JP6124566B2 - Image recognition method and image recognition apparatus - Google Patents

Description

  The present invention relates to an image recognition method and an image recognition apparatus.

  Conventionally, an image recognition method is known in which an identifier is created from a registered image registered in advance, the attribute of a new image input using the created identifier is determined, and the image is recognized (for example, Patent Document 1). In Patent Document 1, a registered image (for example, an image of a workpiece) registered in advance is decomposed into partial images, and a luminance difference between any two points on the partial image is extracted as a feature amount to learn a registered image. Based on the matching between the learned registered image (partial image) and the partial image of the input new image (for example, the image of the workpiece), recognition of the new image (for example, recognition of the position and orientation of the workpiece) is performed. It is configured to be In addition, when using the brightness | luminance difference of two arbitrary points on a partial image as a feature-value, since this feature-value is not invariant with rotation of a partial image, an optical axis (a workpiece | work is image | photographed to the registration image registered previously. It is necessary to include an image obtained by rotating a work around an axis perpendicular to the lens.

JP2011-22991A

  However, in the image recognition method described in Patent Document 1, an image obtained by rotating the workpiece around the optical axis (an image of the workpiece rotated at a predetermined angle) is necessary to obtain the position and orientation of the workpiece. Therefore, the number of registered images registered in advance increases, and there is a problem that it takes time to recognize an input new image (matching between a registered image and a new image).

  The present invention has been made to solve the above-described problems, and one object of the present invention is to provide an image recognition method and an image recognition apparatus capable of suppressing time-consuming image recognition. Is to provide.

To achieve the above object, an image recognition method according to a first aspect uses a step of extracting a plurality of feature points from a learning image, and using a rotation-invariant feature amount for the extracted feature points. A step of calculating a feature amount, a step of creating a discriminator for determining the attribute of the feature point based on the calculated feature amount of the feature point of the learning image, and a plurality of feature points for the estimated image A step of extracting, and a step of recognizing the estimated image by aggregating the attributes of a plurality of feature points of the estimated image extracted using the classifier to determine the position of the estimation target .

  In the image recognition method according to the first aspect, as described above, a feature amount that is not rotation-invariant is obtained by including a step of calculating a feature amount using a rotation-invariant feature amount with respect to the extracted feature points. Unlike the case where the feature amount of the feature point is extracted by using, it is not necessary to learn a learning image rotated at every predetermined angle. That is, it is possible to create a discriminator from fewer learning images than when creating a discriminator for determining the attribute of a feature point based on a feature amount that is not rotation invariant. When determining the attribute of the feature point of the estimated image, the number of feature points of the learning image matched with the feature point of the estimated image can be reduced, and the amount of calculation can be reduced accordingly. As a result, it can be suppressed that it takes time to recognize the image.

An image recognition apparatus according to a second aspect is configured to extract a feature amount using a first feature point extraction unit that extracts a plurality of feature points from a learning image, and a rotation-invariant feature amount with respect to the extracted feature points. A feature amount calculating means for calculating, a discriminator creating means for creating a discriminator for determining the attribute of the feature point based on the calculated feature amount of the feature point of the learning image, and a plurality of the estimated images Second feature point extracting means for extracting feature points, and recognition means for recognizing the estimated image by determining the position of the estimation target by aggregating the attributes of the plurality of feature points of the estimated image extracted using the classifier With.

  In the image recognition apparatus according to the second aspect, as described above, the feature amount calculating means for calculating the feature amount using the rotation-invariant feature amount is provided for the extracted feature point, so that the rotation is not invariant. Unlike the case where the feature amount of the feature point is extracted using the feature amount, it is not necessary to learn the learning image rotated at every predetermined angle. That is, it is possible to create a discriminator from fewer learning images than when creating a discriminator for determining the attribute of a feature point based on a feature amount that is not rotation invariant. When determining the attribute of the feature point of the estimated image, the number of feature points of the learning image matched with the feature point of the estimated image can be reduced, and the amount of calculation can be reduced accordingly. As a result, it is possible to provide an image recognition apparatus capable of suppressing the time taken for image recognition.

  By comprising as mentioned above, it can suppress that recognition of an image takes time.

