JPH07220090A - Object recognition method - Google Patents

Abstract

PURPOSE:To perform pattern recognition based on the limited number of feature elements and the relative arranging information of the feature element. CONSTITUTION:An inputted image is held by recording at an image input part S11, and a local feature element in the image is extracted at a local feature element extraction part S12, and the arranging information of the local feature element is generated at an extracted feature element arranging data generating part S13, and the combination arranging information of the local feature element is stored in a local feature element model arranging data storage part S14 as storage information, and the generated arranging information of the local feature element and storage information are judged by collating at a matching processing part S15, then, an existence area in the image of judged recognition information can be detected and extracted at an adaptive image area extraction part S16.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method for picking up an image of a specific subject and editing the image.

[0002]

2. Description of the Related Art As one of conventional display methods of graphic patterns, there are various methods such as the graphic input method of Japanese Examined Patent Publication No. 5-36830 in which a curved portion of a stroke of a geometric graphic symbol is prepared in advance. A method is known in which a curved portion is represented by one of bending patterns (shape primitives) and a curved portion is approximated by an arc, and it can be applied to geometrically simple figure recognition.

Further, as a method of object recognition, Japanese Patent Publication No. 5-
In the object recognition device of No. 23463, the contour of the recognized object is traced, divided into shape primitives such as a straight line portion or a circular arc portion, and the respective attributes and the attributes of each vertex are registered in the memory as a dictionary, and the dictionary memory is stored. First, recognition is performed by searching each shape primitive of the unknown object.

[0004]

However, since the spatial arrangement relationship of the shape primitives is not extracted for recognition in the above-mentioned conventional example, even if the same object has different image patterns depending on the viewpoint position. Or, if the shape or size changes due to some factor, it is necessary to store a huge number of two-dimensional pattern information for the same object and to perform matching with the huge number of pattern information at the time of pattern recognition. However, there is a problem in that the calculation cost becomes large.

Further, generally, when a plurality of objects exist in an image, it is necessary to appropriately perform area division beforehand, and the area division is performed so that only one object exists in one area before recognition. The treatment was applied.

The area division and the recognition of the target pattern are two sides of the same face, and it is very difficult to perform them completely automatically.

In view of such a point, an object of the present invention is to perform pattern recognition based on a limited number of characteristic elements and relative arrangement information of the characteristic elements.

[0008]

According to an object recognition method of the present invention, an input image is recorded and held, local feature elements in the image are extracted, and arrangement information of the local feature elements is generated. Then, the combination arrangement information of the local feature elements of the object to be recognized is stored as storage information, and the arrangement information of the generated local feature elements and the storage information are collated and determined, and the determined recognition is performed. The existing area of information in the image is determined and extracted.

The object recognition method of the present invention records and holds an input image, extracts local feature elements in the image, and extracts the color, local spatial frequency, intensity, etc. of the local feature element neighborhood region. Region-based information is extracted, placement information of the local feature element and the region-based information is generated, and combined placement information of the local feature element of the object to be recognized is stored as storage information and generated. The placement information and the stored information are collated and determined.

The object recognition method of the present invention records and holds an input image, extracts a local feature element in the image, and extracts a model graphic element of the local feature element of the object to be recognized. An intermediate graphic element of the local feature element is extracted from the extracted local feature element and the first stored information stored as the first memory information, and the arrangement information of the intermediate graphic element is generated. , Storing the combination layout information of the model graphic elements of the object to be recognized as second storage information, and determining by comparing the generated layout information of the intermediate graphic elements with the second storage information. .

As the local feature elements, an intersection pattern of edge segments in a plurality of directions, all or a part of a curve having a constant curvature, and edge segments are extracted.

The arrangement information of the local feature elements is represented as a two-dimensional array or a three-dimensional array of digitized elements to which numerical values discretized by a predetermined method are assigned to the local feature elements.

The combination arrangement information of the local feature elements is represented by a pattern of feature elements obtained by rearranging the extracted local feature elements on a lattice space constituted by a predetermined size and a predetermined shape unit.

The process of extracting the local feature element is performed for each of a plurality of scaling parameters having different sizes.

