 7 that performs vector quantization by comparing the compressed feature amount with a codebook 6 that has been created in advance, A vector quantized value memory 8 for holding vector quantized values, a recognition unit 10 for performing a recognition process by collating the memory with a dictionary 9 of an identification target, and a control unit 11 for controlling the entire process. Has been done.

<Embodiment 1> An input color image signal is divided into a predetermined small area (mesh) by a mesh dividing section 2. For each of the divided small areas (mesh),
The feature amount of the input color image signal is extracted. FIG. 2 shows a diagram in which the original image is divided into small areas (mesh), and in this example, the small area has a size of 64 pixels × 64 pixels.

In this embodiment, the chromaticity signal of the input color image signal RGB is used as the characteristic amount. That is, the feature amount extraction unit 3 extracts the RGB chromaticity signals of the input color image signal for each small area (mesh), and the conversion compression table 4
The data compression processing is performed by referring to, the chromaticity histogram in the small area is created for each compressed RGB chromaticity signal, and the result is held in the feature amount memory 5.

FIG. 3 is a diagram for explaining an example of generating a histogram of the characteristic amount converted and compressed from the input color image signal. In the conversion compression table 4, for each input chromaticity signal, the corresponding conversion is performed. The output signal is held and, for example, the input value 255 is subjected to the conversion compression processing of the feature amount corresponding to the output value 15. The compressed feature amount histogram is generated on the feature amount memory 5. By converting and compressing the feature quantity, it is possible to suppress an increase in the memory capacity and speed up the matching process.

As shown in the histogram of FIG. 3, the chromaticity histogram 31 in the small area stored in the feature amount memory 5
The dimension (Hi) of H = (i) of i = 0 to 15 represents the frequency of the compression-converted R signal, that is, R ′, and i =
H (i) in the dimension of 16 to 31 is the compressed G
It represents the frequency of a signal, that is, G ′, and i = 32 to 4
H (i) in the 7th dimension represents the frequency of the compression-converted B signal, that is, B ′.

In the above embodiment, the total number of dimensions of the histogram of the feature quantity (48 in FIG. 3) is determined according to the degree of compression by the conversion compression table. Further, although the histogram is created from the RGB chromaticity signals of the input color image signal, the present invention is not limited to this, and the YMC signal or the Lab signal obtained by converting the input color image signal is used as a feature amount for conversion. It is also possible to perform compression processing and histogram generation processing.

FIG. 4 shows the chromaticity histogram in the small area,
FIG. 4 is a diagram showing an example of vector quantization by referring to a codebook, 41 shows a chromaticity histogram (Hi) in a small area stored in the feature amount memory 5, and 42 shows a codebook created in advance. The contents of 6 (C0, C1, C2 ...) are shown.

Here, the codebook 6 is typical in that a large number of identification objects or general images (original documents) are input, a large amount of chromaticity histogram data is created under similar conditions, and these are clustered. It is created in advance by obtaining a chromaticity histogram (codebook).

The vector quantizer 7 calculates the distance (DCj) by matching the chromaticity histogram (Hi) 41 and the content 42 of the codebook, and calculates the distance (generally as the square of the Euclidean distance). Code (6) of Codebook 6 with the smallest defined squared strain measure
j) is assigned as the vector quantized value (VQ value) of the small area.

FIG. 5 shows the vector quantization value (V
An example of (Q value) is shown. Each cell corresponds to one small area mentioned above, and the numerical value in each cell is a vector quantized value (VQ value).
Is. The vector quantized value memory 8 holds the vector quantized value of each small area, and a histogram is created for the vector quantized value (VQ value) of the input image. The recognition unit 10 performs a comparison process on the histogram of the vector quantized values by matching with the dictionary 9 of the recognition target.

FIG. 6 is a diagram for explaining matching between the histogram of the vector quantization value of the input image and the histogram of the vector quantization value of the dictionary at the time of image recognition.
Reference numeral 61 in the figure shows an example of a histogram of the created vector quantization value. Reference numeral 62 indicates the dictionary contents of the identification object, which is created in advance by being represented by a vector quantization value histogram as in the input image. In the case of the figure, for example, the identification objects A, B, C. ． The vector quantized value histogram of (the number of codebooks is 64) is stored.

The recognition unit 10 matches the vector quantized value histogram 61 of the input image with the vector quantized value histogram 62 of the dictionary 9 of the object to match the distance (DT).
k) is calculated, and the identification target (k) having the smallest distance (DTk) is determined to be the color image of the recognition target held in the dictionary 10.

The control unit 11 controls each stage in the overall image recognition processing such as management of image data to be processed and memory, distance calculation of matching processing, and the like.

<Embodiment 2> In Embodiment 1 described above, the histogram of the feature amount is created for all the pixels of the input image and the comparison processing with the codebook is performed. By changing the comparison processing with the codebook based on the histogram information of the feature amount generated for each area, the recognition processing for the background portion (background portion) and the noise image that are not necessary for image recognition can be controlled. It is possible to improve the recognition rate and the processing speed.

In general, the background portion, which is not necessary for image recognition, has a substantially constant density in the small area, so the histogram of the feature amount generated for each small area is concentrated in a specific portion. Therefore, the width of the frequency distribution is narrow, and the maximum value of the frequency tends to increase. Therefore, when the maximum value of the histogram frequency exceeds a preset threshold value, it is determined that the area is a solid area of an image with uniform density, and 0 is assigned to the vector quantization value. Vector quantization is applied only to As another embodiment, the comparison processing with the codebook may not be performed when the frequency of the histogram exceeds a preset threshold value.

