JP2005013363A - Image analysis apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately recognize a relatively complicated image by recognizing a similar image as similar image data in processing the similar image. <P>SOLUTION: This image analysis apparatus comprises a memory means 101 for storing a template group consisting of a plurality of first vectors, a cut-out means 102 for cutting out a first region from an inputted image, a moving means 103 for moving a first region within a second region, a converting means 104 for converting a partial image inside the first region to second vectors, an evaluation means 105 for calculating the correlation between the first vectors and second vectors and determining an evaluation value expressing the similarity between them, and a specifying means 106 for specifying a third region based on the evaluation value. The position of a characteristic point within the third region is determined by using the positional relation to a predefined origin in the third region, independently determined for each characteristic point. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an image analysis apparatus, and is particularly suitable for use in image analysis for extracting image features typified by dental X-ray images.
[0002]
[Prior art]
Conventionally, in order to recognize an anatomical feature point of a patient from a cephalometric X-ray image, an orthodontist who has received specialized medical training observes the cephalometric X-ray image. I was aware.
[0003]
An example of a head X-ray image is shown in FIG. An orthodontist or the like needs to recognize each anatomical feature point indicated by + in the drawing. Here, among each anatomical feature point, N is a connection point of the frontal bone (bone of the forehead) and the nasal bone (bone of the base of the nose), S is a point of the center of the pituitary gland, Or is a left or right eye socket (eyeball) Is the midpoint of the lowest point of the bone), MPo is the uppermost point of the ear rod (outer ear), APo is the center point when the ear hole is viewed from the side, and L1 is the lower central incisor (1 U1 is the point of the tip of the upper middle incisor (first tooth), Gn is the point in front of the chin, Go is the point of the mandibular angle, and Ar is the shadow image of the lower skull base edge And Cd is the highest point of the condylar head, Ba is the lowest point on the front edge of the large occipital foramen, and li is the highest protruding point of the lower lip. The anatomical feature points of the soft tissue part are shown in white letters.
[0004]
[Patent Document 1]
Japanese Patent Application No. 10-326253
[Patent Document 2]
Japanese Patent Application No. 2000-326158
[Non-Patent Document 1]
Masakazu Yagi, Masayoshi Adachi, and Tadashi Shibata, “A Hardware-Friendly Soft-Computing Algorithm for Image Recognition,” Proceeding ^th European Signal Processing Conference (EUSIPCO 2000), pp. 729-732, Tampere, Finland, Sept. 4-8, 2000.
[Non-Patent Document 2]
Masakazu Yagi and Tadashi Shibata, "A Human-Perception-like Image Recognition System based on PAP Vector Representation with Multi Resolution Concept," in the Proceedings of 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2002), Vol. I, pp. 1041-1048, Florida, May 13-17, 2002.
[Non-Patent Document 3]
Masakazu Yagi, Tadashi Shibata and Kenji Takada, "Optimizing Feature-Vector Extraction Algorithm from Grayscale Images for Robust Medical Radiograph Analysis," in press in The Proceedings of Fourth International Conference on Multimedia and Image Processing (IFMIP 2002), Orland, June 9- 13, 2002.
[Non-Patent Document 4]
Masakazu YAGI, Tadashi SHIBATA, and Kenji TAKADA, "Human-Perception-Like Image Recognition System Based on the Associative Processor Architecture," in the Proceedings of XI European Signal Processing Conference, Vol. I, pp. 103-106, Sep. 3-6, 2002 Toulouse, France
[Non-Patent Document 5]
Masakazu Yagi and Tadashi Shibata, "An Associative-Processor-Based Mixed Signal System for Robust Image Recognition," in the Proceedings of 2002 IEEE International SyMPosium on Circuits and Systems (ISCAS 2002), pp. V-137-V-140, Arizona, May 26-29, 2002.
[Non-Patent Document 6]
T.A. Yamazaki and T.K. Shibata, “An Analog Similarity Evaluation Circuit Featuring Variable Functional Forms,” Proceedings of The 2001 IS IEEE International SymSum III-561-564, Sydney, Australia, May. 6-9, 2001.
[Non-Patent Document 7]
Toshihiko Yamasaki, Ken Yamamoto, and Tadashi Shibata, “Analog Pattern Classifier with Flexible Reciprocating Circadian Preference ^th European Solid-State Circuits Conference (ESSCIRC 2001), Ed. by F.E. Diegocher and H.C. Grunbacher, pp. 212-215 (Frontier Group), Villach, Austria, September 18-20, 2001.
[Non-Patent Document 8]
T.A. Yamazaki and T.K. Shibata, “Analog Soft-Pattern-Matching Classifier Using Floating-Gate MOS Technology,” Neural Information Processing Systems 14, in press.
[0005]
[Problems to be solved by the invention]
In recent years, there has been an increasing demand for automatically executing recognition of patient anatomical feature points to generate data. However, image analysis technology for accurately and promptly analyzing extremely complicated and delicate images such as cephalometric X-ray images for dentistry is currently under development, and is awaiting its immediate realization.
[0006]
The present invention has been made to solve such a problem. When image processing is performed on a similar image, the image can be recognized as similar image data. An object of the present invention is to provide an image analysis apparatus capable of accurately and quickly recognizing an image feature represented by an X-ray image for dentistry.
[0007]
[Means for Solving the Problems]
An image analysis apparatus according to the present invention is an image analysis apparatus that finds a feature point representing a position of a specific part defined in advance in an image of a specific target object from an input image and outputs a coordinate value of the feature point. Means for storing at least one first template group consisting of a plurality of first vectors representing the feature points for each of the feature points, and a first of a predetermined size from the input image. Means for cutting out the first area, means for moving the first area in a second area having a predetermined size in the target input image, and a partial image in the first area as a predetermined area. Means for converting to a second vector according to a procedure; a first evaluation value for performing a correlation operation between the first vector and the second vector; And means for determining Means for specifying a third region corresponding to the first region including the feature point therein based on the evaluation value of the feature point, and determining the position of the feature point in the third region. This is determined by using a relative positional relationship uniquely determined for each of the feature points with respect to an origin defined in advance in the region.
[0008]
An image analysis apparatus according to the present invention is an image analysis apparatus that finds a feature point representing a position of a specific part defined in advance in an image of a specific target object from an input image and outputs a coordinate value of the feature point. Means for storing at least one first template group consisting of a plurality of first vectors representing the feature points for each of the feature points, and a first of a predetermined size from the input image. Means for cutting out the first area, means for moving the first area in a second area having a predetermined size in the target input image, and a partial image in the first area as a predetermined area. Means for converting to a second vector according to a procedure; a first evaluation value for performing a correlation operation between the first vector and the second vector; And means for determining Means for identifying a third region corresponding to the first region including the feature point therein based on the evaluation value, and expressing at least one feature point an image different from the feature point A means for obtaining a second evaluation value representing the degree of dissimilarity using the second template group and both the first evaluation value and the second evaluation value are used to determine the third evaluation value by a predetermined specifying method. Means for specifying the region.
