JPS61240384A - Image processing device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Field of Application of the Invention] The present invention provides an Idj system for long-term identification of images from an object.
The present invention relates to a processing device, and particularly relates to so-called region segmentation, which extracts a region to be identified based on features of a density distribution obtained from an input image.

[Background of the invention]

In fields such as FA, there is a strong need to use image processing to automate visual inspections such as checking printed circuit boards and inspecting foreign substances in drugs, and in response to this, various image processing (g#I&) devices have been developed. There is. These image processing apparatuses generally perform preprocessing such as noise removal from an input image, extract a region to be identified, and perform grating on the identified target. For example, in the case of inspecting a drug for scratches, the tablet is first extracted from the input image, and then the scratched area within the tablet is extracted to recognize the scratch.

However, in such an image processing apparatus, developing an object recognition procedure (algorithm C) and a program requires manpower and time. As an example of this kind of software development, Noriko Sato, Toshiyuki Goto, and two others presented "Development of General-Purpose Image Processing Devices - Software" at the 28th Information Processing Society of Japan (Information Processing Society of Japan).
First half of 1981) Collection of papers from national conference lectures, 4N-98, 9
95-996, but the development of the software shown here required an enormous amount of time even for experts.

Particularly, among the above-mentioned processes, developing a program for so-called area segmentation, which extracts areas to be identified from an input image, is a very painful, tedious, and time-consuming task for the user of the image processing apparatus.

The above-mentioned area segmentation methods used for visual inspection etc. generally include α) Normalization processing of concentration frequency distribution (2) Extraction processing using feature values of local concentration distribution (
3) It consists of three steps: final extraction processing using geometric features.

The density frequency distribution normalization process (1) is to perform preprocessing such as noise removal from the input image.

In other words, in order to track changes in the intensity of the illumination irradiated to the inspection object and the contrast between the inspection object and the background, the density frequency distribution of the input image is adjusted to have characteristic peaks such as the average value and maximum frequency. Using the concentration value at the location as a guide,
Some almanacs provide characteristic points and use them as a guide to enlarge or reduce the reference state.

Characteristic points used in the above translation include the minimum/maximum density level of the image, the midpoint of the minimum/maximum density level, the average density level, the density level showing the maximum frequency, and the middle of the peaks in the density frequency distribution. There are concentration levels of existing peaks, concentration levels of peaks that indicate the minimum or maximum concentration among the peaks in the concentration frequency distribution, etc., and multiple suitable feature points to be used for scaling are selected from the above-mentioned feature points. The number of cases may be larger than the above, depending on the combinations to be used. Furthermore, when searching for peaks in the concentration frequency distribution, it is conceivable that only large peaks are searched, only small peaks are searched, or both are searched. Therefore, the user alternately develops and experiments the program, searches for the characteristic points as described above, and completes the step of normalizing the fraction.

(2) Extraction processing using feature quantities of local density distribution refers to the image after performing the density normalization processing (0).
The identification target is extracted as a candidate area from the unique density or the two-dimensional distribution of the unique density. Normally, filtering involves calculating the feature amount of the density distribution within a certain local image, and storing that value as a new density at the center of the local image.

A method is used in which processing such as spatial product-sum calculation (C, hereinafter referred to as local density feature calculation processing) is performed on each pixel of the target image, and then binarized at a predetermined threshold level.

The features of the density distribution calculated in the above local image include the average density in the local image, the number of pixels with a certain density, and the above average after performing local difference processing in each horizontal, vertical, and diagonal direction. density or the number of pixels waiting for a specific density,
The number of the specific density pixels in the average density image, locally -fll! There are various maximum, minimum, and medium densities in one local image, such as maximum and minimum densities in one local image, and similarly, users alternately develop programs and experiments to find out what kind of feature is suitable for one local image. Then, the step of extracting candidate regions based on the characteristics of the concentration distribution is completed.

The final extraction process using geometric features in (3) is (2)
From the candidate regions obtained in ), only those whose unique geometrical features, such as area, perimeter, or C perimeter)2/area value, are within a predetermined range are selected. This step has been well researched and
Various practical methods have also been proposed, and their explanations will be omitted.

In other words, the program development work in steps α) and (Oi) takes one to two months, even for experts with extensive knowledge of programming and image processing.
When a general user with no knowledge of image processing created an application program through trial and error, it was difficult to understand the inspection line of the target product.