[0015]

[Action]

(A) A local feature element in the input image is extracted to generate layout information, and the recognition information is determined by collating it with the combination layout information of the local feature elements of the object to be recognized, which is stored in advance. The area in which the recognition information exists is determined and extracted. At this time, as local feature elements, an intersection pattern of edge segments in a plurality of directions, all or part of a curve having a constant curvature, and edge segments are extracted for each of a plurality of scaling parameters having different sizes. Also, the arrangement information of the local feature elements is represented as a two-dimensional array of digitized numerical elements of the local feature elements. Furthermore, the combination arrangement information of the local feature elements,
The extracted local feature elements are represented by a pattern of feature elements obtained by rearranging the extracted local feature elements on a lattice space composed of a predetermined size and a predetermined shape unit. By the above method, the memory capacity required for the recognition target image data is saved,
The efficiency of recognition processing can be improved.

(B) Object information from an arbitrary viewpoint position of the same object corresponding to a change in the viewpoint position with respect to an image and image pickup by expanding the arrangement information of the local feature elements into a three-dimensional array of digitizing elements At the time of object recognition corresponding to changes in illumination conditions, the types of local feature elements to be extracted do not change sensitively, and it is possible to perform object recognition less susceptible to the deformation of objects in the image.

(C) Local feature element By extracting region base information such as color, local spatial frequency, intensity, etc. of a region near the local feature element and generating arrangement information of the local feature element and the region base information, Even if there are multiple objects and there is a strong deformation such as the original shape of the object being lost or hidden due to parts of the multiple objects overlapping or touching each other, robust without region division in advance. Can recognize.

As a result, which position in the image is to be recognized and which target should be recognized are output, and an image centering on that position or a partial image centering on the target image is extracted from the original image to identify the specific target. Image editing such as centering imaging or combining images containing specific objects with other images,
It is possible to output information necessary for efficient and robust operation.

(D) By extracting the intermediate graphic element of the local feature element and generating the arrangement information of the intermediate graphic element, recognition based on the hierarchical feature extraction can be performed, and a plurality of objects can be recognized by each other. Even in images captured by overlapping, robust recognition that is not easily affected by the influence can be performed.

[0020]

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

FIG. 1 is a block diagram of the processing unit in the first embodiment of the present invention. In FIG. 1, the image input unit S ₁₁ is
Image data obtained by the image pickup means is recorded and held in a predetermined recording medium. The local feature element extraction unit S ₁₂ determines a finite number of local feature elements of a plurality of scales (sizes) preset by the scaling parameter σ in each region in the image, for example, an intersection pattern of various edge segments (L type). , T-shaped, X-shaped, Y-shaped intersections, etc.), and local feature element patterns composed of line segments (curve segments) such as curved line segments having various curvatures (constant) and orientations are extracted and extracted. Data other than local feature elements are not stored. The extraction feature element array data generation unit S ₁₃
From the image data of the local feature elements extracted in S ₁₂ , each local feature element is converted into a predetermined two-dimensional array structure (cell array) by a predetermined data format, and the array data maintaining an approximate arrangement relationship. To generate.
In addition, the local feature element model array data storage unit S ₁₄
Model array data (plural possible) is stored as a local feature element pattern of an image to be recognized. The model array data stored in S ₁₄ is used as a kind of template in the matching processing unit S ₁₅ . S ₁₅ evaluates the error amount represented by the sum of squares of the difference between the S ₁₃ sequence data and the S ₁₄ model sequence data, and recognizes the model sequence data such that the error amount is equal to or less than a threshold as a recognition pattern. Is determined as. Further, the matching image area extraction unit S ₁₆ determines and extracts the existing area of the recognition pattern in the original image.

The processing contents of each processing unit after S ₁₂ will be described below.

FIG. 2 is a diagram showing an example of the extracted local feature element pattern. As a method of extracting a crossing pattern of edge segments which are local feature elements to be extracted in S ₁₂ , Deriche, R., Giraudon, G. (1993) (Internationa
l Journal of Computer Vision, Vol.10, 101-124), Ro
hr, K. and Schnoerr, C. (1993) (Image and VisionCo
mputing, Vol.11, 273-277), Iso, Shizawa (1993) (Technical Bulletin, Vol.IE92-125, pp.33-40). is not. In FIG. 2, a finite number of different directions (L ₁ , L ₂ ,
......,

[0024]

[Outer 1] Elements (8 here). Crossing angle β is 0 ° <β
<180 °, not the type of L-shaped intersection depending on the intersection angle, but the direction of the L-shaped intersection (diagonal angle line direction of the intersection)
Are divided into 8 types. For T-shaped, X-shaped, Y-shaped, and arrow-shaped intersections obtained by combining L-shaped intersections,
It can be extracted by the method presented in Rohr, K. and Schnoerr, C. (1993).

