JPS58181179A - Extracting device of picture feature - Google Patents

Abstract

PURPOSE:To extract features of a picture easily and efficiently, by using masks where patterns are generated by values indicated by the Hermite polynomial weighted with Gaussian function. CONSTITUTION:Different mask patterns indicated by the Hermite polynomial weighted with Gaussian function are stored in mask memories 13a-13n, and the read of them is controlled by an address controlling part 12. Information of mask patterns read out from memories 13a-13n are selected through a selector 14, and one of them is selected and given to a product sum circuit 15. In this circuit 15, information of the selected mask pattern and picture information of a specific area read out from an original picture memory 11 are subjected to mask operation. The processing result is distributed to processed picture memories 17a-17n corresponding to memories 13a-13n through a selector and is stored in a pertinent address.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to an image % image extracting device that can easily and effectively extract % characteristics of a given image.

[Technical background of the invention and its problems]

Handwritten characters are recognized by extracting features of an image containing the characters. Conventionally, mask calculation, line element tracing, and moment-based methods have been exclusively used for image feature extraction. Among these, the mask calculation described above has been used more widely than ever because of its wide applicability.

Now, this mask operation basically uses the mask information H(r, τ) for the image information gb), and performs the following congolution operation to obtain the information h of another image.
It means to create (r, τ). However, τ in the above formula is a gelameter related to the nature of this operation. By this mask calculation processing, the percentage characteristics of the image information g(r) are emphasized and appear as the other image information h(r, τ).

Well, a con like this? The above-mentioned mask calculation indicated by the equation is actually performed by a sum-of-products process as shown in FIG. That is, the image information tube of the target position 2 and its surrounding area 3 of the image stored in the image memory 1 is accessed and read out, and the mask 9 turn information read from the mask memory 4 is applied to the multipliers 5m, 5b to 5n. A sum-of-products process is performed in which the outputs are multiplied in accordance with the respective pixel positions at , and then the outputs are added at an adder 6 . Then, this processing result is applied to the pixel of interest i12 in the processed image memory 7.
The masking process for the image information of the target pixel position 2 is completed. Thereafter, this process is repeated one after another, and masking operations (product-sum processing) are performed for each of all pixel information of the image, thereby completing the masking process for the image stored in the image memory 1.

However, in this mask calculation, the processing results vary greatly depending on the value of the mask pattern, and therefore, what kind of mask to use is a big technical problem. Conventionally, as this type of mask, for example,
Lines such as those shown in Figures (a) to (d) are often used to extract the directionality of line elements. However, even when performing % image extraction using such a mask, the results were not completely satisfactory in recognition processing. In other words, as a mask for feature extraction, (1) the necessary information can be extracted, (II) the extracted information does not overlap with each other, and
In addition, it is required to have characteristics such as high efficiency and stability against (ij+) image noise. However, conventionally, it has been extremely difficult to extract image features using a mask that satisfies these requirements.

[Purpose of the invention]

The present invention was made in consideration of these circumstances, and its objectives are: - A highly practical image percentage that can efficiently and easily extract image features using an appropriate mask. An object of the present invention is to provide an image extraction device.

[Summary of the invention]

The present invention uses a plurality of mask patterns represented by a value expressed by a Hermy (H@rmit) polynomial of different orders weighted with a Gaussian function as masks used in mass multi-operations. This is how it was done.

〔Effect of the invention〕

As a result, according to the present invention, it is possible to extract image features efficiently and stably against image noise, and it is possible to effectively obtain the features required for recognition processing, which is of great practical importance. be effective.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

Each lI! that constitutes the mask pattern of the mask used in this example! The prime value is expressed by a Hermitian polynomial weighted by a Gaussian function. FIGS. 3(a) to 3(d) each show an example of this, in which the values on each lattice point are approximated. The value Pm, n of each lattice point expressed as this Hermitian polynomial
(reτ) is expressed as follows.

5- However, in the above formula, r is given as r:Xi yl, and Hm#Hn is known as a Hermitian polynomial.

This Hermitian polynomial is basically n with respect to
It is expressed as the following polynomial.

Thus, according to the mask pattern represented by the value expressed by the polynomial of EL1−) weighted by this strict Gaussian function, the function pm'! 'n has the following gender 6 quality. That is, first, m and n shown as Hm and Hn indicate the order and correspond to the vibration component of the mask/mother turn. The value can be arbitrarily set from "0" to an appropriate value. Therefore, by using these mask patterns, it becomes possible to effectively extract complex information (features) according to necessity.

Secondly, the above functions (P,=,) have a mutually In-intersecting relationship with respect to different values of m and n. That is, now, the adjoint function (Pm, .) of the above function (P:n) is defined as follows. At this time, the inner product is * (Pm-n @Pm', B /) * (Pm, n "pm'n')"'δmm'・δ
It becomes nn'. In other words, this means that the functions (pm-) are orthogonal as described above.

Therefore, if mask processing is performed using a mask pattern indicated by the values of such orthogonal functions, it becomes possible to extract each image feature independently without duplication, and the efficiency is extremely high. expensive. Thirdly, according to the above function, when the values of m and n are both "0"', information regarding one time in the area around the image position of interest can be effectively extracted.

Furthermore, when m = 0 and n = 2, it is possible to extract the % image information regarding the horizontal component of the image, and when m = 2 and n = 0.