In many cases, the application program is no longer needed.

Hereinafter, the steps α) and (2) will be collectively referred to as region division based on the characteristics of the density distribution.

[Purpose of the invention]

An object of the present invention is to suitably perform region segmentation in which regions to be identified are extracted based on the characteristics of density distribution, by simply capturing an input image containing an object separated by Km a number of times λ. The objective is to easily determine various conditions.

[Summary of the invention]

The present invention provides an image processing apparatus for extracting an identification target from the input image, which includes an area dividing means for extracting an area in which density distribution characteristics satisfy a predetermined condition from an input image consisting of multiple gradations. A region exemplifying means for specifying an identification object and a background so as to include the identification object is provided, and a predetermined condition for the area dividing means is determined from characteristics of the identification object and the background in the illustrated example region. This is what I did.

[Embodiments of the invention]

An embodiment of the present invention will be described below with reference to the drawings.

First, FIG. 1 shows the basic concept of the present invention.

As the image processing application in FIG. 1, an example of inspecting tablets for medicine or the like for scratches is used.

Conventionally, for an input image 11 of a multi-gradation object C (flaw on a tablet) to be identified, the user (C operator) determines that the characteristic point to be searched for in the density distribution normalization process is the peak showing the maximum frequency. There are various ways to do this, or to use local average processing for local density feature calculations.

The parameter is set after being predicted based on the operator's knowledge, experience, or the operating manual.

After creating a program in which the set parameters are reflected in the area dividing means 12, the operator executes the program and evaluates the binary image 13 obtained as a result. Block 14 in Figure 1) was repeated through trial and error until the result was obtained.

This type of work has the problem of being very time consuming and costly, and of creating ambiguity in the algorithm due to qualitative evaluation by the operator.

Therefore, in order to solve the above-mentioned problems, the present invention involves the operator setting conditions and making a trial as shown in block 14.
The step of evaluation is deleted, and a region exemplifying means 15 is newly provided for specifying a region including the identification target (scratches on a pill) in the input image, and the region exemplifying means 15 is used to specify the area exemplified by the region illustrating means 15 from the density distribution of the entire input image. By limiting the amount of image information to only the density distribution of the background of the identified target, we attempted to set conditions for the area dividing means 12 to separate the identified target from the background.

The purpose is to automate the process using the evaluation-based condition determination means 16.

The area dividing means 120 condition settings based on the characteristics of the concentration distribution are as follows:
It would take an enormous amount of time, several weeks, or even months, for the operator to try and evaluate each program using programming.

At this point, an area example means 15 is provided, and based on the relationship between the identification object and the background illustrated therein, the conditions under which the identification object and the background are best separated are determined according to a predetermined evaluation scale from among those prepared in advance. Originally, a computer was used to perform calculations and make decisions.

Next, the overall configuration of the image processing apparatus according to this embodiment will be explained using FIG. 2.

This embodiment includes a television camera 21 that takes an image of an identification target, a signal converter 22 that converts an analog signal of the television camera 21 into a digital signal, or conversely, a digital signal into an analog signal, and a multi-level input of the identification target. A gradation and binary image memory 23 that stores tone images, binary images, etc., and an area where the characteristics of the density distribution satisfy a predetermined condition from the input multi-gradation image stored in the gradation and binary image memory 23. an area dividing means 24 for extracting the image and storing it again as a binary image in the grayscale and binary image memory 23; Optimal conditions are determined by testing and evaluating various conditions necessary for the density frequency distribution normalization processing of the illustrating means 25 and the region dividing means 24 by assuming a plurality of the above conditions for a plurality of input images. The various conditions necessary for the local density feature calculation processing and final binarization of the density distribution normalization evaluation unit 26 and the area dividing unit 24 are tested by assuming a plurality of the above conditions for a plurality of input images. Consists of a local concentration feature evaluation section 27 that determines optimal conditions through evaluation, an example management section 28 that mainly performs overall management, a console monitor 29 that displays middle-class information, and a keyboard 30 that allows the operator to input data. has been done.

