JP4541995B2 - Figure recognition method - Google Patents

Description

  The present invention relates to a graphic recognition (pattern recognition) method, and more particularly, relates to graphic information embedding for recognition or embedding category information.

  2. Description of the Related Art Digital cameras, which have been popular in recent years, have excellent characteristics as graphic pattern input devices. Its portability and operability with a single button are considered to be easily accepted by users who feel that the scanner is difficult to use. It is also attractive that characters that cannot be moved in a real environment (for example, a signboard) can be input.

  However, when an input graphic pattern is used for recognition, for example, when a character in an actual environment is to be recognized, there are various problems from the viewpoint of graphic processing technology. For example, there are various distortions caused by shooting conditions. That is, characters subjected to distortion such as geometric distortion (particularly projective transformation distortion), low resolution, non-homogeneous illumination, blur / blurring cannot be handled by a conventional recognition method developed for OCR.

  Several methods for dealing with distortion caused by such a photographing situation have been proposed (for example, see Non-Patent Documents 1 and 2). For example, for geometric distortion, an elastic matching method for each character and a dewarping method for the entire document have been proposed (for example, see Non-Patent Document 3). For non-homogeneous illumination, methods such as local binarization have been studied (for example, see Non-Patent Documents 4, 5, and 6). All of these studies can be said to be a device for feature extraction and preprocessing for recognizing characters in the real environment as much as possible by the conventional OCR framework.

  From a viewpoint different from the above-described various methods, an approach has been made in which additional information (recognition information) that reinforces machine readability is embedded in the character pattern itself in advance. In other words, the character pattern side is devised to facilitate machine recognition in a real environment. Such an attempt has already been made in the early days of OCR and MICR (Magnetic Ink Character Recognition). That is, it is a font having a unique shape for machine reading (see, for example, Non-Patent Documents 7 and 8). For example, C.I. M.M. C. The MICR font called 7 is designed by using seven vertical line segments having two types of wide and narrow intervals, and these six intervals define a character code.

Or in recent years, DataGlyph is known (for example, refer nonpatent literature 9). DataGlyph is a data embedding method in characters that eliminates the influence on the appearance as much as possible. Specifically, the character figure is composed of a texture pattern consisting of fine “/” and “\”.
Kise, Omachi, Uchida, Iwamura: “Current status and issues of character recognition and document image analysis using cameras”, IEICE Technical Report, PRMU 2004-246 (2005) D. Doermann, J. Liang and H. Li: "Progress in camera-based document image analysis", Proc. ICDAR'03, pp. 606-616 (2003) S. Uchida and H. Sakoe: "A survey of elastic matching techniques for handwritten character recognition", IEICE Trans. Info. & Syst., E88-D, To appear (2005) Otani, Shio: "Character pattern extraction and recognition from scene images", IEICE (D), J71-D, 6, pp. 1037-1047 (1988) Senda, Nishiyama, Asahi: “Methods and Realization of Japanese Character Recognition Using Mobile Cameras”, IEICE Technical Report, PRMU 2004-124 (2004) Hirayama, Omachi, Aji: “High-precision extraction of character strings in scene images using color information”, IEICE Technical Report, PRMU2004-247 (2005) The British Computer Society: "Character Recognition 1967", Unwin Brothers Limited (1966) Hashimoto: “Introduction to Character Recognition”, Ohmsha (1980) DL Hecht: "Printed embedded data graphical user interfaces", IEEE Computer, 3, pp. 47-55 (2001)

  As described above, research on conventional pattern recognition including character recognition has been aimed at somehow machine recognition of patterns that have only high human readability. On the other hand, recently, as the research on ubiquitous computing suggests, photographing devices such as video cameras and portable cameras are ubiquitous, and there are many opportunities for machine recognition of various patterns. Considering such a situation, it seems that it will become important in the future to generate a pattern with high human readable and machine readable characteristics and to use it as a media that mediates both human and machine.