[0009]
An image analysis apparatus according to the present invention is an image analysis apparatus that finds a feature point representing a position of a specific part defined in advance in an image of a specific target object from an input image and outputs a coordinate value of the feature point. Means for storing at least one first template group consisting of a plurality of first vectors representing the feature points for each of the feature points, and a first of a predetermined size from the input image. Means for cutting out the region, means for specifying the second region using a predetermined region dividing method, means for moving the first region within the second region, and within the first region Converting the partial image of the second image into a second vector according to a predetermined procedure, and performing a correlation operation between the first vector and the second vector, and calculating the similarity to each of the feature points Determine the first evaluation value to represent That includes means, and means for identifying a third region corresponding to the first area including the feature points therein based on the first evaluation value.
[0010]
DETAILED DESCRIPTION OF THE INVENTION
DESCRIPTION OF EMBODIMENTS Hereinafter, preferred embodiments to which the present invention is applied will be described in detail with reference to the drawings.
[0011]
(First embodiment)
-Basic outline-
FIG. 1 is a schematic diagram illustrating a schematic configuration of the image analysis apparatus according to the first embodiment.
This image analysis apparatus finds a feature point representing the position of a specific part defined in advance in an image of a specific object from an input image, and outputs a coordinate value of the feature point. In the present embodiment, for example, a head X-ray image for dentistry is used as an input image, and for example, when recognizing a face eye contact point, the eye contact point is not always used as a central part, but appropriately according to the case. Set the center. As an example, let the point of the central part of the eye be the central part. Here, not only the eyes of the face but also parts such as the nose tip and the corner of the mouth may be recognized.
[0012]
The image analysis apparatus includes a storage unit 101 that stores at least one first template group including a plurality of first vectors that represent the feature points for each of the feature points, and a predetermined one of the input images. Cutting means 102 for cutting out a first area of a size; moving means 103 for moving the first area within a second area of a predetermined size in the target input image; The conversion means 104 for converting the partial image in the region into the second vector according to a predetermined procedure, and the correlation calculation between the first vector and the second vector is performed, and the similarity to each of the feature points Evaluation means 105 for determining a first evaluation value representing the above, and specifying means 106 for specifying a third area corresponding to the first area including the feature point based on the first evaluation value The third territory The position of the feature point in the inner, is determined using the relative positional relationship defined uniquely to feature points of each against predefined origin in the third region.
[0013]
The storage unit 101 has a function of creating a first vector constituting the first template group by performing learning that always uses a predetermined relative positional relationship, and uses edge information of image information. At this time, when extracting edge information, a median value of image information may be used. Moreover, you may comprise so that the shape of a 1st area | region may differ from the thing of another feature point about at least 1 feature point.
[0014]
When extracting the edge information, the conversion unit 104 uses, for example, the median value of the information in the fourth area having a predetermined size in the first area. When this conversion means 104 is used, it is preferable to use a predetermined integrated circuit during the processing to improve the processing efficiency. Similarly, when using the determination means 105, a predetermined integrated circuit may be used during the processing to improve the processing efficiency.
[0015]
The image analysis apparatus according to the present embodiment is applied to, for example, a medical X-ray image, in particular, an X-ray image for dentistry, and finds a predetermined feature point medically defined as a feature point. The predetermined feature point includes a point on a soft tissue. Here, the soft tissue is a soft tissue other than bone such as muscle and fat.
[0016]
For example, it is conceivable to find Point A and Point B as the predetermined feature points (here, anatomical feature points). Further, it is conceivable to find N (Nation) as the predetermined feature point. In this case, when N is found as the predetermined feature point, it is utilized that the pixel values in the first region including N are close to homogeneous pixels with a white bone portion and a black background portion. It is also conceivable to find S (Sella) as the predetermined feature point. In this case, when S is found as the predetermined feature point, it is first found at a predetermined resolution, and is found with higher resolution and higher accuracy than the predetermined resolution. Furthermore, when S is found as a predetermined feature point, it is recognized using a partial feature of S. It is also conceivable to find Or (Orbitale) as the predetermined feature point. Furthermore, it is conceivable to find MPo and / or APo as the predetermined feature points. When finding MPo and / or APo as the predetermined feature point, the X-ray image for dentistry is binarized and found. Furthermore, it is conceivable to find L1 and U1 as predetermined feature points. When finding L1 and U1 as the predetermined feature points, tooth enamel information is used. Further, when U1 is found as the predetermined feature point, the target point is set as the upper right region in the third region as a relative positional relationship uniquely determined for each feature point.
[0017]
-Specific examples-
FIG. 2 is a schematic diagram illustrating the image analysis apparatus according to the first embodiment.
The image analysis apparatus extracts image data of a region (x, y) defined corresponding to a predetermined position in an input image, and generates a vector generation unit 1 that represents the image data as a vector, and at least a reference pattern. Similarity in which the storage unit 2 having a plurality of pattern groups including one, the vectorized image data and each reference pattern included in each pattern group are collated, and the similarity with the image data for each reference pattern is evaluated A degree calculation unit 3 and a winner take all circuit 4 that performs a predetermined calculation on each evaluation value of similarity and determines at least one evaluation value. However, the winner-take-all circuit is not always used here. Further, it is not always necessary to extract only one evaluation value, and a plurality of evaluation values may be determined depending on circumstances.
[0018]
The storage unit 2 has, for example, 15 first template groups as pattern groups. However, the number of templates is not limited to 15 and may be 100 or 1000. It is also possible to reduce the number of templates by applying a learning algorithm to many sample groups. Each first vector is a vector expression generated by a PAP (Principal Axis Projection) method (see Patent Document 1) described later. Here, the vector representation held by the template is not necessarily such an X-ray image, and the conversion from the partial image to the vector representation need not be the PAP method.
[0019]
In this image analysis apparatus, first, search processing is performed in the second region in the given input image (step 1). Next, a partial image included in the first area having a size of 64 × 64 pixels having the reference coordinates (x, y) in the given input image is cut out (step 2). Here, the position of (x, y) in the partial image is not necessarily the center, and is changed depending on the recognition target. Further, the size to be cut out is not necessarily 64 × 64.
[0020]
Then, the partial image is input to the vector generation unit 2 and converted into a vector representation by the PAP method (see Non-Patent Documents 1 to 4) (step 3). Here, the PAP method is not always used when converting a two-dimensional partial image into a vector representation (a vector representation generation method from a two-dimensional image).