Incidentally, the condition determining means 16 based on trial and evaluation in FIG. The exemplary means 25 (15 in FIG. 1) is generally composed of a coordinate input medium such as a CRT, a tablet, or a mouse, but a combination of a digitizer, a trough ball, a light pen, and more recently developed devices are also available. A man-machine interface such as an input-integrated flat display as a coordinate input medium may be used.Also, coordinate values may be directly specified from a keyboard.In short, any means that can exemplify the area to be identified may be used. do not have.

Before explaining the overall operation, the internal structure of the region dividing means 24 and the functions and operations of each part thereof will be explained using FIGS. 3 and 4.

As shown in FIG.
The example management unit 2 uses information such as input images in the value image memory 23.
8, the region to be identified is extracted. Describing the internal configuration in detail below, the density level of a characteristic part of the density frequency distribution (for example, the peak indicating the maximum frequency) used for the density frequency distribution normalization process is stored in the gradation/binary image memory 23. A code for a feature location to be calculated by the density frequency distribution feature location extraction unit 31 and the density frequency distribution feature location extraction unit 31 obtained from the input image and the reference image,
and a normalization parameter storage unit 32 that stores the feature parts extracted by the density frequency distribution feature part extraction unit 31 from the reference image according to the code of the feature part mentioned above, of the input image that matches the feature locations found in the reference image stored in the normalization parameter storage unit 32! A density frequency distribution normalization processing unit 33 that performs frequency distribution normalization processing once, and a local density feature calculation that performs local density feature calculation processing on a new image output by the density frequency distribution normalization processing unit 33. a processing unit 34; a feature calculation parameter storage unit 35 that stores a code for local density feature calculation (for example, an average density code) to be executed by the local density feature calculation processing unit 34;
, a binarization processing unit 36 that performs predetermined binarization processing on the new image output by the local density feature calculation processing unit 34;
A binarization parameter storage unit 37 that stores the binarization code to be executed by the binarization processing unit 36 and its threshold value.
It is composed of

Next, the operation of the area dividing means 24 will be explained using FIG. 4.

First, in step 41, the necessary conditions for the area division means are set, namely, 1 in the normalization parameter storage section 32, a code of the feature location to be extracted, a code for the local @ degree feature operation in the feature calculation parameter storage section 35, and a binary value. The binarization code and threshold value are input into the parameter storage unit 37, and the reference a! After storing a reference image serving as a reference for density 5rJt distribution normalization processing in the i* recording section 231 and storing a target image in the input image storage section 232, the area dividing means 24 is activated.

　　.

Therefore, in step 42, the density frequency distribution feature extraction unit 31 first obtains the density level (pts9m) for the reference image in the reference image storage unit 2310, and the code (pts9m) of the feature location stored in the normalization parameter storage unit 32 ( For example, a feature point is extracted according to a code indicating that two peaks with large frequency values are supplied as feature points among peaks found in a relatively large range, and its concentration level (pt, ps)'fI: The code of the feature location is stored in the normalization parameter storage unit 32, and the feature location is similarly extracted from the input image in the input image storage unit 232, and its density level (pt', pz') is calculated based on the density frequency distribution normal. The data is transferred to the conversion processing unit 33.

Next, in step 43, the concentration frequency distribution is normalized.

The normalization processing unit 33 performs density normalization processing to match the feature parts of the input image in the input image storage unit 232 with the feature parts of the reference image, and stores the resulting new image in the density/binary image memory 23. The image data is stored in the normalized image storage unit 233 of . The density conversion performed by the density frequency distribution normalization processing unit 33 is performed, for example, by converting the density level of a characteristic point in the reference image into pz
e p2, and the characteristic parts of the input image are 91' + 9
When the density level p' of the input image is linearly converted to a new density level p according to the conversion straight line expressed as (where N is the maximum value of the density gradation), one feature point is defined as z'. For pl and p1', ri = 1)' + (
pt-1)t').

Next, in step 44, the local density feature calculation processing unit 3
4 is sent to the normalized image storage unit 233, and the local density feature calculation code stored in the feature calculation parameter register 1 unit 35 [for example, the minimum density in a 3×3 local image is stored in the center image. A predetermined local density feature calculation is performed according to the operation St-code indicating a filtering process that is performed over all pixels, and the result is
! [Stored in the local feature image storage unit 234 in the image memory 23.

