JPH07287753A - Article discrimination system - Google Patents

Abstract

PURPOSE:To surely discriminate the brand of an article such as a tablet by extracting the feature of each article without requiring human visual observation. CONSTITUTION:The operation of this system is divided into a learning phase for the preparation of a dictionary and a discrimination phase for executing discrimination in the learning phase, many tablet samples of each brand are photographed by a camera 1 and a learning part 106 prepares a dictionary 108 by using the photographed picture. In the discrimination phase, a tablet to be discriminated is photographed by the camera 1 and a discrimination part 107 discriminates the brand of the tablet by using its photographed picture and the dictionary 108. A gray level value vector extracted from an actually photographed picture is used as a feature vector of an article. RGB/gray picture conversion using inverse gamma correction and a G component out of RGB components is applied to the actually photographed picture as preliminary processing so that the gray level feature of the article is effectively reflected on the gray level value vector. A gray picture is divided into small sections and the gray level center value of each small section is sampled to prepare a gray level value vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an article identification system for identifying articles using image processing / identification technology,
In particular, the present invention relates to an article identification system suitable for identifying a brand of an inputted drug tablet or capsule (hereinafter, simply referred to as a tablet).

[0002]

2. Description of the Related Art Currently, about 6000 kinds of tablets are sold in Japan, and in general hospitals handling a lot of tablets, patients are requesting discrimination of the above tablets. That is caused by the anxiety that "Is it okay to take the medicine I received at the hospital I used to go and the medicine I am receiving at the same time?"

In response to this identification request, the pharmacist now reads the identification code written on the surface of the tablet to be identified, identifies the brand with reference to the color and shape of the tablet, and identifies the identified brand. It searches the drug collections based on the original information and provides the patients with information on their efficacy and side effects.

[0004] In a large hospital, this discrimination work occurs as many as 50 times a day, and visual identification of the tablet makes a very heavy burden on the pharmacist. The problem of tablet brand identification is discussed, for example, in “Kitazawa Shikibun: Present Situation and Problems of Drug Identification, Monthly Yakuji Vol. 31, pp. 23-25, 1989”.

Therefore, an interactive simple computer system aimed at reducing the burden on the pharmacist in the discrimination work has been experimentally operated. In these, the pharmacist visually extracts the characteristics of the tablet, searches for candidate tablet brands from the tablet brands registered in the system in advance using the characteristics as keywords, and displays the search results on the display unit. Is configured to display. The details of this computer system are discussed, for example, in "Masayuki Ishijima; A Study on Identification and Selection by Linguistic Expression, Proceedings of the 11th Joint Conference on Medical Informatics, pp. 587-588, 1991".

By using these computer systems, the burden on the pharmacist in the discrimination work can be reduced and the quality of medical treatment can be improved.

[0007]

In the computer system for the above discrimination work, it is necessary for the pharmacist to visually extract the characteristics of the tablet. However, in recent years, from the viewpoint of providing tablets that are easy for humans to take, overwhelmingly many tablets have engraved circles, white colors, and identification codes. There are also tablets without identification code.

[0008] For these reasons, it is difficult to visually extract features.

Therefore, an object of the present invention is to enable feature extraction of tablets without human eyes,
An object of the present invention is to provide an article identification system capable of surely identifying tablet brands.

Further, the more general object of the present invention is to
The object of the present invention is to make it possible to carry out the identification inspection of the article automatically and accurately by using the image processing / identification technology.

[0011]

An article identification system according to the present invention has, as a basic configuration, image input means for inputting an image of a sample article and an image of an article to be identified according to each category, and from this input image, Feature extraction means for extracting a gray value vector, learning means for generating a reference vector for each category based on the gray value vector extracted from the input image of the sample article, a dictionary file for accumulating the reference vector, and an identification target An identification unit that identifies the category of the identification target article by referring to the reference vector in the dictionary file based on the gray value vector extracted from the input image of the article.

The system according to the first aspect of the present invention further comprises an inverse gamma correction means for performing an inverse gamma correction for removing the influence of the gamma correction on the input image, in addition to the above basic structure. The image after the inverse γ correction is input to the feature extraction means.

Further, in the system according to the second aspect of the present invention, in the above basic structure, the image input means is an R of the article.
It further comprises RGB gray conversion means for inputting a GB color image and converting this RGB color image into a gray image by preferentially using the G (green) component thereof, and the gray image from this conversion means is feature extraction means. It is designed to be input to.

Further, in the system according to the third aspect of the present invention, in the above basic configuration, the feature extracting means divides the predetermined area including the article area from the input image into small sections, and the grayscale value of the pixel in each small section. The median value of is sampled, and the gradation value vector having the median gradation value of the sampled small section as an element is extracted.