  However, the above-described OCR / MICR font, DataGlyph, and the like are all designed on the premise of imaging with a scanner, and are not premised on imaging with a camera. Therefore, for example, when extracting information from DataGlyph imaged by the camera, a method of estimating the original DataGlyph while correcting distortion at the time of imaging must be taken.

  A suitable information embedding method or recognition method assuming imaging not only by a scanner but also by a camera is desired. In other words, a technique that can extend the OCR / MICR font to character recognition in a real environment is desired.

  The present invention embeds information that reinforces machine readability in a figure using a technique suitable not only for scanning with a camera but also a camera, and a figure in a real environment in which the information is embedded, for example, a character pattern It is intended to provide a recognition method capable of recognizing the image with high accuracy.

In order to solve the above problems, the present invention provides means for solving the following two points.
First, a method of embedding additional information that is not affected by projective transformation distortion is provided. As a more specific example of this method, a figure in which a horizontal stripe pattern (hereinafter referred to as a cross ratio pattern) is embedded is provided.
That is, the present invention provides a recognition information embedding method characterized in that recognition information used for recognizing a graphic is created in a graphic for embedding recognition using the invariant.

  As a matter of course, the character pattern is included in the figure in this specification. The cross ratio calculated from the spacing of the stripes (for example, see Reference 2) is an invariant to the projective transformation, and even if the figure is taken from any angle, it can be quantized or the like. If there is no influence, the same cross ratio can always be extracted.

  Here, all types of figures to be recognized are classified in advance into a plurality of categories. If the cross ratio value is associated with each category, the category to which each figure belongs can be embedded as the cross ratio value. For example, an alphabetic character pattern can be classified into 26 categories so that one character belongs to each category. In this way, it is possible to create a recognition character pattern by embedding the category to which each character pattern belongs as a cross ratio value. The category embedded in the created recognition character pattern is embedded as invariant information even if it receives projective transformation distortion.

Secondly, embedded information is extracted from the created recognition graphic, for example, a recognition character pattern, and the extracted information and normal pattern recognition, for example, the result of normal character recognition based on the shape of the character pattern, are obtained. By integrating, it is possible to provide a recognition method capable of highly accurate recognition even if the recognition figure is subjected to projective transformation distortion.
That is, this invention extracts a graphic embedded in a recognition graphic using invariants, and recognizes an original graphic corresponding to the recognition graphic based on the extracted information. I will provide a.
(Reference 2) Sato: “Computer Vision”, Corona, Tokyo (1999)

  The recognition information embedding method according to the present invention embeds the recognition information in a figure using an invariant and creates a recognition figure. For example, even if the recognition figure is subject to projective transformation distortion, a projective transformation is performed. A recognition graphic created by embedding the recognition information using an invariant with respect to can give high recognition accuracy.

  The graphic may be a character image. In this way, a recognition character pattern that provides high character recognition accuracy can be created.

  The invariant may be an invariant for the projective transformation of the recognition graphic. Further, the invariant may be a cross ratio. In this way, even if the recognition figure is subject to projective transformation distortion, the embedded recognition information can be accurately extracted, and thus the created recognition figure gives high recognition accuracy. Can do.

  Furthermore, the information may be information in which a plurality of parallel stripe patterns are embedded in the figure so as to be obtained from the ratio of the widths of the stripes. In this way, recognition information can be embedded in a figure by a simple process of superimposing parallel stripe patterns, and the recognition information embedded by a simple process can be extracted from the width of each superimposed stripe. it can.

  Further, the graphic recognition method of the present invention extracts information embedded in a recognition graphic using invariants, and recognizes the original graphic corresponding to the recognition graphic based on the extracted information. When the recognition graphic is input after being subjected to projective transformation distortion, if the information is embedded using an invariant for the projective transformation, accurate information can be extracted to achieve high recognition accuracy.

