JPH08212290A - Feature amount preparing device in pattern identification - Google Patents

Abstract

PURPOSE: To obtain the feature amount of a high identification rate of a recognition result by preparing one feature amount by incorporating a distance feature amount and a projection feature amount. CONSTITUTION: Data picked up by a TV camera 1 is stored as still picture data in a picture storage part 2, and a number plate extraction part 3 extracts the part of a number plate and segments a character next. Segmented character data is given a density normalization processing by a density normalization part at first and the preparation processing of a feature amount being a feature constitution element by a distance/projection feature preparation part 5a in a feature amount preparation part 4 next. In this case, the distance and projection feature amounts from a prescribed reference position are prepared. Then, the feature amount obtained by incorporating the distance and projection feature amounts is prepared at last. The normalized feature amount is used as the input cell of an identification processing by a neural network in a next character recognition part 6 to recognize a branch station character.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to, for example, an automobile for a TV.
The present invention relates to a feature quantity creating device in pattern identification when pattern recognition is performed using a detected feature quantity when the contents of a license plate for identifying a vehicle are read from the image data obtained by shooting with a camera or the like. .

[0002]

2. Description of the Related Art FIG. 3 shows a license plate P of an automobile C.
4 is a schematic configuration diagram of the image processing device a when reading the contents of the above, that is, the contents of the land transportation branch office (hereinafter referred to as “branch office”), the vehicle number, and the like, and FIG.

The image processing apparatus a is constructed so that the data obtained by photographing with the TV camera 1 provided above the road surface can be stored in the image storage unit 2 as a still image. Reference numeral 1'denotes an infrared illuminator, which is devised so that good photographing data can be obtained even at night when the illuminance is insufficient.

The data stored in the image storage unit 2 is
It is digitized and the portion corresponding to the license plate P is extracted by the license plate extraction unit 3. And
After the feature quantity creating unit 4 creates the feature quantity for identifying a character (including numbers and symbols), the character recognizing unit 6 recognizes the character using the feature quantity. The monitor 7 is provided to monitor each of these processes. Further, the distance / projection feature creation unit 5a and the size normalization unit 5b included in the feature quantity creation unit 4 are characteristic component parts of the present invention, which will be described later.

Each of the above parts is a microcomputer (CP
U) and the obtained character recognition output is sent to the management center and used for grasping traffic volume and speed control. Further, when the image processing device a is provided in the parking lot, it is used for the entry / exit management of the vehicle.

FIG. 5 shows the license plate extracting unit 3 of FIG.
FIG. 5A shows a procedure for creating a distance feature amount of a station character (“Adachi” in the example shown in the figure) cut out by character cutting from the image data of the license plate extracted in FIG.
reference).

The cut-out station character is subjected to density normalization processing as the original multivalued data shown in FIG. 5A in order to perform the binarization processing as a preprocessing for obtaining the distance feature amount. (See FIG. 5B).

Next, the density-normalized image data is binarized, and if the density is higher than a predetermined value, it becomes "1" (black portion in FIG. 5C), and if the density is lower than the predetermined density. Is set to "0" (white part in FIG. 5C). Note that FIG.
In (c), a thin rectangular box surrounding the station character "Adachi" is a reference line for character extraction, and the reference line is a reference position in four directions when creating the distance feature amount.

With respect to the local character that has been binarized as described above, the distances from the above four directions are obtained and the distance feature amount a'is created. This distance feature amount a ′ is “1” (black portion in FIG. 5C) measured in the direction perpendicular to each reference position.
It is expressed as a total of the number of "0" s (white portions in FIG. 5C) until hitting (the number of dots hitting "0") (FIG. 5).
(See (a)).

The obtained distance feature quantity is, for example, Japanese Patent Laid-Open No.
Using a neural network as disclosed in Japanese Patent Laid-Open No.-303668, size normalization is performed so that an input cell of a predetermined size for character reading is obtained, and a distance feature amount a of a predetermined size is generated. (See FIG. 5 (e)).

[0011]

However, when pattern recognition is performed on the basis of the distance feature amount of the above-mentioned conventional apparatus, the original image data taken is distorted. Not from the front
When the license plate is tilted or when the license plate is imaged obliquely from the front, there is a drawback that an accurate distance feature amount cannot be obtained and the recognition rate of the recognition result decreases.

More specifically, a TV camera is provided on the road side and the lower part of the license plate is extremely located on the vehicle side so that the image processing apparatus is used for parking lot management. In this case, the image data is distorted, and the station character “Adachi” in FIG. 2 becomes a single continuous character in the horizontal (horizontal) direction, which makes it impossible to obtain an accurate distance feature amount.

Therefore, the present invention has been made to solve the above-mentioned drawbacks, and the purpose thereof is to basically use a distance feature amount, but another feature amount that can complement the distance feature amount. It is to provide a feature quantity creation device in pattern identification in which a feature quantity in which the recognition result is added is created and the recognition rate of the recognition result is high.

[0014]

In order to achieve the above-mentioned object, a feature quantity creating apparatus for pattern identification according to the present invention is capable of processing binary image density data obtained by processing image data from a predetermined reference position. Distance feature amount creating means for creating a distance feature amount by measuring up to the position of "1" and projective feature amount of the binarized density data obtained by processing the imaging data from the predetermined reference position. The present invention is characterized by having a projective feature quantity creating means and a feature quantity creating means for creating one feature quantity by combining the created distance feature quantity and the projected feature quantity.