[0021]
Manhattan distance calculation of similarity evaluation values between vector representations generated from partial images cut out from the input image and all first vector representations present in all first template groups stored in the system (Step 4). Here, the calculation of the similarity evaluation value is not necessarily performed on all the first vectors existing in the first template group. For example, the first vector for calculating the similarity may be a part of the first vector. Further, the similarity calculation does not always use the Manhattan distance calculation. It is possible to use Euclidean distance calculation, histogram intersection, cullback distance, or the like.
[0022]
Among the evaluation values of the similarity calculated by the above-described distance calculation method between vectors, the evaluation value having the highest similarity (the one having the lowest evaluation value) is determined (step 5). Then, the evaluation value information is held in the position information (x, y) in the two-dimensional distribution diagram (step 6). As the evaluation value recorded here, the one having the highest similarity among the similarity evaluation values is selected, but this method is not always followed. For example, the evaluation value may be calculated and recorded as a group using the evaluation values of all the first vectors that have been subjected to the correlation value calculation, or the average of the evaluation values having the highest similarity among the first vectors. May be calculated and recorded.
[0023]
Such a system was applied to L1, U1, and N among anatomical feature points that the orthodontist must identify the position from the cephalometric X-ray image. With respect to L1 and U1, as shown in FIG. 3, a method of cutting out with L1 and U1 as the center can obtain almost the same image, and recognition accuracy that can distinguish these two points cannot be obtained. In addition, if U1, information on the lower teeth is included in the first region if L1, information on the upper teeth is included in the first region, and variations thereof increase and recognition accuracy decreases. In this method, as to U1, the method of clipping is defined so that U1 is at the lower right of the first region and L1 is at the upper right of the first region.
[0024]
FIG. 4 shows how to cut out N. The definition of N is the connecting part of the skull and nasal bones, and the part where the line is slightly inside. As for N, there is a similar part around the eyelid that is often mistaken for new orthodontic students, and the system is prone to misrecognition. When it is cut out at the center, it is very similar as shown in FIG. However, a very different image can be obtained by bringing N to the middle of the upper side of the square of the first region.
[0025]
Then, 100 images are converted into PAP vectors and learning is performed to create a first template group consisting of 15 first vectors, and recognition is performed for 150 new X-ray images. Went. By using the cutting method described above, the recognition rate was improved from 72% to 85% and U1 from 92% to 97%. For N, 95% of the types of misrecognition mentioned above were eliminated and the recognition rate improved from 82% to 88%.
[0026]
Here, the input image does not necessarily need to be a head X-ray image, and the recognition target is not necessarily an anatomical feature point used by the orthodontist. It may be used for finding cancerous tissue in chest X-rays, or may be applied for use in extracting a facial region from a photographic image. The way to cut out is not necessarily defined in such a way that the recognition target point is the upper right and lower right of the region. There may be a position designation in units of pixels such as a few pixels from the bottom right to a few pixels.
[0027]
[Configuration of vector generator]
Here, the VLSI technology was implemented for the vector generation unit 1 in FIG. 2 (see Non-Patent Document 5). The circuit configuration diagram is shown in FIG.
A PAP (Principal Axis Projection) conversion VLSI is roughly divided into two blocks. First, an edge feature extractor 11 that extracts edge features from input two-dimensional image data and generates a feature expression flag is provided, and a vector generator 12 that receives the feature expression flag is provided. A PAP conversion VLSI was configured with such a configuration. A chip photograph is shown in FIG.
[0028]
The measurement result of this circuit is shown in FIG.
As shown in FIG. 7, it takes several minutes to convert a 64 × 64 pixel image into a 64-dimensional vector sequence for a workstation, but 340 msec. The function to calculate with is realized.
[0029]
-Configuration of storage unit and similarity evaluation unit-
A system function unit for storing the pattern group group in FIG. 2 is realized. This functional unit also includes a similarity calculator. However, such a configuration is not necessarily required.
[0030]
First, FIG. 8 shows a basic circuit that holds data for one element of a vector and performs similarity calculation (see Patent Document 2 and Non-Patent Documents 6 to 8).
As shown in FIG. 8, this basic circuit has an input voltage V _GG It has a function of reducing power consumption by changing the input voltage, and a function of changing the calculation evaluation method of the similarity calculator flexibly by changing the input voltages A, B, and C. This characteristic is shown in FIG. V _GG It can be seen that the current value of the peak is reduced by lowering the value of, and that evaluation functions having various sharpnesses are realized by changing the input voltages A, B, and C.
[0031]
In this basic circuit, one element of a template vector to be stored at the time of the first reset operation is inputted as a voltage. Then, one element of a vector for which similarity evaluation is performed is input as a voltage. Similarity information is converted to current and I _OUT Is output from. The higher the similarity is, the more current is output. This I _OUT Is used as the similarity evaluation value between the template information and the input vector information.
[0032]
This basic circuit realizes a function as shown in FIG.
First, a piece of knowledge is first stored as a voltage, and then a similarity with one element of a vector input as a voltage is output as a current value. The higher the similarity is, the more current flows.
[0033]
The function of this basic circuit is only the similarity calculation of one element of the vector, but by taking the sum of the output currents as shown in FIG. A circuit that outputs the degree can be easily realized. For example, in the case of a vector generated by the PAP method, since the number of dimensions of the vector is 64, 64 outputs may be connected. However, this number does not necessarily need to be 64, and the number changes according to the number of dimensions of the vector to be used.
[0034]
A functional block having a storage and similarity calculation function is realized as shown in FIG.
The circuits of FIG. 11 are arranged in parallel, and the input vector X is input to all the circuits simultaneously. With such a configuration, it is possible to simultaneously calculate all the similarities between the input vector and the template vectors in the plurality of pattern group groups simultaneously.
[0035]
An example of realizing such processing is shown in FIG.
In this system, the number of vector dimensions is four. The pattern stored in this circuit is shown in the upper part of FIG. The lower side shows the presented pattern group. The graph shows the similarity between the stored pattern and the presented pattern. The blue line indicates the theoretical value, and the red line indicates the measured value. When exactly the same pattern 7 is input, a large current flows, indicating a high degree of similarity. Also, when a pattern 1 similar to the pattern 7 is input, a very high similarity is shown. However, the low similarity is shown with respect to the pattern 6 which is not similar. Also, the power consumption is about 160 μA even when the current flows with the highest degree of similarity, which is realized with very low power consumption. Here, an example in which the number of dimensions is four is shown, but the number of dimensions is not necessarily four. If it is a vector generated by PAP, it becomes 64, and when other vector generation methods are used, it changes according to the number of dimensions of the vector.
[0036]
(Second Embodiment)
-Basic outline-
As a background of the present embodiment, there is a problem that if there is something that is actually similar to an object, the location is mistakenly recognized. Therefore, in order not to misrecognize such a similar thing, the object is to increase the recognition accuracy by putting a negative template into a negative template and performing a negative evaluation. In this embodiment, the basic recognition system is defined by removing the “position shift” functional configuration from the image analysis apparatus described in the first embodiment.
