JPH0632078B2 - Image signal processing method - Google Patents

Description

TECHNICAL FIELD The present invention relates to an image signal processing method applied to a pattern recognition device such as an OCR.

(Prior Art) A recognition method of a pattern recognition device such as an OCR can be roughly classified into an analog matching method (template matching method) and a structure analysis method. The former is a method of recognizing a pattern by superimposing a standard pattern and an unknown pattern, which is excellent in versatility and is particularly advantageous for reading a printed character.

In the paper "Neural circuit model of selective attention mechanism in visual pattern recognition" by Fukushima published in Technical Report of IEICE, 1986, MBE, vol.85, p.225, hierarchical pattern matching is performed by neural network with feedback The method is described. In this method, pattern matching that allows positional deviation is performed in each layer by preparing a set of characteristic patterns whose positions are slightly shifted. The candidate pattern obtained by hierarchically performing such a matching process is returned further along the top-down signal path, and the pattern matching is performed while correcting the input pattern according to the candidate pattern. As a result, it is possible to perform strong pattern matching against positional deviation or background noise, which is a drawback of template matching.

(Problems to be Solved by the Invention) In this conventional method, it is necessary to prepare a template whose position is shifted for each characteristic pattern. Therefore, when trying to improve the performance of pattern recognition, a large number of templates are stored. There was a difficulty that I had to keep it. An object of the present invention is to realize pattern matching that is strong against positional deviation and background noise without using a large number of templates.

(Means for Solving the Problem) In order to solve the above problems, the present invention divides an input image into many small regions, and the input image is Fourier-transformed and the Fourier transform power of each input region for each small region. It is configured such that it is converted into a spectrum pattern and pattern matching is performed between the Fourier transform power spectrum pattern for each small region and a standard pattern prepared in advance. In order to realize strong matching against background noise, the candidate patterns are fed back to the image input unit along the top-down signal path, and the pattern matching is performed while modifying the input image according to the candidate patterns.

(Operation) The input image is divided into many small areas, and the input image image is converted into the Fourier transform and the pattern of the Fourier transform power spectrum for each small area. Next, the pattern of this power spectrum is subjected to template matching with a standard pattern prepared in advance, and a candidate pattern is selected. As is well known, since the Fourier transform power spectrum of the pattern is invariant to the translational movement of the pattern, this matching process is strong against the displacement of the input image. Moreover, in the method according to the present invention, it is not necessary to prepare a large number of sets of templates obtained by slightly shifting the positions of various characteristic patterns.

Next, the candidate pattern selected in this way is used as a mask to filter the Fourier transform pattern obtained for each small region, and the input image is corrected by the pattern obtained by inverse Fourier transforming the output. When the input image matches the candidate pattern, the pattern obtained by the inverse Fourier transform matches the input pattern and is not modified. When background noise is superimposed on a region where no candidate pattern exists, such as the small regions (1) and 6 shown in the explanatory diagram of FIG. 3, filtering is performed regardless of the spatial frequency component of the noise. Sometimes the components in that region are cut, so the pattern obtained by the inverse Fourier transform is a small region.
As shown in (1) and 10, the background noise is suppressed. In addition, like the small areas (2) and 7 in FIG.
Even if the background noise overlaps the candidate pattern in position, if the main spatial frequency component of the background noise is on a different frequency range from that of the candidate pattern, the noise component is also cut during filtering and the inverse Fourier transform is performed. Noise is removed from the pattern obtained by the method as shown in the small areas (2) and 11. In the filtering mask 8 in FIG. 3, a hatched portion shows a spatial frequency region to be cut.

Therefore, if a candidate pattern is included in the input image,
By repeating the above matching process while making such corrections, the background noise is removed from the input pattern, and the similarity of matching increases, so in the present invention, pattern matching that is strong against noise is performed. realizable.

(Example) Hereinafter, one example of the present invention will be described with reference to the drawings.

FIG. 1 is a flow chart showing an embodiment of the present invention.

The image information I ₀ of the pattern input from the television camera or the image scanner is stored in the image memory in which the brightness is digitized for each pixel.

In the process shown as process 1 in FIG. 1, the image I used in the following process is calculated, but the input image information I ₀ is taken as I as the first process. Next, in the portion shown as process 2, the following conversion is applied to pattern I using a processor or the like.

Now, the number of pixels of the input image is N × N. First, this input image is divided into n ₂ × n ₂ small areas each having the number of pixels n ₁ × n ₁ . Here, when the images are divided so that they do not overlap each other, N = n ₁ × n ₂ , but they may be divided so that they overlap each other.

Next, the processing for obtaining the Fourier transform and the Fourier transform power spectrum is performed on each image divided into these small regions as follows.

Here, each small area is represented by a subscript X = (X, Y) (X, Y = 1 to n ₂ ),
The position of the pixel in each small area is the subscript x = (x, y) (x, y = 1 to n
Expressed as ₁ ). The value of the pixel at the position of x in the small area represented by X is represented by I (x; X).

F (k, X) = Σ x I (x; X) exp (-i2 II kx / n 1) k = (k x, k y), (k x, k y = 1~n 1) x = ( x, y), (x, y = _{1 to} n ₁ ) …… (1) P (k, X) ＝ | F (k, X) | ² …… (2) Equation (1) is calculated as I (x ; X) represents a two-dimensional discrete value Fourier transform with respect to the subscript x. For calculation of equation (1), it is possible to use a fast Fourier transform algorithm.
P (k, X) in equation (2) is the Fourier transform power spectrum. Further, in order to eliminate the discontinuity at the boundary of the small area, a window function such that the boundary becomes zero, for example, Gaussian or Hamming window may be used in the Fourier transform.

