JP4336755B2 - A signboard recognition method using color images - Google Patents

Description

The present invention relates to a signboard recognition method using a color image for recognizing an object such as a road sign using a captured color image.

Conventionally, when recognizing an object such as a road sign using a color image, it is recognized based on lightness information of red, blue, or green, or by normalizing shifted hue information. .
As described above, for example, Patent Document 1 discloses a technique for recognizing a marker that is a recognition target based on lightness information and hue information of a color image.
However, when recognizing an object to be recognized by the conventional technique, it is necessary to set a use color in advance, and it has not been possible to target a signboard having a variety of use colors, shapes, and sizes. .
Also, when using color images acquired in a natural environment such as outdoors for analysis, the color information of the recognition target is not necessarily uniform, and even if the recognition target is the same, data such as weather and time can be acquired. Depending on conditions, it may happen that the color is recognized as different.
For this reason, when recognizing a recognition target object based only on hue information by a conventional technique, a boundary portion between regions using different colors becomes ambiguous.
Furthermore, when considering pedestrians who are road users, the positional relationship (distance / angle) between pedestrians and observation targets (signboards and signs) is undefined, unlike when observing from the position of the driver of the vehicle. The size and shape of the recognition object in the color image are different.
Accordingly, the recognition performance of the recognition object in the color image is deteriorated.
JP 2004-213127 A Japanese Patent Laid-Open No. 9-185703

Therefore, in order to solve the above-described problems, the present invention provides a signboard recognition method using a color image that improves the recognition performance of an object such as a road sign using a captured color image. Objective.

The sign recognition method using a color image according to the present invention divides the hue range of color information into three in the captured color image and has a hue value in the range of 0 to π / 6, 3π / 2 to 2π. An area having a saturation value equal to or greater than a predetermined threshold, a hue value in a range of π / 6 to 5π / 6, and a hue value in a range of 5π / 6 to 3π / 2. extracting signboard by the saturation and circularity values labeling area of 0.3 or more and less than 0.8 information in 3 divided regions chroma value is not smaller than a predetermined threshold value has a signboard with information circularity Recognize the signboard by detecting the edge using the Laplacian filter and extracting the character string area in the signboard, focusing on the fact that the contrast between the background color and the character color inside the signboard is high Is.

According to a method of recognizing signs using the color image of the present invention, it is possible to use color, shape and size to recognize signboard is indefinite, a wide variety of environmental information forward from the color image road Can be provided to users.
Furthermore, since it is possible to consider the effects of the data acquisition condition (imaging time, height, angle and shading information, etc.), without lowering the recognition performance of Nursing plate, it is possible to provide useful information to road users.
Moreover, since the recognition target object is recognized using the hue information, the recognition target object can be recognized from color images with different data acquisition conditions without degrading the recognition performance.
Furthermore, since the circularity and the ratio of pixels that occupy the divided area are used as indices as shape information, road users and recognition objects (signboards and signs) do not necessarily face each other, and are installed at high positions. The recognized recognition object can also be recognized.

Next, an apparatus for carrying out a signboard recognition method using a color image of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic configuration diagram of an apparatus for performing a recognition method according to an embodiment of the present invention.
FIG. 2 is a flowchart of the sign / signboard identification unit of FIG. 1, and FIGS. 3 and 4 are display image diagrams showing example results obtained by the present invention.

