JPS62130480A - Encoding method for picture information - Google Patents

Abstract

PURPOSE:To keep a bit of pattern information required for a pattern recognition from being defected by performing the expansion/compression conversion of a code length in a data code string corresponding to a magnifying/reducing magnification to normalize the size of an input image. CONSTITUTION:An image inputted to an image input part 7 is segmented with a digitization processing circuit 8 and a segmentation circuit 9, then being changed to a digital image. An image variable power ratio can be found from a ratio of the size of the digital image to that of a sample image. A segmented image is converted to an encoding chain-coded by a thinning processing circuit 11 and a chain coding circuit 12. A normalization processing circuit 16 performs the variable power processing of the length of the code string corresponding to the image variable power ratio obtained at the segmentation circuit 9. A check processing circuit 14 performs a check processing using a bit of sample image information at a dictionary part 13, and outputs a recognition result 15.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field] The present invention relates to a method for encoding image information, and in particular to a technique that can be applied to a size normalization processing method during preprocessing for pattern recognition.

[Prior Art] Particularly in pattern recognition, normalizing the size of images is important for measuring the similarity between images.

Conventionally, image size normalization has been performed by image enlargement processing or image reduction processing.

In the case of digital images, the enlargement process cuts connected lines as shown in FIG. Due to the reduction process, fine patterns cannot be resolved as shown in FIG. Furthermore, as shown in FIG. 3, image information may be damaged, such as when separated patterns are combined. In pattern recognition, in order to measure not only the similarity based on the simple superimposition of the input image and the sample image, but also the similarity related to detailed pattern structures such as end points, branch points, and intersection points, the enlargement processing and reduction processing described above are used to measure similarity. Losing the information of the pattern structure is
It has the disadvantage of lowering the recognition rate.

FIG. 4 shows an example of conventional pattern recognition. The operation will be explained below.

The image input section 7 includes devices for inputting image data such as a T, a V, and a camera, and receives analog images. The analog image is converted into a binary level digital image by the digitization processing circuit 8. Since the digital image has margins, the segmentation circuit 9 performs processing to extract only the pattern portion, and the size normalization processing circuit performs processing to normalize the segmented image by enlarging or reducing the segmented image to a predetermined size. Do it in 10.

In a line image such as a character pattern, the (center) line is information necessary for pattern recognition, and the line width is unnecessary. For this reason, the thinning processing circuit 11 thins the image (thinning).
Convert to image.

The Cheno coding circuit 12 uses end points, branch points, and intersection points in the thinned image as starting points or end points during Cheno coding.

Converts image information into information encoded with a direction code. The dictionary section 13 stores information on sample images.

The matching processing circuit 14 matches the input image and the specimen image,
When they match, it is output as recognition result 5.

In the above pattern recognition method, the size normalization processing circuit 1
As mentioned above, the digital image enlargement or reduction processing using 0 destroys information on the structure of the pattern, resulting in a reduction in the recognition rate.

[Objective] The present invention eliminates the drawbacks of the above-mentioned conventional examples. At the same time, an easy-to-implement method for expanding/compressing the code length of a Cheno-coded code string is provided.

[Example] Hereinafter, the present invention will be described in detail with reference to the drawings.

FIG. 5 shows an embodiment of the present invention.

The features of the present invention are that the size of the segmented image is directly thinned without size normalization processing (that is, enlargement/reduction conversion), and that the size of the segmented image is normalized using the Cheno-coded code string. This is to perform expansion/compression conversion according to the image conversion magnification. With this point, it is possible to obtain a Cheno-coded code string equivalent to normalizing the image size without damaging the information on the pattern structure.

Examples of the present invention will be described in detail below.

The image input to the image input section 7 becomes a digital image segmented by the digitization processing circuit 8 and the segmentation circuit 9. The image conversion magnification is determined from the ratio of the size of the digital image and the size of the sample image.

The segmented image is converted into a code string subjected to Cheno coding by a thinning processing circuit 11 and a Cheno coding circuit 12.

FIGS. 6 and 7 show examples of thinned images and Cheno-coded code strings.