Next, in the process 3, the similar conversion is performed in advance with the pattern of the power spectrum thus obtained for each small region, template matching is performed with the standard pattern, and the similarity S is calculated. Determine the candidate pattern. If the degree of similarity is smaller than a predetermined threshold S ₁ here, the input is rejected and the process ends. If S is larger than a predetermined threshold value S ₂ ,
After that, the process ends as a result of the determination. In other cases, the matching process is continued while correcting the input image according to the candidate pattern as follows.

Here, the candidate pattern is P '(k, X). In the process 4 portion, the following filtering operation is performed on the Fourier transform pattern for each small region obtained in the process 2 using the candidate pattern.

F ′ (k, X) ＝ F (k, X) ₁ P ′ (k, X) / P (k, X) P (k, X) ≠ 0 = 0 P (k, X) = 0 If. (3) In the processing 5, the modified Fourier transform F ′
Inverse transformation is applied to (k, X).

I ′ (x; X = 1 / n ₁ ² ΣF ′ (k, X) exp (i 2 _II Kx / n ₁ ). Here, the process returns to the process 1 again, but from the second time onward, the feedback signal I Compute the modified I (x, X) by taking the weighted average of ′ (x, X) and the first input image I ₀ (x, X).

I (x, X) ＝ aI ′ (x, X ′) + (1-a) I ₀ (x, X) …… (4) a is a parameter that determines the magnitude of the top-down effect due to the feedback signal. Set between 0 and 1.
Further, in processing 1, threshold processing is performed, and when the obtained I takes a negative value, it is replaced with 0.

The conversion of the process 2 and the pattern matching of the process 3 are repeated for the I thus modified. This loop is repeated until it exceeds the threshold value S ₂ in which the degree of similarity during the matching is predetermined, if not have sufficient similarity be reached repeat count n ₀ that has been determined in advance,
In the process 6, the candidate pattern is changed to the next one, and the same operation is continued.

FIG. 2 is a diagram showing an example of an OCR device in which the embodiment of FIG. 1 is realized by a multiprocessor configuration.

The OCR device shown in FIG. 2 is an image scanner 1 which is a character input device, an A / D converter 2 for digitizing the input pattern from the scanner, and a pattern in a corresponding small area provided for each small area. Of n local processors 3 in charge of the processing, and a pattern judgment unit 4 for selecting a candidate pattern by template matching with a standard pattern and controlling the number of repetitions of matching.
Each local processor of 3 has its own local memory, Fourier transform of the pattern in the corresponding small area, calculation of the Fourier transform power spectrum, filtering and inverse of the Fourier transform pattern according to the candidate pattern established in the pattern determination unit. The input pattern is corrected based on the transformed and inverse transformed patterns.

Many parts of the process according to the present invention can be executed in parallel for each of the divided small areas. In the apparatus of FIG. 2, the local processors provided for each small area operate in parallel, so that high speed processing can be performed.

In the present embodiment, the case where the present invention is applied to OCR has been described. However, by replacing the input device with a television camera or the like, for example, a pattern recognition device for images for identifying a part from a two-dimensional shape. Also, the present invention can be easily applied.

As is well known, the Fourier transform power spectrum of a pattern is invariant to the translational movement of the pattern. Therefore, in the method according to the present invention, the positional shift of the pattern (the positional shift of about the size of each of the divided small regions). ), A strong pattern recognition method can be realized.

(Effects of the Invention) As described above, the present invention performs matching by using the Fourier transform power spectrum pattern for each small region of an image. It has an effect that it can be realized without preparing a large amount. Also, according to the candidate pattern, the matching is performed while correcting the input pattern by the top-down signal, so the background noise of the spatial frequency component that the candidate pattern does not have is removed in the process of correction, so it becomes a background noise. Strong pattern matching can be realized.

[Brief description of drawings]

FIG. 1 is a flow chart showing an embodiment of the present invention, and FIG.
FIG. 3 is a block diagram in the case where the pattern recognition apparatus is applied with the processing according to the present invention in a multiprocessor configuration, and FIG. 3 is an explanatory diagram for explaining the operation of the present invention. In the figure, 1 ... Scanner, 2 ... A / D converter, 3 ... Local processor, 4 ... Pattern determination unit, 5 ... Input image, 6 ... Small area (1), 7 ... Small area ( 2), 8 ... filtering master, 9 ... inverse Fourier transformed pattern,
10 …… Small area (1), 11 …… Small area (2)

Claims (1)

[Claims] 1. An input image is divided into small regions that overlap each other or do not overlap each other, and the input image is converted into a Fourier transform and a Fourier transform power spectrum pattern for each of these small regions to obtain a power spectrum pattern. Template matching is performed between the pattern and the standard pattern prepared in advance, the candidate pattern thus obtained is used as a mask to filter the Fourier transform pattern for each of the small regions, and the output pattern is inverse Fourier transformed. A method for processing an image signal, characterized in that an input image is corrected by the obtained pattern, and pattern matching is performed on the corrected input image while repeating the above processing.