[0037]
FIG. 14 is a schematic diagram illustrating a schematic configuration of the image analysis apparatus according to the second embodiment.
Similar to FIG. 1 of the first embodiment, this image analysis apparatus finds a feature point representing a position of a specific part defined in advance in an image of a specific object from the input image, and calculates the coordinate value of the feature point. Storage means 101 for storing at least one first template group consisting of a plurality of first vectors representing the feature point for each feature point, and a predetermined one of the input images Cutting means 102 for cutting out a first area of a size of; a moving means 103 for moving the first area within a second area of a predetermined size in the target input image; The conversion means 104 for converting the partial image in the region of the image into a second vector according to a predetermined procedure, and performing a correlation operation between the first vector and the second vector, and similarities to each of the feature points 1st evaluation value representing degree Including an evaluation means 105 for determining, and identifying means 106 for identifying a third region corresponding to the first area including the feature points therein based on the first evaluation value. Further, for at least one feature point, an evaluation means 111 for obtaining a second evaluation value representing the degree of dissimilarity using a second template group that represents an image different from the feature point; A specifying unit 112 that uses both the evaluation value and the second evaluation value to specify the third region by a predetermined specifying method is included.
[0038]
Here, the specifying unit 112 specifies using, for example, fuzzy logic or information on the difference between the first evaluation value and the second evaluation value as a predetermined specifying method. In addition, it is preferable to use an image whose position is shifted from the target as the second template group.
[0039]
-Specific examples-
[Example 1]
FIG. 15 is a schematic diagram illustrating an image analysis apparatus according to the second embodiment.
The image analysis apparatus includes a first storage unit 11 that holds a first template group, and a second storage unit 12 that holds a second template group. The first template group is created by using a vector of anatomical feature points S in the same manner as the basic recognition system. Further, the second template group is created according to a method for creating the first template group in the basic recognition system by collecting partial images included in the third region identified by erroneous recognition in the basic recognition system. Here, the creation method does not necessarily apply the learning algorithm, and the number of the first template group and the second template group to be held may be different.
[0040]
The first region cut out from the second region is converted into a PAP vector representation by the first winner-take-all circuit 13, and the similarity evaluation circuit 15 performs matching with the first template group to evaluate the similarity. Calculate the value. At the same time, a vector expression such as S misrecognition is converted into a PAP vector expression by the second winner-take-all circuit 14, and matching with the second template group is performed by the similarity evaluation circuit 15. Calculate the evaluation value.
[0041]
Then, recognition is performed by holding (evaluation value of similarity) − (evaluation value of dissimilarity) × 0.1 as an evaluation value at the coordinate value (x, y).
As a result of recognizing 150 head X-ray images, the recognition rate of Sella was improved from 72% to 87%. Here, the vector expression is not necessarily a PAP vector expression. Further, the weight applied to the dissimilarity value in the evaluation value calculation method is not necessarily 0.1, but can vary depending on the object. It may be 0.2 or 0.5. The similarity calculation method is not always a simple weighting method. It is possible to use fuzzy logic.
[0042]
[Example 2]
When the recognition is performed by the basic recognition system, the transition of the evaluation value becomes a shape close to a trapezoid as shown in FIG.
Further, when recognition is performed by the basic recognition system, as shown in FIG. 16, the transition of the evaluation value becomes a shape close to a trapezoid, and there is a problem that accuracy of a target point is deteriorated.
[0043]
For the second template group, a partial image whose position is shifted by half from the target point as shown in FIG. 17 is used. The unit shifted here is not necessarily half.
[0044]
The partial image in the cut out first region is expressed as a PAP vector, matched with the first template group, and the similarity is calculated. Here, it is not necessarily said that matching is always performed by converting to the PAP vector representation.
[0045]
At the same time, matching with the second template group is performed to calculate the dissimilarity.
FIG. 18 shows the transition of the evaluation value when the similarity evaluation value is (similarity evaluation value) − (dissimilarity evaluation value) × 0.3. In this way, the evaluation value rises at the correct position and relatively decreases in the vicinity. Here, in the calculation of the evaluation value, the weight 0.3 multiplied by the dissimilarity value is not a fixed value and may change.
[0046]
(Third embodiment)
-Basic outline-
In the background of the present embodiment, when performing a search, there is a problem that the calculation cost of vectorization processing and matching processing increases in proportion to the expansion of the search area. Therefore, it is very important how to perform efficient recognition by appropriately limiting the search area.
[0047]
FIG. 19 is a schematic diagram illustrating a schematic configuration of the image analysis apparatus according to the third embodiment.
Similar to FIG. 1 of the first embodiment, this image analysis apparatus finds a feature point representing a position of a specific part defined in advance in an image of a specific object from the input image, and calculates the coordinate value of the feature point. Is output. This image analysis device finds a feature point representing the position of a specific part defined in advance in an image of a specific object from an input image, and outputs a coordinate value of the feature point. On the other hand, storage means 101 for storing at least one first template group composed of a plurality of first vectors representing the feature points, and a cutout for cutting out a first area of a predetermined size from the input image Means 102, specifying means 121 for specifying the second area using a predetermined area dividing method, moving means 103 for moving the first area in the second area, and a portion in the first area A conversion unit 104 that converts an image into a second vector according to a predetermined procedure, and a correlation operation between the first vector and the second vector are performed, and a first representing the similarity to each feature point Determine the evaluation value of That the evaluation means 105, and is configured and a specifying unit 106 for specifying a third region corresponding to the first area including the feature points therein based on the first evaluation value.
[0048]
The specifying unit 121 specifies using, for example, a third template group different from the first template group in the target input image. As the third template group, for example, a template group for finding information on the skull base, lower jaw, and upper jaw is used. Alternatively, at least one of the feature points whose position is determined may be specified by designating a region having a predetermined positional relationship from the feature point. Here, when the position of the predetermined feature point cannot be determined in the second region, it is preferable to enlarge the second region so that the second region includes at least a part of the adjacent region. It is.
[0049]
-Specific examples-
[Example 1]
An orthodonticist needs to recognize an image analysis apparatus in which a configuration of “region limitation” using the method of region segmentation of the second region is added to the “basic recognition system” from a cephalogram in FIG. It was applied to the recognition problem of the indicated anatomical feature points (S, Or, Gn). Here, the input image is not necessarily an X-ray photograph image, and the recognition target is not an anatomical feature point.
[0050]
As a first template group, the PAP method was applied to the vector representation from 100 X-ray photograph data, and 15 first vectors were created for each anatomical feature point by applying a learning algorithm. (That means that there are a plurality of first template groups.) Here, the number of first template groups held by the system is not necessarily 15 and may be 100 or 1000. In addition, a learning algorithm is not always used. Also, the conversion from the partial image to the vector representation need not be the PAP method.