Next, in step 45, the binarization processing unit 36 applies the binarization code C stored in the binarization parameter storage unit 37 to one local feature image storage unit 2340 image, for example, if it is lower than a certain threshold value T. is set to 1 - other values are set to φ)) and the threshold value is used to perform binarization, and the result is stored in the binary image storage unit 2 in the grayscale/binary image memory 23.
35. The above is the operation of the area dividing means 24.

Next, the operation of the entire image processing apparatus in this embodiment shown in FIG. 2 will be explained using FIGS. 3 to 6.

As shown in FIG. 5, the operation of the image processing apparatus of this embodiment is largely comprised of determining conditions for density frequency distribution normalization processing shown in steps 51 to 54 and local density feature calculation shown in steps 55 to 57. The process is divided into two cycles: and determining conditions for binarization processing.

In the former condition determination cycle for density frequency distribution normalization processing, first, in step 51, the operator takes an image of the identification target to be used as a reference image on the television camera 21 according to a message displayed on the console monitor 29, and then inputs the image from the keyboard 30. Specify the reference image to be captured. Then, under the management of the example management section 28, the above-mentioned identification target is the signal converter 2.
2, the image is stored as a multi-tone image in the reference image storage section 231 inside the grayscale/binary image memory 23.

Next, in accordance with the message displayed on the console monitor 29, the operator uses the area illustrating means 25 to view the identification target displayed on the area illustrating means 25.
As shown in FIG. 6, the area to be identified and the background area including the area are illustrated by stroking the screen of the area illustrating means 25 with an input medium.

An example method for the operator is shown in FIG. 6 (a) screen 61 and (b) Ii! Two types of FC are prepared as shown in the i-side 62, and the method shown in the screen 61 of (a) is suitable for illustrating the identification object 63 and the background 64 when the identification object 63 and the background 64 are sufficiently large and simple. The enclosed area is illustrated as shown in FIG. On the other hand, the method shown in the screen 62 of (b) is such that when the identified object 65 or the background 66 is small or complex, the area is filled in as an example. It goes without saying that if the characteristics of the density distribution of the object to be identified and the background are extremely uniform, it is sufficient to exemplify only a part of each region.

When the operator finishes illustrating the region, the example management section 28 stores the example regions of the identification target and the background exemplified by the region exemplification means 25, and then sends them to the density distribution normalization evaluation section 2.
6, the concentration frequency distribution per unit area in each of the above regions is calculated and stored.

Here, the above concentration frequency distribution per unit area refers to a concentration frequency distribution obtained by multiplying each frequency value of the concentration frequency distribution by the reciprocal of the area of the region.

The density frequency distribution per unit area of the identification target and background in this reference image will be used later for evaluation of various normalization conditions.

In the next step 52, the operator adjusts the illumination illuminated on the identification target, the contrast between the identification target and the background, etc. according to the message displayed on the console monitor 29.
The object to be identified is imaged again after being changed to the extent expected during actual identification, and a command to capture the image is input using the keyboard 30.

Then, under the management of the example management section 28, the image to be identified is stored in the input image storage section 232 inside the gray scale/binary image memory 23.

Next, in accordance with the message displayed on the % console monitor 29, the operator uses the area illustrating means 25 to exemplify the identification target and background in the same manner as in step 51, and the illustrative result C) is the concentration frequency distribution per unit area. is newly stored in the concentration distribution normalization evaluation unit 26.

Next, in step 53, the concentration distribution normalization evaluation unit 26
However, the currently optimal concentration frequency distribution normalization condition, that is, the optimal feature location code is derived and stored in the normalization parameter storage section 32 of the region dividing means 24. Next, the example management section 28 activates the area dividing means 24 as is, and then displays the normalized image storage section 233 inside the gray scale/binary image memory 23 on the area example means 25.

Next, in step 54, the operator decides whether or not to continue the process in accordance with the message displayed on the console monitor 29, and indicates his intention on the keyboard 30. Here, if the key human power continues, the process returns to step 52, and if not, the condition determining cycle for the density frequency distribution normalization process is completed and the process proceeds to the next step.

Here, the internal configuration and operation of the concentration distribution normalization evaluation section 26 will be explained using FIGS. 7 and 8.