In addition, another local characteristic element, which is constant curvature,
Koenderink, J. and R
ichards, W. (1988) (J. Opt. Soc. Am. A, Vol.5, pp.
1136-1141), Li, S. Z. (1990) (International Journa
l of Computer Vision, Vol.5, pp.161-194) etc.
Has been done. In FIG. 2, a curved element with a constant curvature,
That is, the direction of the arc is the inward normal vector at its midpoint.
Finite number (C _v1, C_v2, ……,

[0026]

[Outside 2] Elements (8 here).

Further, a finite number of scaling parameters σ (for example, σ = 2, 4, 8, 16, 32 pixels) at the time of extracting the intersection pattern or the curvature element are set, and each scaling parameter is set. Extract local feature elements. This σ is the smoothing (for example, the smoothing performed when the intersection pattern or the curvature element is extracted).

[0028]

[Equation 1] Of the Gaussian function and the convolution operation)). S ₁₂ includes up process of encoding by extracting the closest among the local feature elements set in advance.

FIG. 3 shows a face using the local feature elements of FIG.
It is a figure which shows the example of encoding of an image, S ₁₂The face image
It is encoded with the scaling parameter σ.

Next, in S ₁₃ , the spatial arrangement relationship of the coded local feature elements is expressed by mapping on a lattice space composed of cells of a preset size and shape. FIG. 4 is a diagram showing an example of the encoded local feature element array display grid space. In FIG. 4, the lattice space is set to 8 of N, E, W, S, NE, NW, SW, and SE.
The grid size of the rectangle is set to the same degree as the scaling parameter σ. In this way, by reconstructing the rough arrangement relationship between local feature elements,
Obtain an image representation that is invariant to deformations of the original image. Furthermore, by extracting such an invariant expression form of the image for each scaling parameter σ, the image to be recognized is
It can be stored in advance as a common local feature element pattern model that does not depend on the spatial size of the relative arrangement relationship between the encoded local feature elements.

As described above, the first embodiment expresses an image by a combination of a smaller number of preset local feature elements and a limited number of spatial arrangement relations in a matrix form, thereby performing the object recognition process. This makes it possible to improve the efficiency (that is, reduce the calculation cost) and recognize an object that is not easily affected by the size change and deformation of the image of the object.

Next, the encoding of array data necessary for recognizing the local feature element array mapped in the lattice space will be described. In the first embodiment, in S ₁₅ , the recognition is executed by matching the model array data with the extracted feature element array data generated from the actual image. However, in order to execute this recognition by the numerical operation of the computer, the local feature element is used. It is necessary to quantify each pattern of. Therefore, assuming that the total number of local feature elements is M, each local feature element is numbered from 1 to M, for example. The numbering method is not particularly limited, but it is desirable that local feature elements (for example, L-shaped intersections with different orientations) in the same category be consecutive or have similar values. The value of the cell (array) in which the local feature element does not exist may be 0 or a value other than the numbered numbers. An ordinary template matching method may be used as an example of the recognition processing after the numerical encoding of the local feature element. However, the model array data is different from the conventional template-based method in that it does not depend on the image size. That is, when the local feature elements are coded with the scaling parameters σ ₁ , σ ₂ , ..., σ _n from the image to match with the model array data, the lattice size of the model array is virtually calculated from the actual image. Reduce or expand to match the grid size of the extracted data. Therefore, it is not necessary to prepare model array data of local feature elements of an image to be recognized for different grid sizes.

For example, in the case of recognizing a face image, size-invariant model mask data is created in advance for the eyes, mouth, etc., which are the parts required for recognition, by local feature elements such as L-shaped intersections and curve elements. In addition, size-invariant model array data that retains the relative positional relationship between the eyes and mouth (however, is reduced or expanded according to the scaling parameter σ when extracting local feature elements) is stored as a mask pattern, and the local feature is stored. Each area of the image after element extraction is scanned, and the degree of matching with the model array data is calculated by the least square method or the like. Ie

[0034]

[Outside 3] Is a cell value (corresponding to a local feature element) at a position on the lattice space (i, j) normalized by the scaling parameter σ,

[0035]

[Outside 4] On the lattice space with the scaling parameter σ (i, j)
Given the value of the cell at the position of, the calculation process of recognition is
For example, it is defined by Equation 2.