1 is an overall view of a robot system according to an embodiment of the present invention. 1 is a block diagram of a robot system according to an embodiment of the present invention. It is a flowchart at the time of learning of the image recognition method by one Embodiment of this invention. It is a conceptual diagram of the classification tree of the image recognition method by one Embodiment of this invention. It is a flowchart at the time of the estimation of the image recognition method by one Embodiment of this invention. It is a conceptual diagram at the time of the estimation of the image recognition method by one Embodiment of this invention. It is a perspective view of the workpiece | work in the experiment 1 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the learning image in the experiment 1 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the presumed scene (work piled up) in the experiment 1 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the voting surface (voting result) at the time of the estimation in the experiment 1 performed using the image recognition method by one Embodiment of this invention. It is a perspective view of the workpiece | work in the experiment 2 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the learning image (surface of a workpiece | work) in the experiment 2 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the learning image (the back surface of a workpiece | work) in the experiment 2 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the presumed scene (work piled up) in the experiment 2 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the voting surface (voting result) at the time of estimation of the surface of the workpiece | work in the experiment 2 performed using the image recognition method by one Embodiment of this invention. It is a figure which shows the voting surface (voting result) at the time of the estimation of the back surface of the workpiece | work in the experiment 2 performed using the image recognition method by one Embodiment of this invention.

  Hereinafter, the present embodiment will be described with reference to the drawings.

  First, the configuration of the robot system 100 according to the present embodiment will be described with reference to FIGS. 1 and 2.

  As shown in FIGS. 1 and 2, the robot system 100 is provided with a robot 1, a robot controller 2, and a sensor unit (image sensor unit) 3. The sensor unit 3 is an example of the “image recognition device” in the present invention.

  As shown in FIG. 1, the robot 1 includes a base 11, a robot arm 12 attached to the base 11, and an end effector 13 attached to the tip of the robot arm 12. The robot arm 12 is configured with six degrees of freedom. The robot arm 12 has a plurality of arm structures, and an arm structure 12 a is rotatably connected to the base 11 around a rotation axis A 1 perpendicular to the installation surface of the robot 1. The arm structure 12b is connected to the arm structure 12a so as to be rotatable around a rotation axis A2 perpendicular to the rotation axis A1. The arm structure 12c is connected to the arm structure 12b so as to be rotatable around a rotation axis A3 parallel to the rotation axis A2. The arm structure 12d is connected to the arm structure 12c so as to be rotatable around a rotation axis A4 perpendicular to the rotation axis A3. The arm structure 12e is connected to the arm structure 12d so as to be rotatable around a rotation axis A5 perpendicular to the rotation axis A4. The arm structure 12f is connected to the arm structure 12e so as to be rotatable around a rotation axis A6 perpendicular to the rotation axis A5. Here, “parallel” and “vertical” have a broad meaning including not only “parallel” and “vertical” in a strict sense but also those slightly deviated from “parallel” and “vertical”. Each rotary shaft A1 to A6 is provided with a servo motor (joint), and each servo motor has an encoder for detecting the respective rotational position. Each servo motor is connected to the robot controller 2 and is configured such that each servo motor operates based on a command from the robot controller 2.

  As shown in FIG. 2, the sensor unit 3 is provided with a camera 31 that captures a two-dimensional image and a laser scanner 32. A sensor controller 33 including an image processing unit 34 and a memory 35 is provided inside the sensor unit 3. In addition, the sensor unit 3 irradiates laser light onto the workpiece 200 (estimation target, see FIG. 1) stacked from the laser scanner 32 and shoots the light reflected from the workpiece 200 by the camera 31. The three-dimensional shape of the workpiece 200 is measured. Further, the sensor unit 3 can measure the three-dimensional shape of the workpiece 200 (estimation of the distance to the workpiece 200 and the detailed position and orientation of the workpiece 200). Before estimating the posture, the image processing unit 34 estimates the approximate position of the workpiece 200 (and the workpiece 201 described later, see FIG. 11) based on the two-dimensional image captured by the camera 31. It is configured as follows. The image processing unit 34 is an example of “first feature point extracting unit”, “feature amount calculating unit”, “classifier creating unit”, “second feature point extracting unit”, and “recognizing unit”. The workpieces 200 and 201 are examples of “estimation targets”. Hereinafter, an image recognition method (a method for estimating the approximate position of the workpieces 200 and 201) according to the present embodiment will be described.