As shown in FIG. 7, the density distribution normalization evaluation section 26 uses the image storage section 2 in the shading/binary image memory 23 in the example regions of the identification target and the background stored in the example management section 28.
31 or the input image storage unit 2320, the density frequency distribution calculation unit 71 that calculates the density frequency distribution per unit area, and the reference image target density distribution that stores the density frequency distribution per unit area of the area to be identified in the reference image storage unit 231. A storage unit 72, a reference image background density distribution storage unit 73 that stores the density frequency distribution per unit area of the background area in the reference image storage unit 231, a reference image background density distribution storage unit 73 that stores the density frequency distribution per unit area of the background region, which is normalized and obtained using a large number of feature location codes prepared in advance. A normalized image target density distribution storage unit 74 stores the density frequency distribution per unit area of the area to be identified in each of the normalized image storage units 233 and 233. Normalized image storage unit 23
3. Normalized image background density distribution storage unit 75 for storing the density frequency distribution per unit area of the background area in each of K.
Shift amount calculation for calculating the shift amount of the normalized image target density distribution storage unit 74 and the normalized image background density distribution storage unit 75 with respect to the reference image target density distribution storage unit 72 and the reference image background density distribution storage unit 73; The deviation amount calculation unit 76 detects the minimum value of the deviation amounts stored in the deviation amount storage unit 77 that stores the deviation amount corresponding to each feature point code calculated. and a deviation amount comparison processing unit 7 that stores the feature point code corresponding to the minimum value in the normalization parameter storage unit 32 of the area dividing means 24.
It is made up of 8 parts.

The operation will be described below with reference to FIG. 8 as well.

First, when activated at step 51 in FIG. The density frequency distribution per unit area of the storage unit 231 is calculated, and the results are stored in the reference image target density distribution storage unit 72 and the reference image background density distribution storage unit 73, respectively.

8r if activated in step 54 in FIG.
As shown in gJ, first, in step 8, the example management section 28 selects one of the feature location codes prepared in advance and stores it in the normalization parameter storage section 32 of the area division means 24.

After setting the normalization parameters, that is, step 81, the process moves to step 82.

Next, in step 82, after the example management section 28 activates the area division means 24, the concentration frequency distribution calculation section 71 of the concentration distribution normalization evaluation section 26 calculates the identification target and the number stored in the example management section 2.8. The density frequency distribution per unit area of the normalized image storage unit 233 in the gradation/binary image memory 23 in the example area of the background is calculated, and the normalized IMgI! corresponding to the feature location code is calculated. It is stored in the target density distribution storage unit 74 and the normalized image background density distribution storage unit 75.

Next, in step 83, the normalized image background density distribution storage unit 74, the normalized image background density distribution storage unit 75, the reference image target density distribution storage unit 72, and the reference image background density corresponding to the currently bundled feature location codes are stored. The amount of deviation from the distribution storage unit 73 is calculated and stored in the deviation amount storage unit 77 as a new amount of deviation this time in addition to the amount of deviation 77 determined up to the previous example.

After performing the above processing for all feature location codes under the management of the example management unit 28, in step 84,
The deviation amount comparison processing section 7B detects the minimum value of the deviation amount 77, and stores the feature location code corresponding to the minimum value in the normalization parameter storage section 32 of the area dividing means 24.

The above is the operation of the concentration distribution normalization evaluation section 26.

Returning now to FIG. 5, the operation of the condition determination cycle for local density feature calculation and binarization processing will be explained.

In Mastela 7'551C, console monitor 29
In accordance with the message displayed in step 11C, the operator photographs the object to be identified in the same way as in step 52, captures the image,
Again, the regions are illustrated in the same way as in step 51.

Next, in step 56, the local density feature evaluation unit 27 derives the currently optimal conditions for local density feature calculation processing, that is, the optimal local density feature code, binarization code, and threshold value, and calculates the characteristics of the region dividing means 24. The calculation parameter storage section 35 and the binarization parameter storage section 374C each store the parameters. Next, the example management unit 28
4 is activated, the binary image storage section 235 inside the gradation/binary image size memory 23 is displayed on the area illustrating means 25.

Next, in step 57, the operator decides whether or not to continue the process based on the message displayed on the console monitor, and indicates his/her intention using the keyboard 30. If the key force continues here, return to step 55 again,
If not, the condition determination cycle for local density feature calculation and binarization processing ends.

Here, the internal configuration and operation of the local density feature evaluation section 27 will be explained using FIGS. 9 and 10.