In a preferred embodiment of the present invention, the learning means comprises:
Principal component analysis is performed on the grayscale value vector of the learning sample to obtain eigenvalues and eigenvectors, the grayscale value vector is converted to a significant principal component vector using these eigenvalues and eigenvectors, and this principal component vector is used as the reference vector. It is like this. Further, the identifying means converts the gray value vector of the article to be identified into a principal component vector using the eigenvector, and refers to the reference vector based on the principal component vector for identification.

[0016]

In the article identification system of the present invention, the images of the sample articles of each category are input to the system when creating the dictionary, and the images of the identification target articles are input to the system when identifying the dictionary. The feature vector of the article is automatically extracted, and based on this feature vector, the reference vector is created when the dictionary is created, and the identification target article is identified when the identification is performed. As a result, it becomes possible to automatically carry out the identification inspection of the article without relying on the visual inspection of a person.

Here, as the feature vector, a gray value vector having the gray values of the pixels of the input image as elements is used.
The identification using the gray value vector is nothing but the calculation of the degree of coincidence by superimposing the reference image and the identification target image. Such an identification method is effective for the identification of large categories (there are many categories) because the image information can be utilized to the maximum extent.

In the system according to the first aspect of the present invention,
Before the grayscale value vector is extracted from the input image, the inverse γ correction is performed on the input image to remove the γ correction inherent in the input image. As a result, the density change and contrast in the density area where the density change and contrast were suppressed more than the true product image by γ correction are closer to the true product image, so the gray value vector reflecting the true product image is extracted. Is facilitated, and as a result, the identification accuracy is improved.

In a system according to the second aspect of the invention,
An RGB color image is input as an article image, and the G component is preferentially used to convert this color image into a gray image having only grayscale values. In image input means such as a TV camera, the G component is generally another R (red) component and B (blue) component.
Since the dynamic range is wider than that of the components, it is possible to well reflect a subtle difference in density of the article image, particularly a difference in density in a low density region. Therefore, by preferentially using the G component to create a gray image, it is possible to obtain a gray value vector that well represents a subtle difference in density of the article, and as a result, the identification accuracy is improved.

In a system according to the third aspect of the invention,
When obtaining the gray value vector, the area containing the article of the input image is subdivided into small sections, and the median gray value of the pixels in each small section is sampled. Then, a gradation value vector is created with the sampled gradation median value as an element. The median value of each small section reflects the light and shade characteristics of each small section better than the average value. For example, in a small section divided into a white area and a black area, the average value is gray, but when the median is taken, the larger white area is white, and the larger black area is black. The characteristics of the compartment are well represented. Therefore, by using the gray value vector having the median value of the small section as an element, the identification accuracy is improved.

In the preferred embodiment, a principal component analysis is performed on the gray value vector extracted from the image, and the gray value vector is converted to a significant principal component vector based on the result.
Identification is performed using this principal component vector. This allows
Since the identification can be performed without using the feature that does not contribute to the improvement of the identification accuracy or the feature of which the contribution degree is small, the identification processing can be speeded up while maintaining the high identification accuracy.

[0022]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 shows a functional configuration of a tablet identification system according to an embodiment of the present invention.

This tablet identification system operates on a computer 100 equipped with the NTSC TV camera 1 as an external device. The computer 100 includes an image input unit 10
1, inverse γ correction unit 102, RGB gray conversion unit 103, position / orientation correction unit 104, feature extraction unit 105, learning unit 10
6, the identification unit 107 and the dictionary file 108.

The operation of this system can be roughly divided into a learning phase and an identification phase.

The functions and operations of the respective parts will be described below separately for (A) learning phase and (B) identification phase.

(A) Learning Phase The system must first go through a learning phase as initialization or preparation before starting its operation. In the learning phase, a large number of tablet samples (hereinafter referred to as learning samples) are input for each brand of tablets to be identified, the reference vector of each brand is calculated based on this, and stored as the dictionary file 108. It

First, the image input unit 101 inputs a color analog image obtained by photographing a tablet placed on a plain black background with the NTSC television camera 1, and converts it into an RGB digital image.

Next, the inverse γ correction unit 102 removes the influence of the γ correction applied to each of the RGB planes on the RGB digital image so that the input / output characteristics of the camera 1 become substantially linear. Inverse γ correction is performed. At this time, for example, 2.2 is used as the value of γ. This is because the NTSC television camera 1 is usually configured to perform γ correction of γ = 2.2 on a captured image. By performing the inverse γ correction, the contrast between the tablet area and the background area in the image is increased, and the subsequent processing can be effectively performed.

Next, the RGB gray conversion unit 103 causes the RG
Using the grayscale value of the G component of the B component as it is, RGB
Converts the image to a gray image with only grayscale components. Where G
The reason for using a component is that G
This is because the components have a wider dynamic range than the other two components, and therefore it is considered that the subtle difference in white can be reflected in the light and shade in a tablet with an overwhelmingly large amount of white. By converting an RGB image into a gray image, the following processing amount can be reduced to 1/3.