FIG. 8 shows a direction code that specifies the directional relationship between connected pixels.

The relationship between the thinned image and the Cheno-coded code string will be explained below.

Cheno coding uses end points, branch points, and intersection points in an image as starting or ending points of coding, converts the direction between connected pixels into a direction code, and obtains a code string.

For the pattern “T゛° in Fig. 6, the code sequence shown in Fig. 7 is obtained by running the chain code from end point E1 to branch point T, from end point E2 to branch point T, and from end point E3 to branch point T. .

The length of the code string is proportional to the size of the image.

Therefore, increasing the length of the code string by A times is equivalent to increasing the size of the image by A times.

In order to perform Cheno coding on line segments that have been decomposed from an image using one end point, one branch point, and an intersection point,
Even if the length of the code string after coding is converted, the information on the pattern structure is not lost.

The normalization processing circuit 16 performs a process of changing the length of the code string by A times in accordance with the image conversion magnification A determined by the segmentation circuit 9. Verification processing circuit 14. Dictionary section 13.
The recognition result 15 is the same as the conventional example shown in FIG. 4, and the recognition result is output after performing a comparison process with the sample image information stored in the dictionary section 13.

An embodiment of the normalization processing circuit 16 according to the present invention will be described below.

A digital image is an image that has been subjected to sampling processing, and the number of pixels in the horizontal and vertical directions in one image is an integer.

Furthermore, since the number of pixels in each direction of the normalized image is also an integer, the image conversion magnification for normalization is an integer of integers, for example m/n (m and n are integers).

The Cheno-coded code string has an integer number of code lengths consisting of 1 to 8 direction codes. The normalized code string also has an integer code length.

Consider the case where the code string of the code length in the example sentence is multiplied by m/n. This can be done by decompressing by m times and then compressing by 17 n times.

Expansion processing by a factor of m means that each code of a sentence example is multiplied by m.For example, if the code string is 1357 and m=3, the code string after conversion becomes 111333555777.

1/n times compression processing means that n code strings are extracted from fLm code strings and the average value of each is calculated. However, there is no mathematical formula for calculating the average value of n code strings.
Furthermore, when a table is used, it consists of 8n items, which is not practical when n is variable or has a large value.

The principle of an easy-to-implement method for determining the average value of n code strings related to this patent will be described below.

First, select a number (for example, 2 or 3) for which it is easy to calculate the average value of n code strings using mathematical formula processing or table search processing.
Divide into code strings and find the average value of each.

For example, when n code strings are ala2----am, a
1a2. a3a4. -1--, an-tan,
Each average value b1, b2)---, bm is determined. Code string consisting of each average value b 1 b 2---bm
is a code string consisting of n/2 pieces. If the same processing is performed on the code string, n/4 code strings will be obtained, and by repeating the process, one code string will be obtained. This is n
is the average value of the codes.

During the following processing, the average value of the two codes is now a and b
are two signs, and X is the average value.

INT(Y) is determined by the largest integer that does not exceed the value of y.

FIG. 9 shows the configuration of the normalization processing circuit 16.

20 is a CPU, Do-D5. Al-A3 is CPU
This register is used for operations such as normalization processing. The memory 21 is a RAM, and the memory (1) stores the Cheno-coded code string, and the memory (2) stores the code string m
The expanded and converted data and the normalized data are stored in the memory (3) as one code in the memory of one address each.

A flowchart of the processing of the normalization processing circuit 16 is shown in FIG.

Each process will be explained below.

The normalization process is roughly divided into two parts. Steps 1 to 9 are expansion conversion processes that multiply the code string by m, and steps 10 to 35 are expansion conversion processes that multiply the code string by 1/n. Performs double compression conversion processing.

First, the expansion conversion process will be explained. In the flow chart shown in FIG. 10(a), steps 1 and 2 load the respective start addresses of memory (1) and memory (2) into register A1 and register A2. With this, CPU and memory (1) and memory (
2) Make the correspondence.

The step is a magnification (m
) and code length (sentence) values. This sets the initial values for the following processing.