The local density feature evaluation unit 27 calculates unit areas of the identification target and background regions for each local feature image storage unit 234 obtained by performing local density feature calculations using a large number of local density feature codes prepared in advance. A feature image density frequency distribution calculation unit 91 that calculates the density frequency distribution of the hit density frequency distribution calculates the density frequency distribution of the unit area of the identification target and the background region calculated by the feature image density frequency distribution calculation unit 91, and calculates each local density. A feature image target density frequency distribution storage unit 92% and a feature image background density frequency distribution storage unit 93% are stored in correspondence with the feature code. , a target distribution cumulative calculation unit 94 that accumulates each example, a background distribution cumulative calculation unit 95, and a target distribution cumulative calculation unit 9.
4 and a feature image object cumulative density frequency distribution storage unit 96 and a feature image background cumulative density frequency distribution storage unit 97% that store the density frequency distribution corresponding to each local feature code accumulated by the background distribution cumulative calculation unit 95. Cumulative density frequency distribution storage unit 96 and feature image background cumulative density frequency distribution storage unit 9
7, the optimum local concentration feature code, binarization code, and threshold are derived and stored in the feature calculation parameter storage unit 35 and the binarization parameter storage unit 37 of the region dividing means 24. Feature calculation/binarization parameter determination It is composed of N98.

Next, the operation will be explained using FIG. 1θ.

First, in step 101, the example management unit 28 stores one of the many prepared local density feature calculation codes in the feature calculation parameter storage unit 32 of the region dividing means 24.

Next, in step 102, after the example management unit 28 activates the area division means 24, the feature image density frequency distribution calculation unit 91 of the local density feature evaluation unit 27 calculates the identification target and background stored in the example management unit 28. Calculates the density frequency distribution per unit area of the local feature image storage unit 234 in the gradation/binary image memory 23 in the exemplary region, and calculates the density frequency distribution per unit area of the local feature image storage unit 234 in the gradation/binary image memory 23, and stores the density frequency distribution storage unit 92 of the feature image object corresponding to each local density feature calculation code. and stored in the characteristic image background density frequency distribution storage unit 93.

Next, in step 103, the feature image target density frequency distribution storage unit 92 and the feature image background density frequency distribution storage unit 93 corresponding to the local density feature calculation code obtained above are activated by the target distribution cumulative calculation unit 94 and the background distribution, respectively. The cumulative calculation unit 95 adds and accumulates the contents of the feature image object cumulative density frequency distribution storage unit 96 and the feature image background cumulative density frequency distribution storage unit 97.

After repeating the above processing for all local density feature calculation codes, in step 104, the optimum local density feature calculation is performed from the feature image target cumulative density frequency distribution storage unit 96 and the feature image background cumulative density frequency distribution storage unit 97. The code, the binarization code, and its threshold value are derived and stored in the feature calculation parameter storage unit 35 and the binarization parameter storage unit 37 of the area dividing means 24.

Next, the internal configuration and operation of the feature calculation/binarization parameter determination section 98 will be explained using FIGS. 11 to 14.

As shown in FIG. 11, the feature calculation/binarization parameter determination unit 98 stores data for the feature image target cumulative density frequency distribution storage unit 96 and the feature image background cumulative density frequency distribution storage unit 97 corresponding to each local density feature calculation code. A binarization threshold calculation unit 111 that calculates a binarization threshold (note that a plurality of binarization threshold calculation units with different threshold calculation methods are prepared). Binarization threshold calculation unit 11
1'.

111"1, the binarization threshold storage unit 112 that stores the threshold value corresponding to each local concentration feature operation code calculated by the binarization threshold calculation unit 111 (in addition, the binarization threshold calculation unit 2 pieces each corresponding to 1 piece 1' and 111''
(there are valorization threshold storage units 112', 112''),
Evaluation of z-value conversion is performed from the binarization threshold 112 corresponding to each local density feature operation code and the corresponding feature image target cumulative density frequency distribution storage unit 96 and feature image background cumulative density frequency distribution storage unit 97. a binarized evaluation value calculation unit 113 that calculates the value, and a binarized evaluation value storage unit that stores the binarized evaluation value corresponding to each local density feature operation code calculated by the tXz digitized liver value calculation unit 113. 114C, there are also binarization evaluation value storage units 114', 114'' corresponding to each of the binarization threshold calculation units 111', 111''), and the binarization evaluation value storage units 114, 114. ', 11
4" is detected, and the code of the binarization threshold calculation unit corresponding to the maximum value, the threshold value, and the local density characteristic calculation code are stored in a predetermined parameter storage unit of the area dividing means 24. It also includes an evaluation value comparison calculation 115.