[0036]

[Equation 2] By determining the position (k, p) where F (k, p) becomes a minimum value (or a maximum value) below (or above) a predetermined threshold value, which position in the original image should be recognized. Is output. Where ‖x and y‖ are (x−
The absolute value of y) or (xy) ²ⁿ (n = 1, 2, ...)
It is desirable that the non-negative value is an even function related to (xy). In this case, when ‖x and y‖ are below the threshold,
Recognize that y is x. Also, J indicates the range of array elements occupied in the lattice space of the object to be recognized, and is standardized as (i = 1, 2, ..., q; j = 1, 2, ..., R). You can set it.

As the function F (k, p),

[0038]

[Outside 5] When

[0039]

[Outside 6] The correlation with may be calculated. In this case, model array data of q × r block size

[0040]

[Outside 7] The extracted data from the image

[0041]

[Outside 8] While scanning the upper part, (k, p) at which the correlation value becomes the maximum when it is equal to or larger than the threshold value is obtained. Furthermore, the processing result may be output to an image pickup system or an image editing system centering on an object to be recognized to perform a desired functional operation.

FIG. 5 is a structural diagram of a three-dimensional lattice space in the second embodiment of the present invention. In the second embodiment, the three-dimensional spatial arrangement relationship of the local feature elements described in the first embodiment is extracted and modeled. As a method of stereoscopic measurement, a method of extracting corresponding points by image processing from a real image (for example, a stereo photograph obtained by taking images with two cameras separated by a predetermined value) and a reflected light by irradiating a laser beam There is a method of measuring the phase of, or a method of projecting a structural pattern (such as a mesh pattern) and measuring the degree of deformation thereof.

In FIG. 5, the cell shape of the lattice space is a rectangle obtained by equally dividing the spherical surface in the longitude and latitude directions, but another shape unit (for example, a triangle) is used for another solid (for example, a cylinder). It may be obtained by dividing. Like this 3
The dimensional lattice space can be applied when recognizing an image from an arbitrary viewpoint position of a target object. That is, even for the same object, an image obtained from a certain viewpoint is generally different from an image obtained from another viewpoint, and an image pattern when changing the viewpoint position can be obtained only from one two-dimensional image. It is difficult to predict changes, and it is almost impossible to record images from all viewpoint positions and use them for recognition. However, the three-dimensional spatial arrangement relationship of a limited number of local feature elements is mapped to a representative point (one point on the lattice space) that is three-dimensionally discretized as the matching model data, and the same domain (lattice By measuring the degree of matching with an actual image in (space), it is possible to dramatically improve the efficiency of processing and memory saving required for stereoscopic image recognition from an arbitrary viewpoint position.

In the second embodiment, a local feature element is unique to each cell in a region (corresponding to a range visible when the object to be recognized is seen from a certain viewpoint) consisting of a finite number of cells covering the spherical surface. The model array data of the n × m array block obtained by setting various numerical values (or symbols) is scanned on the array data of the N × M array block (N> n, M> m) from the actual image. The same matching process as in the first embodiment is performed.

FIG. 6 is a block diagram of the processing unit in the third embodiment of the present invention. In FIG. 6, the image input unit S ₆₁ , the local feature element extraction unit S _62a , and the matching processing unit S ₆₅ are
The same processes as S ₁₁ , S ₁₂ , and S _{15 of} FIG. 1 are performed, respectively.
In the area information extraction unit S _62b , similarly to S _62a , for each block having a size corresponding to the scaling parameter σ,
Area information such as representative color, average intensity, and local spatial frequency of the neighborhood area including local feature elements is extracted. Image input from the S _61, the predetermined processing is performed in the S _62a and S _62b. The array data generation unit S ₆₃ uses S _62a and S _62a.
Sequence data is generated from the local feature elements and area information extracted by _62b . The model array data storage unit S ₆₄ stores model array data in which local feature elements and area information are extracted for each block of an image to be recognized which is divided in advance by rectangular blocks according to the scaling parameter σ.