  The recognition of the graphic may be performed based on the geometric feature independent of the information and the extracted information. Here, the shape feature independent of the information may be a feature that is not embedded, that is, a shape feature of the original graphic and is extracted as so-called normal pattern recognition. . For example, if the figure is a character pattern, it is a feature of characters recognized by so-called normal character recognition.

  The original figure may be a character image. In this way, it is possible to recognize characters with high character recognition accuracy from the recognition character pattern using the embedded information.

  The recognition graphic may be a graphic photographed by a camera. In this way, it is possible to perform recognition by inputting a figure in an actual environment using a camera or inputting a figure by a simple operation.

  The invariant may be an invariant for the projective transformation of the recognition graphic. Further, the invariant may be a cross ratio. In this way, even if the recognition graphic has undergone projective transformation distortion, the embedded information can be extracted accurately, and thus high recognition accuracy can be obtained.

  The information may be information embedded so as to classify all types of graphics to be recognized into a plurality of categories and indicate which category each graphic belongs to. In this way, high recognition accuracy can be obtained by selecting a correspondence between a preferred category and a cross ratio. For example, the correspondence between the category for the character pattern and the cross ratio is determined in advance, and the category to which the character pattern belongs is embedded in the recognition character pattern. The embedded ratio can be extracted from the input recognition character pattern and used as information for recognition. By using the recognition of the character shape and the category information extracted as the cross ratio value, it is possible to obtain very high recognition accuracy even if the recognition character pattern is subjected to the projective transformation distortion.

  Hereinafter, the present invention will be described in more detail with reference to the drawings. The following description will provide a better understanding of the present invention.

  As described above, the present invention relates to the embedding of additional information using an invariant. As a suitable example of the invariant, a cross ratio that is an invariant with respect to projective transformation distortion is taken up. However, the essence of the present invention is not limited to the projective transformation distortion, and is not limited to the cross ratio. Further, as a preferred example of embedding the cross ratio in the original graphic, a mode of superimposing a horizontal stripe pattern as a pattern on the original graphic is taken, but the mode of embedding additional information is not limited to this. Furthermore, although attention is paid to characters as a target for embedding additional information, the same technique can be applied to figures other than characters as they are. Examples of such graphics include various marks and symbols, company logos, and trademark graphics. Thus, the following description should be considered in all respects as illustrative and not restrictive.

1. First, consider embedding a pattern (pattern) having a certain ratio in a character image pattern. Various forms can be considered for the correspondence between the embedding cross ratio value and the character category. For example, different cross ratios may be embedded in each category, or the same cross ratio may be embedded in a plurality of categories. In the former case, if even the cross ratio can be extracted correctly, it can be identified without using the character shape. Hereinafter, general cases including both cases will be described.

1.1. Various ratios of embedding patterns can be considered. Here, a pattern (hereinafter referred to as a cross ratio pattern) in which a set of five parallel stripes is embedded in the entire character image including the character lines and the background. Think about it. However, the number of bands is not limited to five, and may be increased. In this case, a plurality of cross ratios may be embedded. FIG. 2 is an explanatory diagram illustrating a character image in which a cross ratio pattern is embedded and a projective transformation example thereof. Among these, FIG. 2A is an example of a character in which a cross ratio pattern is embedded. Here, for convenience of explanation, a multi-ratio pattern that is conspicuously colored is used.

Of the five bands, the first and last bands are guides representing the existence range of the cross ratio pattern. This represents the cross ratio embedded by the remaining three bands sandwiched between these two guides. If the width of these three bands is l ₁ , l ₂ , l ₃ , the cross ratio by this pattern is
It is represented by By changing the band widths l ₁ , l ₂ , and l ₃ , the value of the embedding ratio can be controlled.

The cross ratio originally has a continuous value, but here, K values r _k (k = 1, 2,..., K) obtained by quantizing it are considered. The specific setting method of the value r _k, described below.
Note that only a minimum of three bands excluding the guide have been described for the band representing the cross ratio pattern, but there may be a plurality of bands.