Next, the operation will be explained using FIG. 12.

First, in step 121, each binarization threshold calculation unit 111° 111', 11
1'' calculates thresholds using different methods from the feature image object cumulative density frequency distribution storage unit 96 and the feature image background cumulative density frequency distribution storage unit 97, respectively, and binarization threshold storage units 112, 112', 112".

Binarization threshold calculation units 111, 111'.

111" is, for example, shown in Fig. 13 (a), (b), (
A binarization threshold is calculated by the method shown in c).

In (a), the average value and deviation value of the cumulative density frequency distribution storage unit 96.97 of the identification target and the background are Cμt and σ, respectively.
t), (μb, σ-), and a threshold value T is set at the center of the distribution.

That is, Binarization is performed using a threshold value T shown by .

On the other hand, (b) shows the cumulative concentration and gradient distribution storage unit 9 of the identification target.
The average value and deviation value of the part of the cumulative concentration frequency distribution of the background (Cμbi, σkl) which is to the left of the average value μt of 6 (smaller than the C mean value μt), and the cumulative concentration frequency distribution of the background which is to the right of the above μ The average value and deviation value of the part are rμ
hfleσh's).

Center TL and C of Cμt, σt) and (μkl, σbt)
The central T wealth of μm reference σ own) and (μb3 σbj11) is found, and the value between %Tl and T! is binarized.

On the other hand, (C) is the background distribution Cμb, contrary to (b).

The distribution of objects (μ, 1°σ) to the left of the mean value of σb)
11) and the distribution on the right [μm2. σl), and in the same manner, values T1 and T1 are found, and the density smaller than %Tl and the density larger than T2 are binarized.

Next, in Stella 77122, the binarization evaluation value calculation unit 1
13 calculates the evaluation value of each binarization threshold value 112, 112', 112'' for each local concentration feature operation code 9, and stores it in a binarization evaluation value storage unit 114°114', 114.
“Stored in.

As shown in FIG. 14(a), the amorous value obtained here is the portion that is binarized as an identification target when the feature image object cumulative density frequency distribution storage unit 96 is divided by the obtained threshold value. Integral value St (shaded area) and integral value Sb (shaded area) of the portion that is not binarized when the feature image background cumulative density frequency distribution storage unit 97 is divided by the determined threshold value.
expressed as the sum of

It shows that the larger this total value is, the better the #7 & other object and background can be binarized.

In addition, in FIG. 14 (B), the target area according to one example of the operator is 141, the background area is the area surrounded by 142 outside of the target area, and 2
When the binary pattern obtained by value conversion is set to 143,
The ratio At/person 1 of the area At of the binary pattern 143 included in the area 141 and the area At of the area 141 is calculated, and the area 14
Area Ah of the area outside 1 and surrounded by 142
, the binary pattern 1430 area A included in the area
The value obtained by subtracting %Ah-At and AhO ratio (Ah-person I
:) / A Find b, and combine this with the above A '* / A
The value of adding t, As/A*+Ou At )/Ah
The sum of the values for all examples is equal to S*+S b in (a) above.

Next, returning to FIG. 12, in step 123, the evaluation value comparison calculation section 115 converts the binarized favorable values 114, 114' corresponding to each local density feature calculation code.

114'' is detected, and the local density feature calculation code corresponding to the maximum value, the code corresponding to the binarization threshold calculation section, and the threshold are output to the region dividing means 24.

An embodiment of the present invention has been described above, but in short, in performing region division, a region illustrating means is provided.
This is to make it easier for the operator C (user) to set the condition 1 in which the object and background are well separated by limiting the information. be.

Therefore, the present embodiment satisfies the above conditions, that is, the feature location code and local density feature code for normalization.

Binarization codes were explained in Section 1 only when selecting from among those prepared in advance to determine the area exemplification means, but it is also possible to assign priorities to the various codes mentioned above and allow the user to calculate from the code with the highest priority. You may adopt any method.