Hereinafter, the color will be taken as an example of the area information, and the description will be limited to the two-dimensional image recognition. The color vector defined below is used as the representative color for each block to be extracted.

[0047]

[Outside 9] To use.

[0048]

[Equation 3] here

[0049]

[Outside 10] Represents the output intensity of the R pixel of the sensor at pixel position (i, j) in the image,

[0050]

[Outside 11] Similarly represents the output intensity of the G pixel and the B pixel. symbol

[0051]

[Outside 12] Indicates addition of pixel values for each block, and is performed over all pixel positions (i, j) in the same block.

In this way, the array data for the recognition process is generated in S ₆₃ based on the local feature elements extracted for each scaling parameter σ in S _{62 a} and S _{62 b} and the region information such as the representative color. .

The model array data stored in S ₆₄ is for the local feature elements.

[0054]

[Outside 13] For area information

[0055]

[Outside 14] Numerical data unique to the local feature element or representative color in each array. For example, for the color of the block at position (k, p)

[0056]

[Equation 4] You can display a two-dimensional vector like

[0057]

[Outside 15] May be used. The local feature elements are as described in the first embodiment.

As a first method of matching between model array data and array data extracted from an image, that is, a recognition process, first, matching is performed on the basis of area information (color), and then color is used to approximate similarity. A method may be used in which matching is performed on the basis of a local feature element with respect to a region (block) in which a defect has occurred. As a second method, it is possible to reverse the order of matching, first to extract regions where similar correspondence can be obtained on the basis of local feature elements, and then to narrow down the similar correspondence on a color basis for each region. . As a third method, a total evaluation function f f = is obtained by adding a matching evaluation function f _F based on a local feature element and a matching evaluation function f _A based on a region information base with appropriate weights λ. _A position may be obtained such that the value of f _F + λf _A (1) is equal to or less than a predetermined threshold value. However, in the first and second methods, "matching" means model data as described in the first embodiment.

[0059]

[Outside 16] and

[0060]

[Outside 17] And extracted data from the actual image

[0061]

[Outside 18] Appropriate evaluation function for

[0062]

[Equation 5] Is to obtain (k, p) that is less than or equal to a predetermined threshold. Note that ‖x and y‖ are the functions presented in the first embodiment.

Even when a plurality of objects are present in the image by combining the local feature element information and the area information and some of the plurality of objects overlap each other, the areas are divided in advance. It is possible to perform recognition without causing only one object to exist in one area.
FIG. 7 is an explanatory diagram of three regions when the T-shaped intersection is caused by the shielding. In FIG. 7, a T-shaped intersection is detected with a size larger than other local feature elements in the image, and three pieces of area information A ₇₁ contacting the T-shaped intersection with that size are detected.
Attributes of A _72, A ₇₃ (e.g. color), A ₇₂ and A ₇₃ in the like case substantially equal but for very different from the A _71, the object corresponding to the A ₇₂ and A ₇₃ by A ₇₁ is partially shielded When recognizing the image in the area including A ₇₂ and A ₇₃ near the T-shaped intersection, it is possible to use A ₇₁ from the actual image data when matching with the model array data. If a region having the same attribute as A ₇₁ is excluded, or if the threshold level is recognized by detecting the minimum value of the error, the threshold value is increased by a predetermined value, and if the correlation is determined, the value is decreased by a predetermined value. By adding the processing to S ₆₅ , it is possible to perform recognition that is not based on area division.

FIG. 8 is a block diagram of the processing unit in the fourth embodiment of the present invention. 8, the image input unit S ₈₁ , the local feature element extraction unit S ₈₂ , the array data generation unit S ₈₃ , and the matching processing unit S ₈₅ are the same as S ₁₁ , S ₁₂ , S ₁₃ , and S _{15 of} FIG. Perform processing. The intermediate graphic element extraction unit S ₈₇ forms a part of the image of the object and has a meaning as a graphic concept, that is, an intermediate graphic element is extracted. The model graphic element storage unit S ₈₈ stores in advance the model graphic element of the intermediate graphic element. The model graphic element array data storage unit S _{84 previously} stores the model graphic element array data for matching with the array data of S ₈₃ .