[0051]
As a third template group, the learning algorithm is applied by converting the low resolution X-ray photograph data of 100 skulls, eyes, upper jaw, lower jaw and neck shown in FIG. After that, 15 template vectors were created by applying a learning algorithm. Here, the number of first template groups held by the system is not necessarily 15 and may be 100 or 1000. In addition, a learning algorithm is not always used. Also, the conversion from the partial image to the vector representation need not be the PAP method.
[0052]
First, a region is divided into the skull base, eyes, upper jaw, lower jaw, and neck using the fourth template group. Here, the region dividing method is not limited to the method shown here. An area such as a nose or a fixing device may be divided.
[0053]
The second region is recognized by performing recognition processing on the skull base region for Sella, the region surrounded by the skull base, eyes, and upper jaw region for Or, and the lower jaw region for Gn. This search area is not limited to this. Other areas can also be targeted for search.
[0054]
This method was applied to the problem of recognizing 150 new X-ray images. Since the search target is limited in advance, the search area is reduced and the calculation time is reduced by 80% on average. In addition, with respect to an example in which a different location is clearly recognized by the basic recognition system, the correct position is recognized. The recognition rate is 72% to 85% for S, 92% to 95% for Or, and 13% for Gn. To 45%.
[0055]
[Example 2]
Using the same image analysis apparatus as in Example 1, U1 (85%) and N (82%) can be recognized with high recognition accuracy, but S (72%) has relatively high recognition accuracy. Low. An example of the failure is shown in FIG. In addition, FIG. 22 shows two-dimensional representations of similarity evaluation values. Thus, a peak appears in the area even at the correct position. Therefore, it is necessary to efficiently limit the area. On the other hand, orthodontists have information on the average and standard deviation values regarding the distance and positional relationship of anatomical feature points.
[0056]
Therefore, in the present embodiment, recognition of Or and N is first performed. Then, as shown in FIG. 23, a circle having a radius of (average of the distance between U1 and S−standard deviation value) and a circle having a radius of (average of the distance of U1 and S + standard deviation value) is centered on U1. Determine the area enclosed. The same applies to N. By doing so, it was possible to limit the area and reduce the misrecognition. By applying such a region limiting method, the recognition rate of S was improved from 72% to 80%. However, it is not always the method used when S is recognized, and the distance from N and U1 is not always used for S recognition. There is a possibility of using an angle, and the reference point is not necessarily an anatomical feature point, and a point that the computer can easily recognize may be newly defined.
[0057]
[Example 3]
In the recognition process, the limitation of the search area (second area) is very important, but if the limitation fails and the target point does not enter the second area, erroneous recognition is performed. At that time, it is very important how to appropriately expand the area. In the method of the first embodiment, when a valid recognition target candidate is not found in the determined second area, the area outside the area adjacent to the determined second area by 20% is included in the second area. Recognized.
[0058]
As a result, the recognition rate was improved from 85% to 87% for S, 95% to 97% for Or, and 45% to 55% for Gn. The reason why the improvement rate is small is that almost correct area designation is performed in the first determination of the second area. The percentage of correct recognition by extending outward by 20% in the case of failure to determine the second region for the first time exceeded 90% for the three anatomical feature points. Here, the method of extending the second area is not necessarily limited to extending outward by 20%. It may be 40% or 50%. Also, the expansion method is not limited to simply expanding at an equal rate in all directions, and depending on the characteristics of the target point, it may be 50% horizontally and 10% vertically expanded.
[0059]
[Example 4]
In the method of the second embodiment, when a valid recognition target candidate is not found in the determined second area, the second area is averaged ± 2 * standard deviation as shown in FIG. Recognition was performed by re-designating the second region by value. The recognition conditions are the same as in the second embodiment.
[0060]
By applying such a method, the recognition rate S was improved from 85% to 89%. The reason why the improvement rate is small here is that almost correct area designation is performed in the first area determination. In the case of 90% of cases, correct recognition was performed by using the average value ± 2 * standard deviation when the first region determination failed. Here, the method of expanding the region is not necessarily limited to the average value ± 2 * standard deviation. There can be 0.5 * standard deviation and 1.5 * standard deviation.
[0061]
(Fourth embodiment)
-Basic outline-
In addition to the configuration of FIG. 1 of the first embodiment (including a so-called “position shift” function), this image analysis apparatus is one of a plurality of classes by a predetermined classification method for at least one feature point. And a calculation means for performing a correlation calculation with the second vector using a fourth template group different from the plurality of first template groups corresponding to the class. As a plurality of classes, for example, classification is performed using maxillary front protrusion, normal occlusion, and mandibular front protrusion. As this classification method, classification may be performed using soft tissue information. As an example of this classification method, classification is performed by performing a correlation operation with the second vector using a plurality of the first template groups.
[0062]
-Specific examples-
U1 and L1 recognized by orthodontists on cephalometric X-ray images vary considerably in the overall bone structure. Therefore, it is very difficult to recognize with a simple template matching technique. Therefore, first, as shown in FIG. 25, the bone structure is classified into two classes, that is, a maxillary front protrusion (extraction) and a mandibular front protrusion (receiving hole). The classification template is created from a low-resolution image in order to focus on rough overall structure information. 100 X-ray images respectively belonging to the tooth extraction and receiving mouth classes are prepared, converted into PAP vector expressions, and 15 tooth extraction and template class templates are created using a learning algorithm. Then, the class is identified by looking at which of the input image structures is close to.
[0063]
Here, the classification class is not limited to the tooth extraction and the receptacle. It is possible to add a normal class such as normal occlusion. In addition, the classification method may not use a low-resolution X-ray image, but may be classified and input by an orthodontist to determine and input, or may be classified based on information extracted from a profile of a profile photographic image. The recognition target is not necessarily a point on the orthodontic X-ray image. Further, the target is not necessarily a point on the head X-ray image. The number of X-ray images used when creating a template is not necessarily 100, and a learning algorithm is not necessarily applied. Further, the number of templates to be held is 15 and is not fixed.
[0064]
Then, U1 and L1 are recognized by using the fourth template group corresponding to each class by the method shown in FIG. The fourth template group is created by applying a technique for creating the first template group to X-ray image photographs belonging to each class. However, in order to increase the flexibility of recognition, there is a case where an intermediate product of both classes is put in each template group. Here, the recognition target is not limited to U1. In addition, the classification class is not limited to the tooth extraction and the receptacle. It is possible to add a normal class such as normal occlusion. The recognition target is not necessarily a point on the orthodontic X-ray image.
[0065]
By applying such a method, the recognition rate of U1 was improved from 92% to 95%, and L1 was improved from 72% to 90%.