(Rotation invariant feature based on DoG)
First, the rotation invariant feature amount used in the image recognition method will be described. In this embodiment, DoG (Difference-of-Gaussian) is used as a rotation-invariant feature amount. Hereinafter, DoG (DoG value) will be described. First, by convolving a Gaussian function G (u, v, σ) shown in the following equation (1) with respect to a learning image I (x, y) at an arbitrary feature point (x, y), A smoothed image L (x, y, σ) shown in Expression (2) is generated.

Next, a difference image D ^{(i, j)} (x, y) of two smoothed images obtained by the two smoothing parameters σ _i and σ _j is generated by the following equation (3).

Then, according to the above equation (3), σ _i , σ _j ∈ [σ ₁ , σ ₂ ,. . . , Σ _m ] in the range ， σ _i ∀σ _j (σ _i <σ _j ), D ^{(i, j)} (x, y) is obtained and the feature vector V shown in the following equation (4) is obtained. The element is (x, y).

The feature vector V shown in the above equation (4) has DoG (DoG value) as an element. In the present embodiment, DoG is the difference between the sum of the luminance values of two concentric regions with different ranges in the learning image and the estimated scene (estimated image) (the two values obtained from σ _i and σ _j above). Since the difference is a smoothed image, it is a rotation-invariant feature quantity. The feature vector V (DoG) is a feature amount similar to a wavelet feature amount (frequency analysis) using a mother wavelet function (finite long waveform). While the wavelet feature quantity has resolution and orientation as elements, DoG does not have an orientation element. However, the feature vector V (DoG) has many variations (D for various σ) with respect to resolution (size). For this reason, it is possible to select a feature quantity of an appropriate size for the estimation target by using the following technique for reducing the number of dimensions of the feature vector V.

(Reducing the number of dimensions of feature vector V)
Since the feature vector V shown in the above equation (4) has many DoGs (dimensions), the separation capability of the feature vector V is improved. However, it may take a lot of time to generate DoG at the time of estimation (recognition) of the estimation target, or there may be wasted (similar features) DoG. Therefore, in the present embodiment, a plurality of vectors (f to be described later) having DoG (DoG value) as elements are generated, and DoG values having low correlation with each other are selected based on the Hamming distance between the plurality of vectors ( And the classifier is created based on the selected DoG value. Hereinafter, a method for reducing the number of dimensions of the feature vector V will be described in detail.

First, in obtaining the feature vector V, it is necessary to compare the classification performance at all feature points (n points on the learning image) in order to select the optimum element D. For some σ _i , σ _j , (x, y) ε [(x ₁ , y ₁ ), (x ₂ , y ₂ ),. . . , (X _n , y _n )] with respect to ∀ (x, y), a vector f ^{(i, j)} whose elements are D ^{(i, j)} (x, y) is expressed by the following equation (5). Newly defined.

In the above equation (5), d ^{(i, j)} (x, y) is a binary value of D ^{(i, j)} (x, y) according to the above equation (6). . In the above equation (6), D _med ^{(i, j)} is a median value obtained by the above equation (7). Note that a plurality of f ^{(i, j)} are generated. Here, D which is an element of f ^{(i, j)} is a real value. Therefore, the median value of D, which is an element of f ^{(i, j)} , is set as a threshold value, and each element D is binarized to “0” or “1” depending on whether each element D is larger than this threshold value. . As a result, “0” and “1” are equally present in the bits in f ^{(i, j)} , and this binarized element classifies n feature points into two. Therefore, it becomes appropriate information.

Next, an algorithm for determining F, which is an optimal set of f ^{(i, j)} , will be described. First, the first element f ₁ (t = 1) of the set F is randomly selected from all f ^{(i, j)} . Thereafter, while t satisfies 2 ≦ t ≦ T _max , the following processing is sequentially performed. Specifically, when selecting the t-th f _t , H ^{(i, j)} shown in the following equation (8 ⁾ is calculated for all f ^{(i, j)} not included in the set F. Is done.

Here, the function ω _H (f ^{(i, j)} , f ^{(k, l)} ) in the above equation (8) is expressed by the following equation (9).