Next, the position / orientation correction unit 104 performs processing such that the position and orientation of the tablet in the image are always the same position and orientation for the same brand. FIG. 2 shows a detailed processing flow of this position / orientation correction.

As shown in FIG. 2, first, the gray image is binarized by using a predetermined threshold value,
Only the tablet area is extracted (step 201). At this time, for example, Otsu's method based on the discrimination criterion is used as a method of selecting the binarization threshold value. This method is a very effective method for an image that can be clearly divided into two areas, a background area and a tablet area. For details, see “Otsu Nobuyuki: Discrimination and Automatic Threshold Selection Method Based on Least Squares Criterion, IEICE Transactions vol. J63-DN.
o. 4, pp. 349-356, 1980 ".

Next, the center of gravity of the tablet area in the binarized image is obtained, and the center of gravity is moved to the center of the image (step 202). Subsequently, in the tablet area, the gray image is binarized using a threshold value different from the above threshold value to extract the area of the identification number code printed or stamped on the tablet surface (step 203). The method of Mr. Otsu based on the above-mentioned discrimination standard can also be used for the binarization threshold selection method at this time.

Next, the second moment of the extracted identification code area is obtained (step 204), and the entire image is rotated by affine transformation so that the direction in which the second moment becomes maximum coincides with the X axis of the image. (Step 20
5).

By correcting the position and direction of the tablet area in this manner, the identification accuracy described later is improved. This is because the characteristics of the external shape of the tablet depend on its position and direction.

Further, since the same tablet always has the same position and direction, it is not necessary to strictly adjust the position and orientation of the tablet with respect to the television camera 1 at the time of photographing.

Referring again to FIG. 1, next, the feature extraction unit 1
05 performs a process of extracting the feature vector of the tablet from the gray image whose position / direction has been corrected. FIG. 3 shows a detailed processing flow of this feature extraction. 4 to 6 show examples of images for explaining the feature extraction processing step by step.

First, as shown in FIG. 5, a rectangular area 500 of a predetermined size is cut out from the gray image 400 whose position and orientation are corrected as shown in FIG. 4 (FIG. 3, FIG. 3). Step 301). Here, the rectangular area 50
The size of 0 is determined by a statistical method based on the result of performing feature extraction on tablets of various brands in advance as the size that allows the most accurate feature extraction.

Next, as shown in FIG. 6, the rectangular area 5
00 is divided into small sections 600 of arbitrary size N × M (step 302). Here, the size N of the small section 600
× M is determined in proportion to the processing capacity of the computer 100.

Next, for each small section 600, a representative value is calculated from the gray values of all the pixels (step 303). The median is used as the representative value, for example. The median value is, even when the distribution of gray values is obviously biased, such as the tablet area and the background.
This is because the entire distribution can be well reflected.
Then, a grayscale value vector having the representative value of each small section 600 as an element is extracted as a feature vector (step 30).
4).

In the learning phase, a large number of learning samples are supplied to this system for various brands of tablets to be identified, and the above-described processing from photographing to feature extraction is performed on each learning sample.

Referring again to FIG. 1, the learning unit 106 performs principal component analysis on all gray value vectors extracted from a large number of learning samples for each brand to obtain eigenvalues and eigenvectors. Then, based on the eigenvalue, the contribution ratio is calculated, several principal components are selected from the one having the highest contribution ratio, and the all gray value vector is converted into a principal component vector using the eigenvector. in this way,
By selecting a significant feature based on the result of the principal component analysis, the dimension of the feature vector can be compressed, and the amount of processing thereafter can be reduced. Next, the obtained main component vector is averaged for each issue to obtain a reference vector, and this reference vector is used as a reference vector for each issue.
It is stored in 08. Further, the eigenvector of each issue is also stored in the dictionary file 108. Below, the dictionary file 108
The reference vectors of all brands stored in are collectively called a dictionary.

(B) Identification Phase In this phase, an unknown tablet is identified using the dictionary created in the learning phase to identify which brand is closest. Here, the tablet to be identified is input to this system.

Referring to FIG. 1, first, an input tablet is photographed by a camera as in the learning phase, and an inverse γ correction unit 102 and an RGB gray conversion unit 1 are applied to this image.
03, the position / orientation correction unit 104 and the feature extraction unit 106 perform a series of image processes to extract a grayscale value vector.

Next, the identifying unit 107 converts the gray value vector into a principal component vector by using the eigenvector obtained in the learning phase. Next, the distance between this principal component vector and the reference vector of each brand is calculated. At this time, Euclidean distance, similarity, etc. are used for the distance function. For general distance functions, see, for example, “Noboru Funakubo: Visual Pattern Processing and Recognition,” published by Keigaku Shuppan, 199.
0 ".