[0066]
(Fifth embodiment)
-Basic outline-
In addition to the configuration of FIG. 1 of the first embodiment (including a so-called “position shift” function), the image analysis apparatus includes the resolution of the input image and / or the gradation of the pixel before generating the second vector. And changing means for changing at least one feature point. This changing means creates at least one fifth template group composed of a plurality of first vectors representing the feature points of the input image with resolution and / or pixel gradation.
[0067]
-Specific examples-
[Example 1]
FIG. 27 shows the result of recognizing anatomical features Gn, Go, and li recognized by the orthodontist using the basic recognition system. As shown here, similar points in the vicinity are misrecognized. In each figure in the lower stage, + indicates an accurate point, and an arrow indicates a point that is erroneously recognized.
[0068]
When a doctor recognizes an anatomical feature point, first, a rough position candidate is found using rough peripheral information of the target point. Subsequently, the candidate area is recognized with high accuracy. Therefore, as shown in FIG. 28, a first template group that converts 15 images including target points of standard resolution into vector representations by PAP and holds 15 first vectors by applying a learning algorithm. At the same time, a fifth template group holding 15 first vectors is stored in the same manner using an image having a resolution half that of the standard. First, rough recognition is performed using the fifth template group to obtain a candidate point group, and then high resolution recognition is performed using the first template group.
[0069]
Here, the recognition target is not necessarily an anatomical feature point. A technique for converting an image into a vector representation is not necessarily PAP. A template group is created from 100 images. However, the template group is not necessarily 100 images, and a learning algorithm is not always used. Further, although there is a template group holding 15 vectors, it is not necessarily 15 pieces. A low-resolution image is a half of the standard resolution, but not necessarily a half.
[0070]
A recognition example improved with respect to Gn is shown in FIG. Moreover, the recognition rate of Go, Gn, Ar, Cd, and Ba was improved by using multi-resolution search as shown in FIG.
[0071]
[Example 2]
With respect to the points Gn and Go on the hard tissue shown in FIG. 27, the points on the soft tissue corresponding to the correct anatomical feature points are erroneously recognized. Therefore, we have developed a method that significantly reduces the adverse effects of soft tissue information when recognizing points on hard tissue.
[0072]
When the image is cut out from the input image, the grayscale image has 256 gradations. FIG. 30 shows an example in which images in four directions are extracted by the PAP method with gradation down to 64 gradations. In this way, the information on the soft tissue is considerably extracted at 256 gradations, but the soft tissue information is considerably removed at 64 gradations.
[0073]
Here, the gradation is 64 gradations, but is not necessarily 64 gradations. Further, the method of extracting information after that is not necessarily PAP. Gn and Go were recognized using this method. A first template vector group for holding 15 first vectors was created using a learning algorithm by converting 100 images with gradations of 64 gradations into a vector representation using the PAP method.
[0074]
When this template group was used for 150 sheets, the recognition rate was improved from 13% to 97% for Gn and from 48% to 67% for Go as compared with the case where the number of gradations was 256. Here, the recognition target is not necessarily an anatomical feature point. A technique for converting an image into a vector representation is not necessarily PAP. A template group is created from 100 images. However, the template group is not necessarily 100 images, and a learning algorithm is not always used. Further, although there is a template group holding 15 vectors, it is not necessarily 15 pieces.
[0075]
In addition, this technique has a feature that it can be easily realized on a system, particularly on a VLSI system. For example, if gradation is reduced from 256 gradations to 64 gradations, it can be realized by dropping the lower 2 bits of the input image. Here, the gradation is not necessarily reduced to 64 gradations.
[0076]
(Sixth embodiment)
-Basic outline-
This image analysis apparatus has a function of binarizing an X-ray image when finding MPo as an anatomical feature point in addition to the configuration of FIG. 1 of the first embodiment (including a so-called “position shift” function). It has a configuration.
[0077]
-Specific examples-
[Example 1]
The anatomical feature point MPo recognized by the orthodontist is the uppermost part of the ear rod part that fixes the patient's head when X-ray imaging is performed. In this case, however, the recognition rate does not increase due to the addition of shooting conditions and surrounding information. Therefore, as shown in FIG. 31, the threshold value processing is performed with the pixel value 235 in which the ear rod portion clearly appears, and the recognition is performed after binarization.
[0078]
The recognition condition is that a 100-value binarized image is converted into a vector expression using the PAP method, and a first template group holding 15 first vectors is created by applying a learning algorithm. Then, when recognition was performed on 150 images, the recognition rate was improved from 48% to 75% by performing such preprocessing. However, the anatomical feature point to which such a method is applied here is not necessarily MPo, and the pixel value used as the threshold value is not necessarily fixed at 235. Further, the template group creation method is not limited to this.
[0079]
[Example 2]
Although the anatomical feature point S is similar as shown in FIG. 22, there is a tendency to misrecognize a completely different place. Therefore, we developed a technology to improve the recognition rate. S is the central point of the pituitary gland. Therefore, as shown in FIG. 32, whether or not there is a characteristic of the pituitary gland is examined for the candidate S and an evaluation using the certainty factor is performed.
[0080]
For each feature portion, a template group was created by the same method as that described in the image analysis apparatus of the first embodiment. The recognition rate of S was improved from 72% to 80% by conducting the certainty factor investigation using such partial information. However, the recognition target is not limited to anatomical feature points. Even in the case of an anatomical feature point, it is not necessarily S. Further, the partial features to be used are not necessarily the three shown in FIG. Other features may be used.
[0081]
Note that the basic skeletons of the first to sixth embodiments described above may be configured not only individually but also in combination of these two or all (six). For example, an image analysis apparatus that combines the three of the first to third embodiments includes the storage unit 101 to the specifying unit 106 in the first embodiment, and the position of the feature point in the third region. The evaluation unit according to the second embodiment has a configuration in which each feature point is uniquely determined (that is, shifted) with respect to the origin defined in advance in the third region. 111, the specifying means 112, and the specifying means 121 of the second embodiment are employed.
[0082]
(Other embodiments of the present invention)
Each means constituting the image analysis apparatus according to the first to sixth embodiments and examples described above, and each step (steps 1 to 6 and the like) of the image processing method are stored in a RAM or a ROM of a computer. This can be realized by operating the program. This program and a computer-readable storage medium storing the program are included in the present invention.
[0083]
Specifically, the program is recorded in a storage medium such as a CD-ROM, or provided to a computer via various transmission media. As a storage medium for recording the program, a flexible disk, a hard disk, a magnetic tape, a magneto-optical disk, a nonvolatile memory card, and the like can be used in addition to the CD-ROM. On the other hand, as the transmission medium of the program, a communication medium (wired line such as an optical fiber, etc.) in a computer network (LAN, WAN such as the Internet, wireless communication network, etc.) system for propagating and supplying program information as a carrier wave A wireless line or the like.