In addition, the information on the direction component extracted in this way is not interpreted geometrically as in the past, so even if the geometric information is disturbed by noise, it is possible to effectively It is possible to extract In other words, it is possible to stably extract the % mold against image noise.

In this way, as mentioned above, if the % image extraction of the image is performed using a mask that is formed by i+ amount of turns by the value expressed by the Hermitian polynomial weighted by the Uth function, the required information will not be duplicated. It is possible to efficiently and stably obtain the percentage characteristics of image noise.

Now, the present apparatus that performs such feature extraction processing is configured as follows, for example. FIG. 4 is a schematic configuration diagram of the embodiment device. The original image memory 11 stores image information used for feature extraction processing, and the address control unit 12
The image information of the pixel position of interest and its surrounding area is accessed and read out under direct address control. On the other hand, a plurality of mask memories JB&, 13b ~1
3n stores different masks and mother turns each represented by a Hermitian polynomial weighted according to an Us function, and read-out thereof is controlled by the address control section 12. However, these l) mask memories 13m, 13b~
The mask/quantity turn information read from 13n is selected via the selector 14, and one of the mask/quantity turns is selected via the selector 14.
is selected and applied to the product-sum circuit 15. This product-sum circuit 15 performs a mask operation (# sum processing) between the information of the selected mask pattern and the image information of the specific area read out from the original image memory 1. However,
The processing results are distributed and supplied to the processed image memos 11m and 17b to 17n corresponding to the mask memories 13&, 13b to 13n, respectively, via the selector 16, and are stored at the corresponding addresses. ing.

Here, the product-sum circuit 15 is, for example, as shown in FIG.
Multiplier 15*s adder 15b and accumulator 15
It is composed of c. The 15 10-product-sum circuits supply the multiplication value of the image signal at the corresponding position and the mask signal, which has been multiplied by the multiplier 15h, to the adder 15b, and add the sum '! which is the output of the accumulator 15a. J4
After adding it to m, it is stored again in the accumulator 15a, and this process is sequentially repeated to obtain the respective sum value.

Also, when this accumulator 15a starts its processing,
It is reset and initialized by the address control section 12. Thereby, masking processing for the image position of interest is performed independently.

On the other hand, the address control section 12 is configured as shown in FIG. 6, for example. That is, the dimensional counter 20 performs a count corresponding to the mask width (horizontal direction) in the circuit,
Its carry output is given to another dimension counter 21. This dimension counter 21 is designed to count by 7, which corresponds to the mask width (vertical direction).

The mask target area is scanned by these counters 20 and 21, and the count value is stored in the mask memo IJ 13a.
, 13b to 13n are output as address signals. On the other hand, information on the difference between the horizontal position of the original image and the horizontal position of the mask is set in the constant register 22, and this information is given to the offset address register 24 via the adder 23. . This address register 24 is set by the counter 20, and its value is fed back to the adder 23, information on the -F difference is added thereto, and this is registered again. The offset information sent to this register at 24V is
Original image memory 1, which will be described later, via adders 25 and 26 in order.
It is designed to be output as a 1 address signal.

That is, the adder 25 includes a mask position counter 2 which receives a carry signal from the counter 21 and performs an increment operation.
7 is added to the output of the register 24,
Further, the adder 26 further adds the value of the counter 20 to this value. As a result, the original image memory 11 is accessed sequentially for images of a specific area consisting of the image position of interest and its surrounding area. Therefore, the mask memories 13m, 13b to 13n and the original image memory 11 are accessed by corresponding addresses. Further, the constant register 28 stores information regarding the horizontal width and the vertical width of the mask (width x height ÷ 2), and this information is added to the output of the mask position counter 27 via an adder 29. It has become so. This value is the processed image mail E:I) 17a, 17
It is output as an address control signal for b to 17n. still,
The offset address register 24 has a counter 21
This carry signal is output as a reset signal to the product-sum circuit 15.

According to the address control unit 12 configured in this way, the image 13- information in the peripheral area is read out in sequence in correspondence with the image information of the mask, with the image position of interest on the original image memory 11 as a reference. Then, the sum of products is processed.

As described above, according to the present invention, image features can be extracted easily and effectively. Moreover, the required feature information can be extracted stably without duplication and against image noise, so its practical advantage is very great.

Note that the present invention is not limited to the above embodiments. For example, the product-sum calculation may be performed in parallel to speed up the calculation. Further, the mask shift width may be given in steps. Further, when the size of the mask and the image are the same, only an inner product operation may be performed instead of the above-mentioned conrrelation operation. In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an example of a conventional device; FIG. 4 is a schematic configuration diagram of a device according to an embodiment of the present invention, FIG. 5 is a configuration diagram illustrating an example of a product-sum circuit, and FIG. 6 is a diagram showing an example of such a mask lean. It is a block diagram showing an example of 11... original image memory, 12... address control section,
I Ja a 13b~13n-...-q quick memory,
14゜16... Selector, 15... Product-sum circuit, 77
m. 17b to 17n...processed image memory. Applicant's agent Patent attorney Takehiko Suzue 15- Figure 1 Figure 2 (a) (b) (c) (
d) Section: (a) (C) Figure 3 (b) (d)

Claims (1)

[Claims] Multiple different polynomials represented by Hermitian polynomials of different orders weighted with Gaussian functions.
A plurality of masks each having a mother turn, a means for accessing image information of a position of interest in a given image and its surrounding area, and a mask operation using each of the plurality of masks for the accessed image information. An image extracting device characterized by comprising means for extracting an image.