By the way, in the present embodiment, since the gray value vector is used as the basic feature vector, the distance calculation process is performed by superimposing the reference image and the identification target image to calculate the degree of coincidence. It is none other than. Such a feature vector has hardly been used in the conventional pattern identification system, but it is effective when a large category is to be processed because it can maximize the use of image information.

Further, the identifying section 107 outputs the brand having the shortest distance among the distances thus obtained as a correct answer.
Alternatively, some brands may be output as correct answer candidates in ascending order of distance, and the pharmacist may finally select the correct answer.

As described above, in the tablet identification system of this embodiment, the tablet identification inspection can be automatically performed without relying on human visual inspection.

The present invention is not limited to the above embodiments, but can be carried out in various different modes without departing from the scope of the invention. For example, the present invention can be applied to not only tablet identification but also identification of selected parts in an automatic assembly line.

[0050]

According to the present invention, it is possible to automatically identify articles.

Further, since the gray value vector extracted from the image of the article is used as the feature vector of the article, it is effective for identifying a large category.

Further, before the gray value vector is extracted, inverse γ correction is performed, RGB / gray image conversion using the G component is performed, or the gray value vector is created using the gray median value of the small section. In the case of, the grayscale value vector that well reflects the grayscale feature of the article is obtained, so that the identification accuracy is improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a tablet identification system according to an embodiment of the present invention.

FIG. 2 is a flowchart showing processing of a position / orientation correction unit in the embodiment.

FIG. 3 is a flowchart showing a process of a feature extraction unit in the embodiment.

FIG. 4 is a diagram showing an example of a gray image after position / orientation correction according to the embodiment.

FIG. 5 is a diagram showing a rectangular area cut out from the image of FIG. 4 in the embodiment.

FIG. 6 is a diagram showing a state in which the image of FIG. 5 in the embodiment is divided into small sections.

[Explanation of Codes] 1 NTSC TV camera 100 computer 101 image input unit 102 inverse γ correction unit 103 RGB gray conversion unit 104 position / orientation correction unit 105 feature extraction unit 106 learning unit 107 identification unit 108 dictionary file

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Internal reference number FI Technical display location G06T 7/00 9061-5L G06F 15/70 460 B

Claims (4)

[Claims] 1. An article identifying system for identifying which of a plurality of preset categories an identification target article belongs to, image input means for inputting an image of a sample article and an image of an identification target article relating to each category. , An inverse γ correction means for performing an inverse γ correction for removing the influence of an inherent γ correction on the input image, and a gray value vector extracted from the input image after the inverse γ correction Feature extracting means, based on the gray value vector extracted from the input image of the sample article, learning means for generating a reference vector of each category, a dictionary file for accumulating the reference vector, and the identification target article Based on the gray value vector extracted from the input image, by referring to the reference vector in the dictionary file, Article identification system, characterized in that it comprises identification means for identifying a Gori, a. 2. In an article identification system for identifying which of a plurality of categories an identification target article belongs to, a RGB color image of a sample article and an RGB color image of an identification target article are input for each category. An image input means; an RGB gray converting means for converting the RGB color image into a gray image by preferentially using the G component of the RGB color image; and a feature extracting means for extracting a gray value vector from the gray image. Learning means for generating a reference vector of each category based on a gray value vector extracted from the gray image of the sample article, a dictionary file for accumulating the reference vector, and extracted from the gray image of the identification target article The reference vector in the dictionary file based on the gray value vector And an identification unit that identifies the category of the identification target article. 3. An article identification system for identifying which of a plurality of preset categories an identification target article belongs to, and image input means for inputting an image of a sample article and an image of the identification target article in each category. , A predetermined area including the article area in the input image is divided into small sections, the median of the gray values of the pixels in each small section is sampled, and a gray value vector whose elements are the median gray values of the sampled small sections Feature extracting means for extracting, learning means for generating a reference vector of each category based on the gray value vector extracted from the input image of the sample article, a dictionary file for accumulating the reference vector, and the identification target article By referring to the reference vector in the dictionary file based on the gray value vector extracted from the input image of Article identification system characterized in that it comprises, identification means for identifying a category of another target object. 4. The system according to claim 1, wherein the learning unit performs principal component analysis on the gray value vector of the learning sample to obtain an eigenvalue and an eigenvector, and uses the eigenvalue and the eigenvector to perform the The gray value vector is converted into a significant principal component vector, the principal component vector is used as the reference vector, and the identifying means converts the gray value vector of the article to be identified into the principal component vector using the eigenvector. An article identification system, wherein identification is performed by referring to the reference vector based on a component vector.