As shown in FIG. 1, an apparatus for performing a recognition method according to an embodiment of the present invention includes an imaging unit 11, a sign / signboard identification unit 12, a circular sign information extraction unit 13, a circular sign recognition unit 14, and a non-signboard area rejection. Part 15, character string region extraction part 16, and output part 17.
The imaging unit 11 is a device that incorporates an imaging device such as a CCD camera, for example, and images a landscape that develops in the vicinity of the front of an automobile or the like to which a sign recognition device is attached, and acquires a color image.
The sign / signboard identification unit 12 converts a color image into color information of hue, saturation, and brightness by HSI (Hue Saturation Intensity) conversion.
The circular marker information extraction unit 13 targets the inside of the extracted marker region and sets a pixel having a color whose hue value is close to blue and cyan and whose saturation value is equal to or greater than a predetermined threshold as a blue region. A pixel having a color close to red / magenta and having a saturation value equal to or greater than a predetermined threshold is set as a red region, and a region other than blue and red within the marker region is set as a white candidate region.
The circular marker recognizing unit 14 obtains the number of constituent colors and the ratio of the constituent colors in the region using the result obtained by the circular marker information extracting unit 13.
The non-signboard area rejection unit 15 obtains the interclass variance for the signboard candidate area obtained from the sign / signboard identification unit 12 by a discriminant analysis method.
The character string region extraction unit 16 performs binarization processing on the signboard candidate region, and then detects an edge using a four-direction Laplacian filter.
The output unit 17 displays the contents obtained by the circular sign recognition unit 14 and the character string region extraction unit 16 on a display.

The following processing is performed for the purpose of circular label extraction by an apparatus that performs a recognition method according to an embodiment of the present invention.
As shown in FIG. 2, the image pickup unit 11 picks up a landscape developed near the front of a car or the like and inputs a color image, and the sign / signboard identification unit 12 converts the acquired color image into an HSI (Hue Saturation Intensity). By the conversion, the color information of hue, saturation, and brightness is calculated.
In the sign / signboard identification unit 12, pixels whose hue value is close to blue / cyan and whose saturation value is a predetermined threshold (for example, 170) or more are set as blue candidate areas, and the hue value is close to red / magenta. Pixels having a predetermined threshold value (for example, 100) or more are set as red candidate regions.
Next, the binarization process and the contraction / expansion process are performed on the blue candidate area in the image, and the labeling process is performed after the expansion process.
Then, in order to complement the red candidate area and obtain a circular feature, the target pixel in the information (0, 1) having a large number of pixels for the target pixel and eight neighboring pixels (total nine pixels) in the binarized image of the red candidate area Replace the value of.
Finally, after the shrinking / expanding process, a labeling process is performed.
The circularity in each labeling area is calculated, and a labeling area having a circularity of 0.8 or more is set as a label area.
The circularity is a feature value for measuring the complexity of the shape from the area and the perimeter, and the circularity = 4π × the area of the labeling area / (the perimeter of the labeling area) ^2. The shape of the labeling area is a perfect circle. In this case, the value is 1.

Next, the circular marker information extraction unit 13 targets the inside of the extracted marker region, and applies pixels whose hue value is close to blue / cyan and whose saturation value is equal to or greater than a predetermined threshold (for example, 140) to the blue region. And
On the other hand, a pixel having a color whose hue value is close to red / magenta and whose saturation value is equal to or greater than a predetermined threshold (for example, 110) is defined as a red region.
In addition, a region other than the blue and red regions inside the marker region is set as a white candidate region, and binarization processing is performed on the white candidate region.
Next, since the color information of the boundary part of the painted area (white, blue, red) is ambiguous, the Laplacian filter in four directions is applied to the binarized image of the white candidate area to detect edge position information To do.
The detected edge portion is assumed to be a boundary area of two colors (white and blue, white and red, blue and red), and a white candidate area excluding the boundary area is set as a white area.
Further, each luminance value of RGB is automatically acquired from the pixels classified into each region (white region, blue region, red region), and teacher data is created.
By this processing, color information of each region that differs for each target image can be extracted, so that color classification of the boundary region is possible.
Using the created teacher data, each pixel of the boundary portion of the paint color is classified into white, blue, and red, and the RGB luminance values of the pixels classified into the white, blue, and red regions are (255, 255, 255), The values (0,0,255) and (255,0,0) are assigned, respectively.
Next, the circular marker recognizing unit 14 divides the feature region into a predetermined size and obtains the feature amount of each block.
Specifically, a feature quantity a (the ratio of the feature area occupying each block, and a state in which the block is composed of only the feature area is assumed to be 100) and a feature quantity b (a feature area included in one block) However, the ratio of the total area of the feature area, that is, if all the feature areas are included in one block, the value of the block is 100 and the value of the remaining blocks is 0).
Each of the two feature quantities is obtained, and the sign is recognized by comparing this with the prepared teacher data.