  In general, when there are N bands, a multi-ratio can be recorded for a combination in which three of the N bands are selected. By recording a plurality of cross ratios in this way, the following effects can be obtained. First, in the case where the same value is recorded in a plurality of combinations, it is possible to increase the resistance to cross ratio reading errors. Further, when different values are recorded for each combination, the capacity of information that can be recorded increases.

1.2. When extracting multiple ratio from embedded character image extraction method cross ratio r _k of cross ratio is a straight line is drawn so as to traverse it (linear p in FIG. 2 (b)), flanked by the guide of the line The width of the band (l ′ ₁ , l ′ ₂ , l ′ ₃ ) is obtained for the obtained section, and the cross ratio is calculated by the equation (1) as l _i = l ′ _i (i = 1, 2, 3). Calculate it. Without noise such as quantization error, how even a straight line drawn, and how even under projective conversion, determined the same cross ratio r _k and when embedded. Therefore, in principle, the same cross ratio can always be obtained regardless of the angle at which a character image in which this cross ratio pattern is embedded is photographed.

Actually, the extracted cross ratio has an error due to the influence of quantization or the like. As a countermeasure, P straight lines are drawn at random, and the true value is estimated from the P cross ratio values obtained respectively. Specifically, first, we obtain a cross ratio r in the manner described above for each line of the P present, selects the closest r _k to the cross ratio r. After making this selection process for the linear total P present, the most selected r _k, the cross-ratio embedded in the character pattern.

1.3 Design of cross ratio pattern Consider changing the cross ratio value by changing the width of each band of the cross ratio pattern. In order to minimize the influence of quantization error or the like, for example, theoretical analysis as in Reference 3 is considered necessary, but here, the following procedure is simply set.
Let L be the width of the portion excluding the guide from the cross ratio pattern. That is, L = l ₁ + l ₂ + l ₃ . L and l ₁ are common in the K-type cross ratio pattern. _{Therefore, l 2 + l 3 = L} -l satisfy _{1 l 2, l r k (} k = 1,2, ... K) 3 set to by setting K as a to define a.

Specifically, defining a r _k by l _2, l ₃ satisfying the following equation.
Here, ε is a positive constant that determines the minimum width of l ₂ and l ₃ . According to the above equation, the boundary points of l ₂ and l ₃ are set to a dividing point when the section that can be taken is divided into K-1 at equal intervals and a total of K points at both ends of the section.

2. Integration of Character Recognition Result and Cross Ratio Extraction Result As described above, when the type K of embedded cross ratios is the same as the number of character categories | C | (C is a set of character categories), the respective ratios _{c c} Can be identified using only the extracted cross ratio without using the character shape. However, when targeting character sets with many categories such as kanji, it is difficult to define multiple types of cross ratios with limited resolution.

In the more general case of K <| C |, since the cross ratio and the category have a one-to-many relationship, ambiguity remains in the recognition result of the cross ratio alone. Specifically, when a set of character categories embedded the k cross ratio r _k and C _k (⊂C), be cross ratio r _k from a certain input character extraction, the character is included in the C _k I don't know any of the categories. Therefore, it is considered that the category is narrowed down to one as the final recognition result by integrating the recognition result using the character shape (that is, the result of normal character recognition) and the recognition result by the cross ratio.

  In the same way as there are many forms of integration of a plurality of discriminators such as voting, various methods can be considered for this integration. However, general discriminator integration and this integration differ in the following two points. First, as described above, ambiguity remains in the recognition result based on the extracted cross ratio. The second is that the recognition accuracy by the cross ratio (that is, the extraction accuracy of the cross ratio) is overwhelmingly higher than the recognition accuracy by the character shape. As will be apparent from the experimental results described later, if the projective transformation distortion at the time of shooting is large, the recognition rate due to the character shape is significantly reduced. On the other hand, the cross ratio, which is an invariant to the projective transformation, is always correctly extracted if there is no noise. In practice, some erroneous extraction occurs due to noise, but the extraction accuracy of the cross ratio is still considerably higher than the recognition accuracy based on the character shape.