In the fourth embodiment, after extraction in S ₈₂ , in S ₈₇ , areas corresponding to eyes, nose, mouth, eyebrows, ears, etc. in the face image are extracted as intermediate graphic elements. . The extracted intermediate graphic element belongs to a hierarchically middle-level local feature element that constitutes a more complex upper-level image pattern, such as an entire face, and is extracted in the first to third embodiments. The characteristic feature element can be positioned as a lower-level local feature element, and expresses an intermediate graphic element by the spatial arrangement relation in the lattice space.

After the model graphic elements such as eyes and mouth stored in advance in S ₈₈ are extracted in S ₈₇ based on the spatial arrangement of the local feature elements extracted in the lower level in S ₈₂ , the medium level is extracted in S ₈₃ . The array data in is generated by numerical data or symbols unique to each intermediate graphic element.

FIG. 9 is a diagram showing an example of encoding a face image by a part of intermediate graphic elements. In FIG. 9, the intermediate graphic element extracted from the image in S ₈₇ and the model graphic element to be recognized extracted in S ₈₄ are matched in S ₈₅ so that a plurality of objects overlap each other. Even in a captured image, it is possible to perform robust recognition that is unlikely to be affected by it. That is, in face image recognition, intermediate graphic elements such as eyes, nose, and mouth are extracted as preprocessing, and relative positions are encoded in a lattice space as shown in FIG. 9 (here, eyes are 9, Nose is 5,
Although the mouth is quantified as 1, the spatial arrangement of other intermediate graphic elements is inconsistent with the structure of the face image even if any of these elements in the face is an image missing due to the above-mentioned factors. If you don't, you can recognize it as a face.

In the detection of the above-mentioned consistency in the fourth embodiment, it is determined that the matching with the model array data on the grid space at the intermediate graphic element level is the maximum value (or the minimum value) when it is the predetermined threshold value (or less). Is equivalent to detecting such a position.

[0069]

As described above, the present invention has the following effects.

The local feature elements in the input image are extracted to generate the placement information, and the recognition information is determined by collating with the combination placement information of the local feature elements of the object to be recognized which is stored in advance. The area in which the recognition information exists is determined and extracted. At that time, as local feature elements, an intersection pattern of edge segments in a plurality of directions, all or part of a curve having a constant curvature, and an edge segment,
It is extracted for each of a plurality of scaling parameters having different sizes. Also, the arrangement information of the local feature elements is represented as a two-dimensional array of digitized numerical elements of the local feature elements. Further, the combination arrangement information of local feature elements is represented by a pattern of feature elements obtained by rearranging the extracted local feature elements on a lattice space configured by a predetermined size and a predetermined shape unit. The above method has an effect that the memory capacity required for the image data to be recognized can be reduced and the efficiency of the recognition process can be improved.

That is, as shown in the first embodiment, the object recognition processing is performed by expressing the image by a combination of a smaller number of preset local feature elements and a limited number of matrix-like spatial arrangement relationships. The efficiency of (1) (that is, the reduction of the calculation cost) and the object recognition that is not easily affected by the change and deformation of the size in the image of the object are enabled.

Further, when the arrangement information of the local feature elements is expanded to a three-dimensional array of digitizing elements, the object recognition from an arbitrary viewpoint position of the same object corresponding to the change of the viewpoint position with respect to the image and at the time of imaging In the object recognition corresponding to the change of the illumination condition, the type of the local feature element to be extracted does not change sensitively, and it is possible to perform the object recognition that is not easily affected by the deformation of the object in the image.

That is, as shown in the second embodiment, the three-dimensional spatial arrangement relationship of the limited number of local feature elements is mapped to the three-dimensionally discretized representative point (one point on the lattice space). Is used as the matching model data, and by measuring the degree of matching with the actual image in the same domain (lattice space), the processing efficiency required for recognizing a stereoscopic image from an arbitrary viewpoint position and the memory saving can be dramatically improved. Can be planned.

Further, by extracting the region-based information such as the color, local spatial frequency, and intensity of the region near the local feature element and generating the arrangement information of the local feature element and the region-based information, a plurality of images can be displayed in the image. Even if there is a strong deformation such as the original shape of the object being lost or hidden due to the fact that some of the plurality of objects overlap each other or come into contact with each other, as shown in the third embodiment, This has an effect that robust recognition can be performed without performing region division.