Next, the following processing is performed for the purpose of signboard extraction by an apparatus that performs the recognition method according to the embodiment of the present invention.
As shown in FIG. 2, the image pickup unit 11 picks up a landscape developed near the front of a car or the like and inputs a color image, and the sign / signboard identification unit 12 converts the acquired color image into an HSI (Hue Saturation Intensity). By the conversion, the color information of hue, saturation, and brightness is calculated.
The sign / signboard identification unit 12 extracts regions that fall within the hue range obtained by dividing the hue value into three from the color image.
That is, the hue value is in the range of 0 to π / 6, 3π / 2 to 2π, and the saturation value is a predetermined threshold (for example, 80) or more, and the hue value is in the range of π / 6 to 5π / 6. And has a hue value in a range of 5π / 6 to 3π / 2 and a saturation value equal to or greater than a predetermined threshold (eg, 110). Divide into three areas.
Next, the obtained image is subjected to binarization processing and contraction / expansion processing.
Further, a labeling process is performed, and a labeling area having a circularity value of 0.3 to less than 0.8 is set as a signboard candidate area.

The non-signboard area rejection unit 15 obtains the inter-class variance by the discriminant analysis method for the signboard candidate area obtained from the sign / signboard identification unit 12.
As a result, the signboard candidate area whose variance value is equal to or less than a predetermined threshold (for example, 4) is assumed to be an area other than the signboard and rejected.

The character string region extraction unit 16 performs binarization processing on the signboard candidate region, and then detects an edge using a four-direction Laplacian filter.
Since the detected edge portion is a region where the pixel value changes abruptly, it is expected to be a boundary region between the background color and the character color of the signboard.
Therefore, an area excluding a part not including edge information in the signboard candidate area is extracted as a character string area.

The output unit 17 displays the contents obtained by the circular sign recognition unit 14 and the character string region extraction unit 16 on a display.
Note that the output by the output unit 17 is not limited to displaying the content on the display, and for example, sound may be output from a speaker.
Next, examples of display images are shown in FIGS.
The signboard area of FIG. 3B is extracted from the color image of FIG. 3A as the signboard area extraction result, and the character string area of FIG. 3C is extracted as the character string area extraction result. I understand.
In addition, a plurality of signs are extracted from the color image of FIG. 4A as a result of sign extraction, and “maximum speed” on the left of FIG. 4B and “vehicle (combination) closure” on the right of FIG. 4B, respectively. It is confirmed that it is recognizable as a sign.
As described above, the recognition object can be recognized by the present invention even in the case where a plurality of signboards or signs exist in one image.

It is a schematic block diagram of the apparatus which implements the recognition method which concerns on one Embodiment of this invention. FIG. 2 is a flowchart of a sign / signboard identifying unit in FIG. 1. It is a display image figure of an example of the result obtained by the present invention. It is a display image figure in an example of the result obtained by the present invention.

DESCRIPTION OF SYMBOLS 11 Image pick-up part 12 Sign / signboard identification part 13 Circular sign information extraction part 14 Circular sign recognition part 15 Non-signboard area rejection part 16 Character string area extraction part 17 Output part

Claims (1)

In the captured color image, the hue range of the color information is divided into three , a hue value in the range of 0 to π / 6, 3π / 2 to 2π, and a saturation value equal to or greater than a predetermined threshold value, π / An area having a hue value in the range of 6 to 5π / 6 and a saturation value equal to or greater than a predetermined threshold, and having a hue value in a range of 5π / 6 to 3π / 2 and having a hue value equal to or greater than the predetermined threshold extracting signboard by the saturation and circularity values labeling area of 0.3 or more and less than 0.8 information in the area, it performs rejection of the area other than the signboard by the information of the circularity, further background in Signs Recognizing signboards using color images, which recognizes signboards by detecting edges using Laplacian filters and extracting character string areas in signboards, focusing on the high contrast between color and text color Method.

Images