Considering the above points, the categories are narrowed down to some by the extracted cross ratio, and then the category is uniquely determined by the recognition result by the character shape. Specifically, first extracts the cross ratio r _k from the input character, it is considered correct, as the final recognition result category recognition costs character shape (distance) had the lowest among the C _k . In the case of K = | C |, the recognition is performed only by the cross ratio without using the character shape.

In this method, the result of recognition in the character shape recognition cost s even if higher correct answer categories, (i) are cross ratio r _k is correctly extracted, and (ii) C _k in a non correct category category If the recognition cost is greater than s, a correct recognition result can be obtained. On the contrary, if the cross ratio extraction fails, even if the correct answer is obtained by the recognition by the character shape, the recognition ratio is changed to erroneous recognition. However, since the cross-ratio extraction accuracy is high as described above, it is considered that this deterioration is small.

It is a good idea that the category set C _{k to} which the same cross ratio r _k is assigned is composed of categories that are unlikely to be erroneously recognized in recognition by the character shape. This is a device for easily satisfying the condition (ii). In order to construct such an assignment {C _k | k = 1,..., K}, first, the misrecognition characteristics of the technique used for recognition by character shapes are grasped, and category pairs that are not likely to be misrecognized from each other are known. After obtaining, the cross ratio assignment {C _k | k = 1,..., K} may be determined so that such category pairs are included in one set C _k as much as possible. The effect of this assignment will be further described later.

(Experimental example)
3.1 Preparation 3.1.1. Font Image As a character image for embedding the cross ratio, an English capital letter 26 image of the font “Arial” was used. The number of vertical pixels of the font original image is slightly different in each category, and is 194 at the minimum, 212 at the maximum, and 196 on the average. On the other hand, the number of horizontal pixels varies greatly from category to category, with a minimum of 52 (“I”), a maximum of 251 (“W”), and an average of 170. In this experiment, the font type does not affect the extraction accuracy of the cross ratio. This is because the cross ratio pattern is embedded not only in the character line but also in the background portion, that is, the cross ratio pattern can be extracted regardless of the character shape. In the future, if a study of embedding the cross ratio pattern only on the character line is made, it is considered that some differences in the cross ratio extraction accuracy appear depending on the font.

3.1.2 Design of cross ratio pattern For the cross ratio pattern, the width of the guide is 5 pixels, the width L of the entire cross ratio pattern is 150 pixels, and further l ₁ and ε are 15 pixels. Under this _{condition, r k (k = 1,2,} ..., K) according to the method described in the item 1.3 was set. Further, C _k was determined according to the method described later, and then the cross ratio pattern by r _k was embedded in a font of category cεC _k . FIG. 1 is an explanatory diagram showing an example of a character image when different cross ratio patterns are embedded in all categories (that is, K = | C | = 26).

3.1.3 Test pattern generation by projective transformation For a font image in which a cross ratio pattern is embedded, the x and y coordinates of its four corners are displaced by ± δ (δ = 0, 4, 8,..., 48) pixels, respectively. Then, projective transformation was performed, and 2 ⁴ × ² = 256 test patterns were generated for each δ. FIG. 3 is an explanatory diagram showing an example of a test pattern generated by projective transformation. Thus, when δ increases to about 36, a character image distorted to an unrealistic level is generated. That is, the current test pattern includes a character pattern that is more strongly distorted than when intentionally photographed.