[0084]
In addition, the functions of the above-described embodiments are realized by executing a program supplied by a computer, and the program is used in cooperation with an OS (operating system) or other application software running on the computer. When the functions of the above-described embodiment are realized, or when all or part of the processing of the supplied program is performed by a function expansion board or a function expansion unit of the computer, the functions of the above-described embodiment are realized. Such a program is included in the present invention.
[0085]
For example, FIG. 33 is a schematic diagram illustrating an internal configuration of a general personal user terminal device. In FIG. 33, reference numeral 1200 denotes a computer PC. The PC 1200 includes a CPU 1201, executes device control software stored in the ROM 1202 or the hard disk (HD) 1211, or supplied from the flexible disk drive (FD) 1212. To control.
[0086]
【The invention's effect】
According to the present invention, when a similar image is processed, it can be recognized as similar image data, and a relatively complicated image, particularly an image typified by a dental medical X-ray image or the like. Even in the case of extracting features, an image analysis apparatus capable of accurately and quickly recognizing is realized.
[Brief description of the drawings]
FIG. 1 is a schematic diagram showing a schematic configuration of an image analysis apparatus according to a first embodiment.
FIG. 2 is a schematic diagram illustrating an image analysis apparatus according to the first embodiment.
FIG. 3 is a diagram showing X-ray photographs of L1 and U1, which are anatomical feature points.
FIG. 4 is a diagram showing how to cut out N in an X-ray photograph;
FIG. 5 is a circuit configuration diagram in which the vector generation unit is mounted in the VLSI technology.
FIG. 6 is a diagram showing a VLSI chip photograph of a vector generation unit.
FIG. 7 is a characteristic diagram showing a measurement result of a circuit in which the vector generation unit is mounted in the VLSI technology.
FIG. 8 is a basic circuit diagram that holds data for one element of a vector and performs similarity calculation;
FIG. 9 is a characteristic diagram of a basic circuit that holds data for one element of a vector and performs similarity calculation.
FIG. 10 is a characteristic diagram showing functions of a basic circuit that holds data for one element of a vector and performs similarity calculation;
FIG. 11 is an explanatory diagram showing an output current of a basic circuit that holds data for one element of a vector and performs similarity calculation;
FIG. 12 is an explanatory diagram showing functional blocks having storage and similarity calculation functions;
FIG. 13 is a characteristic diagram showing a calculation result by a similarity calculation circuit.
FIG. 14 is a schematic diagram illustrating a schematic configuration of an image analysis apparatus according to a second embodiment.
FIG. 15 is a schematic diagram illustrating an image analysis apparatus according to a second embodiment.
FIG. 16 is a characteristic diagram showing similarity evaluation values.
FIG. 17 is a diagram for explaining the transition of the similarity evaluation value using an X-ray photograph.
FIG. 18 is a characteristic diagram showing similarity evaluation values.
FIG. 19 is a schematic diagram illustrating a schematic configuration of an image analysis apparatus according to a third embodiment.
FIG. 20 is a diagram showing an X-ray photograph of an anatomical feature point to be recognized.
FIG. 21 is a diagram showing an X-ray photograph of a third template group.
FIG. 22 is a diagram illustrating an example of recognition failure.
FIG. 23 is a schematic diagram showing a region limiting method using an X-ray photograph.
FIG. 24 is a schematic diagram showing an example of limiting the second region.
FIG. 25 is a diagram showing a classification result using an X-ray photograph.
FIG. 26 is a diagram showing an example of hierarchical recognition by classification using X-ray photographs.
FIG. 27 is a diagram showing an example of recognition of anatomical feature points using an X-ray photograph.
FIG. 28 is a diagram for explaining multiresolution search using an X-ray photograph;
FIG. 29 is a characteristic diagram illustrating an example of recognition improvement using an X-ray photograph.
FIG. 30 is a diagram illustrating an example in which images in four directions are extracted by the PAP method with a gradation being lowered, using an X-ray photograph.
FIG. 31 is a diagram for explaining an example of recognition by binarizing an image using an X-ray photograph;
FIG. 32 is a diagram for explaining a partial feature of an anatomical feature point S using an X-ray photograph;
FIG. 33 is a schematic diagram showing an internal configuration of a personal user terminal device.
FIG. 34 is a diagram showing an example of a head X-ray image using an X-ray photograph.
[Explanation of symbols]
1 Vector generator
2 storage unit
3 Similarity calculation section
4 Winner take all circuit
11 First storage means
12 Second storage means
13 First winner take all circuit
14 Second winner take all circuit
15 Similarity evaluation circuit
101 Storage means
102 Cutting means
103 Transportation means
104 Conversion means
105,111 evaluation means
106, 112, 121 Identification means

Claims (51)

An image analysis device that finds a feature point representing a position of a specific part defined in advance in an image of a specific object from an input image, and outputs a coordinate value of the feature point,
Means for storing at least one first template group comprising a plurality of first vectors representing the feature points for each of the feature points;
Means for cutting out a first region of a predetermined size from the input image;
Means for moving the first region within a second region of a predetermined size in the target input image;
Means for converting the partial image in the first region into a second vector according to a predetermined procedure;
Means for performing a correlation operation between the first vector and the second vector, and determining a first evaluation value representing the degree of similarity for each of the feature points;
Means for specifying a third region corresponding to the first region including the feature point therein based on the first evaluation value;
Determining the position of the feature point in the third region using a relative positional relationship uniquely defined for each feature point with respect to an origin defined in advance in the third region. A characteristic image analysis apparatus. Means for obtaining a second evaluation value representing the degree of dissimilarity with respect to at least one feature point using a second template group representing an image different from the feature point;
The image analysis apparatus according to claim 1, further comprising means for specifying the third region by a predetermined specifying method using both the first evaluation value and the second evaluation value. . The image analysis apparatus according to claim 2, wherein the means for specifying the third region uses fuzzy theory as the predetermined specification method. The means for specifying the third region performs specification using information on a difference between the first evaluation value and the second evaluation value as the predetermined specification method. The image analysis apparatus described in 1. The image analysis apparatus according to claim 2, wherein an image whose position is shifted from a target is used as the second template group. The image analysis apparatus according to claim 1, further comprising means for specifying the second area using a predetermined area division method. The means for specifying the second region is specified by using a third template group different from the first template group in the target input image. Image analysis device. The means for specifying the second region is specified by specifying a region having a predetermined positional relationship from the feature point using at least one of the feature points whose position has been determined. The image analysis apparatus according to claim 6. When the position of the predetermined feature point cannot be determined in the second region, the second region is enlarged so as to include at least a part of the adjacent region. The image analysis apparatus according to claim 7 or 8. An image analysis device that finds a feature point representing a position of a specific part defined in advance in an image of a specific object from an input image, and outputs a coordinate value of the feature point,
Means for storing at least one first template group comprising a plurality of first vectors representing the feature points for each of the feature points;
Means for cutting out a first region of a predetermined size from the input image;
Means for moving the first region within a second region of a predetermined size in the target input image;
Means for converting the partial image in the first region into a second vector according to a predetermined procedure;
Means for performing a correlation operation between the first vector and the second vector, and determining a first evaluation value representing the degree of similarity for each of the feature points;