[0015]

In the above structure, the feature quantity creating means creates the feature quantity from the distance feature quantity and the projected feature quantity.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to the flow chart of FIG. 1 and the explanatory view of FIG.

The present invention, as shown in FIGS.
A case will be described in which a station character (“Adachi” in this example) is identified from the license plate P of the automobile C. First,
Also in the present embodiment, as in the conventional case, the data captured by the TV camera 1 is stored in the image storage unit 2 as still image data (step 100; hereinafter, step is referred to as S). Then, the license plate extracting unit 3 extracts the license plate P portion (S102), and then cuts out characters (“Adachi” in the illustrated example) (S1).
04). FIG. 2A shows the cut out original pattern of multi-value density.

As described above with reference to FIG. 5, the extracted character data is first subjected to density normalization processing by the density normalization section 5a (S106. FIGS. 2B and 2C). Then, the distance / projection feature creating unit 5a of the feature creating unit 4 performs a feature amount creating process which is a characteristic component of the present invention.

In short, also in the apparatus of this embodiment, the same pretreatment as that shown in FIGS. 5 (a) to 5 (c) is performed. Therefore, FIGS. 5A to 5C correspond to FIGS. 2A to 2C, respectively. Although not shown in FIG. 2, a distance feature amount a ′ from a predetermined reference position is created in the same manner as shown in FIG. 5D above (S108).
a).

Next, the character cutting-out frame line on the upper side is set as a reference position, and the number ("1") of the portions (black portions in FIG. 2 (c)) measured at a right angle from the reference position and hitting "1". The number of dots) is measured, and a projected feature quantity b'as shown in FIG. 2D is created (S108b).

This projective feature quantity b ′ may be created not only in the above-mentioned reference position of the upper side, but also in the projective feature quantities b ′, b ′ ... Of the left and right side reference positions and the lower side reference positions. As will be described later, this projective feature amount is used to complement the above-described distance feature amount, and it is sufficient if there is a projective feature at any one of the reference positions. If the number is small, it is possible to obtain the recognition result without increasing the load of the neural network processing of the next stage. In the case of reading the station characters on the license plates arranged side by side as in the present embodiment, it is desirable that the characters are classified as the projected feature amount from the reference position on the upper side or the lower side.

The projected feature quantity b ′ obtained as described above.
Are merged by making them correspond to the same reference position of the created distance feature with a predetermined interval maintained, and then the size is normalized so as to be the size of the input cell of the neural network in the next step. (S108c), finally, as shown in FIG.
A combined feature amount of the distance feature amount a and the projected feature amount b shown in (e) is created.

The normalized feature quantity is used for the next character recognition unit 6
The local network character (“Adachi”) is identified by the neural network in S as an input cell of the identification process (S).
110). In addition, the identification processing by the neural network,
Since the processing load can be reduced if the number of input cells is small, it is preferable that the number of projective feature quantities B is small.

As described above, the apparatus according to the present embodiment uses not only the distance feature quantity as in the conventional apparatus but also the feature quantity in consideration of the projection feature quantity. Therefore, even if the image data is distorted. The character identification rate can be increased.

In the above-mentioned embodiment, the original two-dimensional data is the image pickup data picked up by the TV camera 1.
This may be recorded on negative paper. Therefore, the term "imaging data" as used in the present invention means that
It also includes dimensional data.

Further, in the above embodiment, "1" is used as the basis for binarization at the time of creating the distance feature quantity, but "0" may be used as the basis. Further, the binarization at the time of creating the projective feature value is based on "1", but this may be based on "0". That is, the reversal information may be used as the basis. Therefore, the present invention includes that "1" may be the opposite "0" or "0" may be the opposite "1".

[0027]

As described above, the feature quantity creating apparatus in the pattern identification according to the present invention measures the distance feature by measuring from the predetermined reference position to the position "1" of the binarized density data obtained by processing the image pickup data. Distance feature amount creating means for creating an amount, projective feature amount creating means for creating a projective feature amount of the binarized density data obtained by processing the imaging data from the predetermined reference position, and the created distance feature Since the feature quantity creating means creates a single feature quantity by combining the quantity and the projected feature quantity, a feature quantity with a high recognition rate of the recognition result can be obtained.

[Brief description of drawings]

FIG. 1 is a flowchart for creating a feature quantity according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram when creating a feature amount of the present invention.

FIG. 3 is a schematic diagram showing an example of an image processing apparatus for identifying characters on a license plate of an automobile.

FIG. 4 is a block diagram of an image processing system.

FIG. 5 is an explanatory diagram when creating a distance feature amount.

[Explanation of symbols]

 1 TV camera 1'Infrared illuminator 2 Image storage unit 3 License plate extraction unit 4 Feature amount creation unit 5a Distance / projection feature creation 5b Size normalization unit 6 Character recognition unit 7 Monitor C Car P license plate a Image processing device

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI technical display location G08G 1/017

Claims (1)

[Claims] 1. A distance feature quantity creating means for creating a distance feature quantity by measuring from a predetermined reference position to a position of "1" of binarized density data obtained by processing image pickup data, and the predetermined feature. Projection feature quantity creating means for creating a projection feature quantity of the binarized density data obtained by processing the imaged data from the reference position of 1 and the created distance feature quantity and the projected feature quantity are combined into one feature. A feature quantity creating device for pattern identification, comprising: a feature quantity creating means for creating an amount.