Here, d _H (f ^{(i, j)} , f ^{(k, l)} ) represents a Hamming distance between f ^{(i, j)} and f ^{(k, l)} . The Hamming distance refers to a different element between the element (“0” or “1”) of f ^{(i, j) and} the element (“0” or “1”) of f ^{(k, l)} . Means number. Then, among all the ^{H (i, j),} corresponding to the minimum value at which was ^{H (i, j)} is calculated source ^{f (i, j) (the} minimum value and which was ^{H (i, j)} F ^{(i, j)} ) to be added as an element of the set F.

The above algorithm sequentially adds f having an element D having the lowest correlation with the group of selected elements D among f having an unselected element D to the set F until t reaches T _max. It is a technique to go. Since each selected f is composed of Ds having different information, the selected f (D) is considered to be a lean feature amount (features having similar features are reduced). It is done.

(Approximate position estimation of estimation target using ensemble classification tree)
Next, estimation of the approximate position of the estimation target (recognition target) using the ensemble classification tree will be described with reference to FIGS.

(During learning)
First, the learning time will be described. In the present embodiment, as shown in step S1 of FIG. 3, feature points are extracted at random from the learning image. In step S2, a feature amount (feature vector V) using DoG (DoG value), which is a rotation-invariant feature amount, is calculated using the above equation (4) at each feature point. Next, in step S3, the number of dimensions of the feature vector V is reduced using the above equations (5) to (9). In step S4, a classification tree is created using the generated feature quantity (feature vector V, selected DoG) as a classification criterion. The procedure for creating a classification tree is described below.

As shown in FIG. 4, first, binarization is performed for all feature points stored in a node (nodes of a binary tree) with a median value for each element (element of feature vector V) as a threshold value. Is called. That is, for each element (v (DoG), see Expression (10)) of the feature vector V (see Expression (10) below), each element is “0” or “ 1 ”is binarized (feature vector V _bin , see formula (11) below).

Next, an arbitrary distance d is generated, and a Hamming distance d _H (V _bin (x _i , x _i ,) between the arbitrary distance d and the binarized feature vector V _bin (see the above equation (11)). y _i ), d) are calculated. Then, the classification of the feature vector V _{bin is performed} by comparing the calculated Hamming distance d _H (V _bin (x _i , y _i ), d) with a threshold value d _{th in} which the number of elements of the child nodes can be equally classified. Done. By performing the above processing recursively, a classification tree is created. In addition, in this embodiment, a plurality of independent classification trees are created by extracting feature points (step S1) every time a classification tree is created. These multiple classification trees are called ensemble classification trees. And the position of the estimation object mentioned later is estimated by using the created ensemble classification tree. The ensemble classification tree is an example of the “discriminator” of the present invention.

(At the time of estimation)
In this embodiment, as shown in FIG. 5, in step S11, the entire scene is extracted in the estimated scene (estimated image), and feature points are extracted (see FIG. 6). That is, for example, a raster scan is performed in the estimated scene, and a feature vector V (see the above equation (4)) is calculated at each scanned point. Next, in step S12, by using the ensemble classification tree created at the time of learning, in step S11, the feature point of the learning image similar to the feature amount of the extracted feature point of the estimated scene is obtained as the corresponding point. (The attribute of the feature point extracted in step S11 is determined, see FIG. 6).

  Next, in step 13, in this embodiment, as shown in FIG. 6, voting is performed at a position on the voting plane ((x, y) plane) corresponding to the corresponding point (attribute). Note that DoG, which is a feature quantity, is a rotation-invariant feature quantity, and therefore the direction of the estimation target in the estimation scene cannot be uniquely determined. For this reason, voting is performed circularly on the voting surface. As a result, it is possible to determine (estimate) that an estimation target (recognition target) exists at a place where votes are gathered on the voting surface (a place where many votes have been placed).

  Next, with reference to FIG. 7 to FIG. 16, the workpieces 200 and 201 are compared with the workpieces 200 and 201 in a state of being stacked in order to confirm the effectiveness of the image recognition method according to the present embodiment. An experiment for estimating the approximate center position will be described.

(Experiment 1)
In Experiment 1, as shown in FIG. 7, an experiment for estimating the approximate center position of the workpiece 200 was performed on a flat workpiece 200 having three holes 200a. Below, the conditions at the time of learning are demonstrated. This condition is the same in Experiment 2 described below.