Step 4 loads the contents of the memory at the address stored in register Al into register DI.

Step 5 stores the contents of register D1 in the memory at the address stored in register A2.

This process is performed m times in steps 6 and 7, and m contents of the register D1 are stored in the memory (2). As a result, the first code in the code string is expanded by m times.

In step 8, 1 is added to the contents of register A1. As a result, in step 4, the memory contents of the next address are loaded into the register D1, and in steps 5 to 7, the second code in the code string is expanded by m times.

In steps 8 and 9, the above processing is performed on the codes of all sentence examples in the code string obtained by chaincoding.

As a result, m=3 in code string 1357. When intersection=4, the code string after conversion becomes 111333555777, and is stored in the memory (2) in this order.

As described above, a code string subjected to decompression conversion processing is obtained. The code string is subjected to a compression conversion process in steps 10 to 35 shown in FIG. 10(b) to obtain a normalized code string.

The compression conversion process will be explained below. In steps 10 to 16, conversion when n=1, that is, no compression is performed, and only transfer from memory (2) to memory (3) is performed. In steps 11 to 13, initial values are set, and in step 14, address A of memory (2) is set.
The contents of 1 are transferred to address A2 of memory (3), and this process is carried out at step 15 while changing the values of Al and A2.
Execute times.

The conversion in the case of n#1, that is, the compression process, is shown in Figure 10 (c
), steps 17 to 35 shown in (d) are performed.

In the loop from step 21 to step 31, two or one code is extracted from the code string, an average value is obtained, and it is stored in the memory (3). Here step 26 is 2
1) Calculate the average value using formulas 2 to (3) for each code. Step 28 stores the average value in the memory (3).
.

Ru.

In the loop from step 20 to step 33, the previous process is repeated to convert n codes into one code.

In the loop from step 20 to step 35, the code string of m sentence examples is converted into n code strings, each of which is converted into one code.

It can be normalized by the expansion conversion processing and compression conversion processing described above.

Note that this technology can be applied not only to recognition technology but also to patterning of graphic information and normalization in interpolation processing technology between graphics.

[Effect] As explained above, by normalizing the size of a code string, there is an effect that the recognition rate can be improved without losing pattern information necessary for pattern recognition.

Because it handles one-dimensional information called code strings, processing is simple and fast.

[Brief explanation of drawings]

Fig. 1 shows an example of enlargement processing, Figs. 2 and 3 show an example of reduction processing, Fig. 4 shows a conventional embodiment, and Fig. 5 shows an embodiment of the present invention. 6 is a diagram showing a thinned image, FIG. 7 is a diagram showing a chain-coded code string, FIG. 8 is a diagram showing a direction code, and FIG. 9 is a configuration diagram of a normalization processing device. FIG. 1O is a flowchart of the normalization process. 8-1--Digitization processing circuit, 12---Chaencoding circuit. Sixth normalization processing circuit. Tomi 1 Figure Lower 20 High School 3 Kasumi EEI Ez
Elephant - one gold, one umbrella, one thing, one army, one life, one car, one spear, one turtle, one turtle, one sword, one bag, one army, one army ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ Engineering ↓E. (El)711 f 1 f r f (T) (Ez)
55 de 55";5'5 ke) (El) 33:3 wo 33333333 u (ding) punishment io [
p(d)

Claims (3)

[Claims] (1) Convert input image information into a chain-coated code string (data code string), and determine the code length of the data code string according to the enlargement/reduction conversion magnification for normalizing the size of the input image. 1. A method for encoding image information, the method comprising expanding/compressing the image information. (2) In claim 1, when the code length of a code string is expanded/compressed, the expansion/reduction conversion magnification is m/
Integers m and n that are n or approximately m/n are found, and the code string is first expanded by m times and then compressed by 1/n times. Encoding method. (3) In claim 2, when performing 1/n times compression conversion, the code string expanded and converted by m times is divided into a code string group consisting of n or less than n codes; An image characterized in that when each divided code string is converted into one code (that is, the average value is determined), division processing and averaging processing are repeated for the divided code strings until the average value is determined. A method of encoding information.