In the case of input image data obtained by reading a document image with a scanner or the like, a phenomenon may occur in which the image on the back side of the sheet causes show-through in the highlight portion of the image. Since the feature amount of the image data in the highlight part other than the image noise due to the show-through is often uniform, the histogram of the feature amount has a large peak concentrated in a specific part, and the maximum frequency becomes large and the frequency increases. The width of the distribution tends to narrow. Therefore, if it is determined that the histogram frequency distribution information has a preset histogram characteristic, the feature quantity histogram of the small area is removed from the preset histogram characteristic and compared with the codebook to obtain the vector quantization value. Is assigned to suppress the influence of image noise, and vector quantization is applied only to the image information to be recognized.

In this embodiment, the maximum value of the frequency and the distribution width of the frequency are adopted as the histogram information of the characteristic amount, but the present invention is not limited to this, and the characteristic of the characteristic amount histogram of the specific image to be recognized is analyzed. It is sufficient if the histogram information is set as follows.

<Third Embodiment> The codebook in the first embodiment is created by inputting a large number of recognition target images, creating a large amount of chromaticity histogram data under the same conditions, and clustering these. In addition, when the input image is not the image to be recognized, it may be far from any codebook. In the third embodiment, when the recognition target image is compared with the codebook and the vector quantization value is assigned, if the compared distance is larger than a predetermined threshold value, the vector quantization value is not assigned.
As a result, when there is no identification candidate as a result of matching, it is possible to quickly determine that the image to be recognized does not exist in the input color image.

<Embodiment 4> When matching with the dictionary in Embodiment 1, a threshold value for the effective distance is set, and the calculated distance is compared with the threshold value. The object to be identified is identified as an identification candidate, but if it is larger than the threshold value, it is not identified. As a result, if there is no identification candidate as a result of matching, it is possible to determine that the image to be recognized does not exist in the input color image.

<Fifth Embodiment> In the fourth embodiment, the above-mentioned threshold value is set for each identification target object, and the calculated distance is compared with the threshold value for each recognition target image. If the distance is less than the threshold value, the recognition is performed. The target image is set as an identification candidate, but if it is larger than the threshold value, the recognition target image is not set as an identification candidate. As a result, when identifying a plurality of recognition targets, matching processing that makes use of the characteristics of the target objects becomes possible. More specifically, a certain object k is a document j other than the object k.
When it is easy to make a mistake, it is possible to distinguish the object k and the document j with high accuracy by lowering the threshold value of the object k, and it is possible to prevent erroneous recognition.

<Embodiment 6> In the image recognition processing of each of the above-described embodiments, as a specific image to be recognized, a characteristic amount such as banknotes or securities is held in a codebook or a dictionary, so that the specific image is input to the input image. It is possible to accurately determine whether or not is included. When it is determined that the input image includes the specific image, the control circuit 11 controls the color image data determined to be the specific image to perform a process of degrading the output image. This embodiment is a color copying machine,
It is suitable when applied to an image processing apparatus such as a color facsimile. As another embodiment, the image data output process may be prohibited.

[0038]

As described above, according to the invention described in claim 1, since the feature quantity extracted from the input image data is compressed and then vector-quantized, the amount of information necessary for color image recognition is reduced. It is possible to reduce the amount of processing data while holding it, improve the processing speed, and perform the image recognition processing with high accuracy.

Further, according to the invention described in claims 2 to 7, when the matching is performed, the matching processing is changed by the threshold processing for comparing the obtained histogram information with the set threshold. Therefore, the processing time required for image recognition can be shortened and the color image recognition process can be efficiently performed.

According to the eighth aspect of the invention, there is a great effect in preventing forgery when the input image includes a specific image such as a bill or securities.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of an exemplary embodiment of the present invention.

FIG. 2 is a diagram showing an example in which an original image is divided into small areas.

FIG. 3 is a diagram showing an example in which a feature amount in a small area is converted and compressed to generate a histogram.

FIG. 4 is a diagram illustrating matching between a feature amount histogram of an input image and a codebook.

FIG. 5 is a diagram showing an example of vector quantized values of an input image.

FIG. 6 is a diagram illustrating matching between a vector quantization value histogram of an input image and a vector quantization value histogram of a dictionary at the time of recognition.

[Explanation of symbols]

 1 Color Image Signal 2 Mesh Divider 3 Feature Extractor 4 Conversion Compression Table 5 Feature Memory 6 Codebook 7 Vector Quantizer 8 Vector Quantized Value Memory 9 Recognition Unit 10 Dictionary 11 Control Unit

Claims (8)

[Claims] 1. An input color image is divided into a plurality of regions, a feature amount is extracted from the color image in the divided regions, and the extracted feature amount is compressed to generate a histogram. Vector quantization is performed by comparing a histogram with a codebook created in advance, and the color image is recognized by matching the vector quantized value of the color image with a dictionary of an identification target. And color image recognition method. 2. The color image recognition method according to claim 1, wherein the comparison process with the codebook is changed according to the information of the histogram. 3. The color image recognition method according to claim 2, wherein a predetermined vector quantization value is given when the frequency of the histogram exceeds a predetermined value. 4. The color image recognition method according to claim 2, wherein when the frequency of the histogram exceeds a predetermined value, the comparison process with the codebook is not performed. 5. The color image recognition method according to claim 2, wherein when the distribution width of the histogram exceeds a predetermined value, a predetermined vector quantization value is given. 6. The color image recognition method according to claim 2, wherein when the distribution width of the histogram exceeds a predetermined value, the comparison process with the codebook is not performed. 7. The color image recognition method according to claim 1, wherein a predetermined vector quantization value is given when a comparison result between the histogram and the codebook is larger than a predetermined distance. 8. When the color image recognition processing is performed and it is determined that the input color image includes a specific image,
The output image is deteriorated, wherein the output image is deteriorated.
The described color image recognition method.