3.1.4 Recognition Method Based on Character Shape As described in item 2 above, in the case of general K <| C | Character recognition method). Although any method can be used, the following two known methods were used in this experiment.
Character Recognition Method 1—Simple Matching This method is a method in which a standard pattern and an unknown input pattern are simply overlapped to obtain a matching cost of both character image patterns, that is, a distance.

Character recognition method 2 -elastic matching This method, if intuitively described, is a method of superimposing one character image pattern while deforming it like a rubber film. As shown in Non-Patent Document 3, there are various types of elastic matching. Here, the following method is used. FIG. 4 is an explanatory diagram showing the elastic matching method used in this experiment. As shown in FIG. 4, each column of the input pattern is associated with the standard pattern as an inclined straight line while maintaining the adjacent relationship. At that time, one-dimensional expansion and contraction is allowed inside each row. Although details are omitted, an algorithm based on dynamic programming (DP) can be used to optimize the entire map. The degree of freedom of this elastic matching method is higher than that of the projective transformation, and therefore, in principle, any projective transformation can be compensated.

  In both methods, linear normalization of character size was performed as preprocessing. The feature amount of each pixel is set to 1 (character line) and 0 (background) very simply.

3.1.5 Configuration of the assignment set C _k of the cross ratio, namely either assign a Fukuhi r _k in which category c, the item 2. As described later, it is very important as a factor that greatly affects the final recognition performance. In this experiment, the following two allocation methods were used.

Allocation method 1 simple assignment scheme which is a method of allocating a cross ratio r _k satisfying k = ((c-1) mod K) +1 to category c.
Allocation method 2—optimal allocation method This is a method (for example, see Reference 4) in which a cross ratio is allocated to each category in consideration of the characteristics of the recognition method based on character shapes. In this method, the confusion matrix of the recognition method based on the character shape is used as a clue for the cross ratio assignment. For example, if it is found from the confusion matrix that categories that can give a recognition result of “H” are “H” and “N”, different cross ratios are assigned to these two categories in order to avoid ambiguity. In accordance with this policy, cross ratio assignment will be established for all categories.

  Table 1 shows the simple assignment and the optimum assignment when K = 4, 12, 30 when the elastic matching method described in the previous section is used. At any K, the same cross ratio is assigned to “C” and “V” by the optimum assignment. This suggests that erroneous recognition is unlikely to occur between these two categories. In determining the optimum allocation, a confusion matrix obtained from recognition results for all data of δ = 0 to 48 was used.

3.2. Recognizing result of character shape alone Using the two methods (simple and elastic matching methods) described in item 3.1.4 above, the recognition experiment using character shape alone is not performed. I did it. As the standard pattern, the font image itself (that is, the character pattern in FIG. 1) that has not undergone projective transformation was used. FIG. 5 shows changes in the recognition rate depending on the degree of projective transformation δ. It can be seen that simple matching is very sensitive to projective transformation, and that the recognition rate rapidly decreases when the deformation amount δ increases even a little. On the other hand, since elastic matching has a deformation absorption capability, it can be seen that a constant recognition rate can be maintained up to about δ = 28. However, if more deformations are added, the recognition rate decreases as in simple matching.

  From this result, it is expected that it is difficult to achieve the same accuracy as a barcode in an actual environment by only recognizing the character shape alone. Of course, much room for improvement remains. For example, this time, a very simple distance evaluation is performed using only binary pixel values as feature amounts. Also, with the matching method, since the degree of freedom of the elastic matching used this time is higher than that of the projective transformation, there is a high possibility that misrecognition due to excessive matching has occurred. Actually, the reason why the maximum recognition rate of elastic matching remains at about 90% is that misrecognition occurs frequently due to a synergistic influence of this simple feature amount and a too high degree of freedom. In particular, in the vicinity of δ to 0, a plurality of categories in which the distance from the input pattern becomes 0 due to over-matching occurred, and as a result of the tie break, an incorrect answer category was often selected. Therefore, it is considered that the recognition rate can be considerably improved from the current situation by using more sophisticated feature amounts and matching methods. However, considering the tendency of the experimental results and the fact that this experiment is a simulation in a computer, the recognition performance of the character shape alone seems to be limited.