Means for identifying a third region corresponding to the first region including the feature point therein based on the first evaluation value;
Means for obtaining a second evaluation value representing the degree of dissimilarity with respect to at least one feature point using a second template group representing an image different from the feature point;
An image analyzing apparatus comprising: means for specifying the third region by a predetermined specifying method using both the first evaluation value and the second evaluation value. The image analysis apparatus according to claim 10, wherein the means for specifying the third region uses fuzzy theory as the predetermined specifying method. The means for specifying the third region performs specification using information on a difference between the first evaluation value and the second evaluation value as the predetermined specification method. The image analysis apparatus described in 1. The image analysis apparatus according to claim 10, wherein an image whose position is shifted from a target is used as the second template group. An image analysis device that finds a feature point representing a position of a specific part defined in advance in an image of a specific object from an input image, and outputs a coordinate value of the feature point,
Means for storing at least one first template group comprising a plurality of first vectors representing the feature points for each of the feature points;
Means for cutting out a first region of a predetermined size from the input image;
Means for specifying the second region using a predetermined region dividing method;
Means for moving the first region within the second region;
Means for converting the partial image in the first region into a second vector according to a predetermined procedure;
Means for performing a correlation operation between the first vector and the second vector, and determining a first evaluation value representing the degree of similarity for each of the feature points;
An image analyzing apparatus comprising: means for specifying a third area corresponding to the first area including the feature point therein based on the first evaluation value. The means for specifying the second region is specified by using a third template group different from the first template group in the target input image. Image analysis device. The means for specifying the second region is specified by specifying a region having a predetermined positional relationship from the feature point using at least one of the feature points whose position has been determined. The image analysis apparatus according to claim 14. When the position of the predetermined feature point cannot be determined in the second region, the second region is enlarged so as to include at least a part of the adjacent region. The image analysis apparatus according to claim 15 or 16. The means for storing the first template group creates the first vector constituting the first template group by performing learning always using the determined relative positional relationship. The image analysis apparatus according to any one of claims 1 to 17. The image analysis apparatus according to claim 18, wherein the means for storing the first template group uses edge information of image information when performing the learning. 20. The image analysis apparatus according to claim 18, wherein the means for storing the first template group uses a median value of image information when extracting the edge information. 21. The method according to claim 1, further comprising means for changing a resolution and / or pixel gradation of the input image with respect to at least one of the feature points before generating the second vector. The image analysis device according to any one of the above. The means for changing the resolution and / or pixel gradation includes at least one of the first vectors representing the feature points of the input image having the resolution and / or pixel gradation. The image analysis apparatus according to claim 21, wherein a fifth template group is created. The image analysis apparatus according to claim 1, wherein the shape of the first region is different from that of the other feature points with respect to at least one feature point. Classifying at least one feature point into one of a plurality of classes by a predetermined classification method, and using a fourth template group different from the plurality of first template groups corresponding to the class The image analysis apparatus according to claim 1, further comprising means for performing a correlation operation with the second vector. 25. The image analysis apparatus according to claim 24, wherein classification is performed by performing a correlation operation with the second vector using the plurality of first template groups as the classification method. 26. The image analysis apparatus according to claim 1, wherein the means for converting into the second vector uses edge information of image information. The means for converting to the second vector uses a median value of information in a fourth area of a predetermined size in the first area when extracting the edge information. Item 27. The image analysis device according to any one of Items 1 to 26. 28. The image analysis apparatus according to claim 1, wherein the means for converting to the second vector uses an integrated circuit during the processing. 29. The means for determining the first evaluation value uses an integrated circuit during a correlation calculation process between the first vector and the second vector. The image analysis apparatus according to any one of the above. 30. The image analysis apparatus according to claim 1, wherein a medical X-ray image is used as the input image, and a predetermined feature point medically defined in advance is found as the feature point. The image analysis apparatus according to claim 30, wherein the X-ray image is for dentistry. 32. The image analysis apparatus according to claim 31, wherein the predetermined feature point includes a point on a soft tissue. The image analysis apparatus according to claim 7 or 15, wherein a template group for finding out information on the skull base, the lower jaw, and the upper jaw is used as the third template group. 26. The image analysis apparatus according to claim 24, wherein classification is performed using a maxillary frontal protrusion, a normal occlusion, and a lower frontal protrusion as the plurality of classes. The image analysis apparatus according to claim 34, wherein classification is performed using soft tissue information as the classification method. 36. The image analysis apparatus according to claim 24, 25, 34, or 35, wherein a template group for maxillary protrusion, normal occlusion, and mandibular protrusion is used as the fourth template group. 37. The image analyzing apparatus according to claim 31, wherein Point A is found as the predetermined feature point. 37. The image analysis apparatus according to claim 31, wherein Point B is found as the predetermined feature point. 32. The image analysis apparatus according to claim 31, wherein a Nasion is found as the predetermined feature point. 40. The method according to claim 39, wherein when finding a Nation as the predetermined feature point, the first region including the Nation uses a pixel value that is close to homogeneity with a white bone portion and a black background portion. Image analysis device. 32. The image analysis apparatus according to claim 31, wherein Sella is found as the predetermined feature point. 42. The image analysis apparatus according to claim 41, wherein when the Sella is found as the predetermined feature point, it is first found at a predetermined resolution, and is found with higher resolution and higher accuracy than the predetermined resolution. 43. The image analysis apparatus according to claim 41, wherein when the Sella is found as the predetermined feature point, the Sella is recognized using a partial feature of the Sella. 32. The image analysis apparatus according to claim 31, wherein an Orbitale is found as the predetermined feature point. 32. The image analysis apparatus according to claim 31, wherein MPo and / or APo is found as the predetermined feature point. 46. The image analysis apparatus according to claim 45, wherein when the MPo and / or APo is found as the predetermined feature point, the dental X-ray image is binarized and found. 37. The image analysis apparatus according to claim 31, wherein L1 is found as the predetermined feature point. 48. The image analysis apparatus according to claim 47, wherein tooth enamel information is used when L1 is found as the predetermined feature point. 37. The image analysis apparatus according to claim 31, wherein U1 is found as the predetermined feature point. The image analysis apparatus according to claim 49, wherein tooth enamel information is used when U1 is found as the predetermined feature point. 50. The method according to claim 49, wherein when U1 is found as the predetermined feature point, the target point is set as an upper right region in the third region as a relative positional relationship uniquely determined for each feature point. The image analysis apparatus described.

Images