〔Effect of the invention〕

According to the present invention, the amount of information in the input image is limited to the identification target and the background other than the identification target by using the region exemplification means. Various necessary conditions can be easily set, and algorithms and programs can be developed not only by experts but also by film-forming users in a short period of time.

[Brief explanation of the drawing]

Fig. fs1 is a conceptual diagram of the present invention, Fig. 2 is an overall configuration diagram of an implementation column of the present invention, Fig. 3 is an internal configuration diagram of the area dividing means, and Fig. 4
FIG. 5 is an explanatory diagram of the operation of the region dividing means, FIG. 5 is an explanatory diagram of the operation of the entire embodiment, FIGS. 6(a) and (6) are explanatory diagrams of the region exemplification method, and FIG. 7 is the density distribution normalization evaluation unit. 8 is an explanatory diagram of the operation of the concentration distribution normalization evaluation section. FIG. 9 is an internal configuration diagram of the local concentration feature evaluation section. FIG. 10 is an explanatory diagram of the operation of the local concentration feature evaluation section. The figure is an internal configuration diagram of the feature calculation/binarization parameter determination section, FIG. 12 is an explanatory diagram of the operation of the feature calculation/binarization parameter determination section, FIG. 13 is an explanatory diagram of the binarization threshold r value calculation method, and FIG. 14 FIG. 2 is an explanatory diagram of a method for calculating a binarization threshold evaluation value.

Claims (1)

[Claims] 1. Region dividing means for extracting regions whose density distribution characteristics satisfy predetermined conditions from an input image consisting of multiple gradations, and using the region dividing means to extract a region to be identified from the input image. In an image processing apparatus for extracting an image, a region exemplifying means is provided for specifying a region including an identification target from the input image, and a predetermined region dividing means is used based on the characteristics of the density distribution of the identification target and the background in the exemplified region. An image processing device characterized by determining conditions. 2. An image processing apparatus characterized in that the area illustrating means described in claim 1 is constructed of a coordinate input medium. 3. The area dividing means described in claim 1 is a density divider that performs processing to make the density value of the input image match the density value of a characteristic part in the density frequency distribution of the reference image set in advance. a frequency distribution normalization processing unit, and a local density feature calculation process that stores the feature amount of the density distribution in a local image centered on each pixel of the normalized image as a new density value for each pixel. An image processing device comprising: a binarization processing unit that performs binarization using a threshold value set in advance; 4. The density values of characteristic locations in the density frequency distribution used in the density frequency distribution normalization processing unit described in claim 3 are based on the density values of the identification target and the background within the example area exemplified for the reference image. Concentration frequency distribution per unit area,
For multiple images obtained by performing normalization processing on multiple input images, identify the sum of deviations from the density frequency distribution field per unit area of the identification target and background within each example region. An image processing apparatus characterized in that the density value is calculated for each of the object and the background, and uses a density value at a location where the sum of the amount of deviation of the identification object and the sum of the amount of deviation of the background is the minimum. 5. The feature amount of the density distribution in the local image used by the local density feature calculation processing unit described in claim 3 is obtained by performing local feature calculation processing on a plurality of input images. For a plurality of binary images obtained by performing binary processing using a preset threshold, calculate the area of each binary image included in the classification target within the example region and the area of each of the classification targets. Calculate the sum of the ratios to the area, and calculate the ratio of each area obtained by subtracting the area of each binary image included in the background from the area of the background in each example region to the area of each background. An image processing apparatus characterized in that a sum is calculated, and a feature amount is used that maximizes a value obtained by adding the sum of the area ratios and the sum of the area ratios related to the identification target. 6. The threshold value used in the binarization processing unit described in claim 3 is obtained by preparing a plurality of threshold calculation means in advance and performing local density feature calculation processing on a plurality of input images. For each image obtained, the density frequency distribution per unit area of the identification target and background in the example region is determined, and the total density frequency is calculated by summing the frequency values for the same component of each density frequency distribution. The distribution is calculated for each of the identification target and the background, and the discrimination is performed in the concentration range divided by the threshold value from among the threshold values for binarizing the identification target calculated by the plurality of threshold value calculation means. An image processing apparatus that selects a threshold value that maximizes a value obtained by adding a frequency sum of a total density frequency distribution of a target and a frequency sum of a total density frequency distribution of the background outside the density range.