3.3. Cross-ratio extraction accuracy For the test patterns in which the cross-ratio was embedded according to simple assignment, the cross-ratio extraction accuracy was measured. FIG. 6 is a graph showing the accuracy of the cross ratio extracted from the test pattern as a result of this experiment. Even when subjected to considerable projection conversion, cross-ratio r _k embedded in it it can be seen that can be extracted in a stable manner. When compared with the recognition result of the character shape alone described in the previous section, there is a difference of about 1 to 2 digits in the number of recognition errors. In particular, the accuracy in the case of K = 4 and K = 12 sufficiently suggests that the cross ratio pattern is useful in aiming at the recognition performance of the barcode accuracy.

The result of K = 26 in the figure corresponds to the character recognition result with the single ratio. If the projection transformation distortion when photographing intentionally is at most δ = 24 or less, even a simple cross ratio pattern as used in this experimental example has a recognition rate of 98% or more without using a character shape. It turns out that it is obtained.
As a cause of erroneous extraction, there is an error included in the width (l ′ ₁ , l ′ ₂ , l ′ ₃ ) of each band of the cross ratio pattern due to quantization. In fact, taking the case of K = 26 as an example, if the extracted erroneously r _k as r _k ± ₁ was about 85% of the total erroneously extracted. The more serious mis-extraction (r _k ± Δ, Δ ≧ 2) was about 10%. The remaining 5% was when the guide could not be captured due to the influence of quantization, and the multi-ratio pattern itself could not be detected.

3.4. Integration Result of Character Recognition Result and Cross Ratio Extraction Result FIGS. 7 and 8 are graphs showing the recognition rate when the recognition result by the character shape alone and the recognition result by the cross ratio are integrated. The former is when simple matching is used, and the latter is when elastic matching is used. In addition, both are the results when simple assignment is used.
From both figures, it can be seen that the recognition rate can be significantly improved by using the result of character recognition by shape and the result of cross ratio extraction in comparison with the result of the former alone (FIG. 5). For example, when the result of elastic matching in the case of δ = 4 is seen, the recognition rate of the character shape alone is 89.8%, but when combined, K = 4 is 97.4% and K = 12. 99.1%, K = 20, improving to 99.97%.

  The reason for this improvement will be discussed with reference to FIG. FIG. 9 is an explanatory diagram showing a confusion matrix for elastic matching when a test pattern of δ = 4 is recognized by a character shape alone. From this confusion matrix, for example, it can be seen that there are two possibilities of “H” and “N” as the true category of a pattern recognized as “H” due to its shape. In other words, it can be said that ambiguity remains in the discrimination between these two categories when the character shape alone is recognized. On the other hand, if the cross ratio is embedded according to the simple assignment method, different cross ratios are assigned to “H” and “N” in both cases of K = 4, 12, and 20 (see Table 1). Eventually, the ambiguity remaining in the recognition of the character shape alone is eliminated by the cross ratio, and “H” and “N” can be correctly recognized. Since the same disambiguation occurred between other categories, it is considered that a significant improvement in accuracy was obtained.

  7 and 8 again, when the projection distortion is not large (about δ ≦ 16), the result obtained by using the recognition by the character shape is more than the result of the recognition by the multi-ratio alone (K = 26). It can be seen that a high recognition rate is obtained. Although the extraction accuracy of the cross ratio itself is much higher than the recognition accuracy based on the character shape, if a large number of cross ratios are embedded to completely narrow down the category, the extraction accuracy of the cross ratio itself deteriorates (FIG. 6). After all, in this experiment setting, it can be said that it was better to secure the extraction accuracy by reducing the number of cross ratios to some extent.