As a result, which position in the image is to be recognized and which target should be recognized are output, and an image centering on that position or a partial image centering on the target image is extracted from the original image to identify the specific target. Image editing such as centering imaging or combining images containing specific objects with other images,
This has the effect of being able to output the information necessary for efficient and robust operation.

In addition, by extracting the intermediate graphic element of the local feature element and generating the arrangement information of the intermediate graphic element, the recognition based on the hierarchical feature extraction can be performed,
Even in images captured by overlapping a plurality of objects with each other, it is possible to perform robust recognition that is not easily influenced by the images.

That is, as shown in the fourth embodiment, an intermediate figure element is extracted as a preprocessing for recognition and the relative position is coded and expressed in the grid space, and any one of these elements is missing due to the above-mentioned factors. Even an image can be recognized if the spatial arrangement of other intermediate graphic elements does not conflict with the configuration of the object to be recognized.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a processing unit according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of extracted local feature element patterns.

FIG. 3 is a diagram showing an example of encoding a face image using the local feature elements of FIG.

FIG. 4 is a diagram showing an example of an encoded grid space for displaying a local feature array.

FIG. 5 is a structural diagram of a three-dimensional lattice space in the second embodiment of the present invention.

FIG. 6 is a configuration diagram of a processing unit in a third embodiment of the present invention.

FIG. 7 is an explanatory diagram of three regions when a T-shaped intersection is caused by shielding.

FIG. 8 is a configuration diagram of a processing unit in a fourth embodiment of the present invention.

FIG. 9 is a diagram showing an example of encoding a face image by a part of intermediate graphic elements.

[Explanation of symbols]

S ₁₁ , S ₆₁ , S ₈₁ Image input section S ₁₂ , S _62a , S ₈₂ Local feature element extraction section S ₁₃ Extracted feature element array data generation section S ₁₄ Local feature element model array data storage section S ₁₅ , S ₆₅ , S ₈₅ Matching processing unit S ₁₆ Compatible image region extraction unit S _62b Region information extraction unit S ₆₃ , S ₈₃ Array data generation unit S ₆₄ Model array data storage unit S ₈₄ Model graphic element array data storage unit S ₈₇ Intermediate graphic element Extractor S ₈₈ Model graphic element storage

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location 9061-5L 460 B 9061-5L 460 F

Claims (7)

[Claims] 1. An input image is recorded and held, a local feature element in the image is extracted, placement information of the local feature element is generated, and the local feature of an object to be recognized. The combination arrangement information of elements is stored as storage information, and the arrangement information of the generated local feature element and the storage information are collated and determined, and the existence area in the image of the determined recognition information is determined. Object recognition method to extract. 2. An input image is recorded and held, local feature elements in the image are extracted, and area-based information such as color, local spatial frequency, and intensity of the area near the local feature element is extracted. Then, the arrangement information of the local feature element and the region-based information is generated, the combination arrangement information of the local feature element of the object to be recognized is stored as storage information, and the generated arrangement information and the storage are stored. An object recognition method that determines by checking information. 3. An image to be input is recorded and held, local feature elements in the image are extracted, and an object to be recognized,
The model feature element of the local feature element is stored as the first storage information, and the extracted local feature element and the first feature information are stored.
The intermediate graphic element of the local feature element is extracted from the stored information of the local feature element, the layout information of the intermediate graphic element is generated, and the combination layout information of the model graphic element of the object to be recognized is stored in the second storage. An object recognition method, which stores the information as information and collates the generated arrangement information of the intermediate graphic element with the second stored information for determination. 4. The intersection pattern of edge segments in a plurality of directions, all or a part of a curve having a constant curvature, and edge segments are extracted as the local feature elements. The described object recognition method. 5. The arrangement information of the local feature elements is
The object recognition method according to claim 1, wherein the local feature elements are represented as a two-dimensional array or a three-dimensional array of digitized elements to which numerical values discretized by a predetermined method are assigned. 6. The pattern of feature elements obtained by rearranging the combined arrangement information of the local feature elements on a grid space formed by the extracted local feature elements with a predetermined size and a predetermined shape unit. The object recognition method according to claim 1, wherein the object recognition method is represented. 7. A process of extracting the local feature element,
The object recognition method according to claim 1, wherein the object recognition method is performed for each of a plurality of scaling parameters having different sizes.