3.5. Effect of Allocation Method FIG. 10 is a graph showing the effect of the allocation method on the recognition rate as a result of this experiment. The results of using simple assignment and optimum assignment are shown for each of cases where the cross ratio type K is 4 and 12. Elastic matching is used as a recognition method by character shape.
From this result, it can be seen that a considerable improvement can be obtained by using optimal allocation. In particular, in the case of δ ≦ 20, it is worth noting that a higher recognition rate is obtained with the optimally assigned four types of cross ratios than with the simply assigned double types of cross ratios. In this way, when assigning a cross ratio, it is possible to achieve significant efficiency by assigning an appropriate cross ratio to each category after grasping the characteristics of the recognition method based on the character shape.

  Finally, it is apparent that there can be various modifications of the present invention in addition to the above-described embodiment. Such variations are not to be construed as not belonging to the features and scope of the invention. The scope of the present invention is intended to include all modifications within the meaning and range equivalent to the scope of the claims.

(Reference 3) Kanazawa, Matsunaga, Kanaya: "Design of the best marker pattern for discrimination by multi-ratio: Theoretical analysis", Jisho Kenji, 99-CVIM-115-13 (1999).
(Reference 4) Iwamura, Uchida, Omachi, Kise: "Realization of 100% recognition rate by adding information-For information transmission that can be understood by both humans and machines-", Image Recognition and Understanding Symposium (MIRU2005) (2005).

It is explanatory drawing which shows the example of the character image at the time of embedding different cross ratio patterns in all the categories. It is explanatory drawing which shows the character image which embedded the ratio ratio pattern, and its projective transformation example. It is explanatory drawing which shows the example of the test pattern produced | generated by projective transformation. It is explanatory drawing which shows the method of the elastic matching used in this experiment. It is a graph which shows the change of the recognition rate by the grade (delta) of projective transformation. It is a graph which shows the precision of the cross ratio extracted from the test pattern. It is a graph which shows the recognition rate at the time of integrating the recognition result by a character shape independent, and the recognition result by a cross ratio, when simple matching is used. It is a graph which shows the recognition rate at the time of integrating the recognition result by a character shape independently, and the recognition result by a cross ratio when elastic matching is used. It is explanatory drawing which shows the confusion matrix of an elastic matching at the time of recognizing the test pattern of (delta) = 4 by the character shape independent. It is a graph which shows the influence which the allocation system has on the recognition rate.

Explanation of symbols

  1 Recognition figure

Claims (7)

For a figure to be classified into any one of a plurality of categories by pattern recognition, a computer determines a plurality of parallel stripe patterns according to the category to which the figure belongs,
Create a recognition figure by superimposing the determined pattern of parallel stripes on the figure,
The method of embedding recognition information , wherein the parallel fringe pattern has a cross ratio as a ratio of widths of the parallel fringes determined according to the category .   The embedding method according to claim 1, wherein the graphic is a character image. The width ratio of the parallel stripes corresponds to a value determined so that all types of figures to be recognized are classified into a plurality of categories and each figure belongs to which category. Embedding method.   In the pattern of the parallel stripes, the stripes at both ends represent the boundary where the pattern exists, a plurality of stripes are sandwiched between the boundaries, and the cross ratio represented by the ratio of the width of the sandwiched stripes corresponds to the category The embedding method according to claim 1, wherein the embedding method is a value. A computer extracts a ratio of the width of the parallel stripes from a recognition figure in which a pattern of parallel stripes is superimposed on a figure to be classified into any one of a plurality of categories ;
Determine the category to which the recognition graphic belongs based on the extracted ratio value or the extracted ratio value and the shape of the graphic ;
The pattern recognition method according to claim 1, wherein the parallel stripe pattern includes a cross ratio of values corresponding to the category as a ratio of widths of the parallel stripes . The recognition method according to claim 5 , wherein the original figure is a character image. The recognition method according to claim 5, wherein the recognition graphic is a graphic photographed by a camera.