  At the time of learning, an image of the workpiece 200 in a virtual environment created based on three-dimensional CAD data was used as a learning image. As shown in FIG. 8, in this embodiment, the learning image is an image in which one work 200 is placed on a plane. The learning image (and the estimated scene, see FIG. 9) is a two-dimensional image. The number of feature points on the learning image used for creating one classification tree is set to 300. The number of classification trees was 16. Further, at the time of estimation, feature points were extracted (feature amounts were calculated) for every 4 pixels without performing a full search for each pixel for an estimated scene of 256 × 256 pixels. That is, in the present embodiment (Experiments 1 and 2), the feature point of the estimated scene is a local image of the estimated scene.

  FIG. 9 is an image of the stacked workpieces 200 used at the time of estimation. Note that numbers 1 to 5 in FIG. 9 indicate places estimated as the center position of the workpiece 200 based on the voting results shown in FIG. FIG. 10 shows the result of voting on the voting plane based on the feature point attributes determined by the ensemble classification tree. Specifically, based on the attribute of the local image of the estimated scene determined by the ensemble classification tree, the result of voting a position where the center position of the workpiece 200 is thought to exist in a circular shape on the voting plane (for each local image The results of voting based on the attributes of the are indicated by contour lines. The numbers in FIG. 10 indicate the heights of the contour lines. Further, the position corresponding to the maximum value of the voting surface was estimated as the center position of the workpiece 200. In FIG. 9, the positions corresponding to the maximum values of the voting plane are ranked in order of the maximal values of the voting plane (voting rank order), and the top first to fifth (numbers 1 to 5). ).

  As shown in FIG. 9, it was confirmed that the higher results of the voting rank (numbers 1 to 4) roughly estimated the actual center position of the workpiece 200 in general. That is, it was confirmed that the image recognition method of this embodiment has high accuracy. Since the workpiece 200 has many flat surfaces, the possibility of the posture of the workpiece 200 is limited even when the workpiece 200 is piled up (the variation in posture is relatively small), and includes three holes 200a. It is thought that it was possible to estimate with high accuracy. On the other hand, when the workpiece 200 has a posture inclined with respect to the stacked surface, such a posture has not been learned, and thus the estimated center position may deviate from the actual center position. It was confirmed that there was.

(Experiment 2)
In Experiment 2, as shown in FIG. 11, an experiment was performed to estimate the approximate center position of the work 201 with respect to a flat work 201 having six holes 201a. As shown in FIGS. 12 and 13, the workpiece 201 has a different shape on the front surface and the back surface. Specifically, the surface of the work 201 has a periodic uneven shape, while the back surface has a flat surface.

  In Experiment 2, as shown in FIGS. 12 and 13, the image of the front surface and the image of the back surface of the work 201 placed on a plane were used as learning images.

  FIG. 14 is an image of the stacked workpieces 201 used at the time of estimation. In addition, the numbers 1-3 in FIG. 14 have shown the place estimated as the center position of the workpiece | work 201 based on the voting result shown in FIG.15 and FIG.16. As shown in FIGS. 15 and 16, voting surfaces were prepared for the front surface and the back surface of the workpiece 201, respectively. An ensemble classification tree was created based on the learning image on the surface of the work 201 (see FIG. 12). In addition, based on the attribute of the local image of the estimated scene determined by the created ensemble classification tree, the positions where the center position of the surface of the work 201 is considered to exist are voted in a circle on the voting surface (see FIG. 15). It was done. Similarly, an ensemble classification tree is created based on the learning image on the back surface of the work 201 (see FIG. 13). Further, based on the attribute of the local image of the estimated scene determined by the created ensemble classification tree, the position where the center position of the back surface of the work 201 is considered to be a circle on the voting surface (see FIG. 16). It was done. In FIG. 14, for each of the front surface and the back surface of the work 201, the positions corresponding to the maximum values of the voting surface are ranked in the order of the maximal values of the voting surface (voting rank order). The first to third positions (numbers 1 to 3) are described.

  As shown in FIG. 14, it was confirmed that the actual center position of the workpiece 201 was roughly estimated accurately. That is, it was confirmed that the image recognition method of the present embodiment has high accuracy even for the workpiece 201 having different front and back shapes. Since the workpiece 201 has many flat surfaces like the workpiece 200, the possibility of the posture of the workpiece 201 is limited even in a stacked state, and the six holes 201a and periodic irregularities are formed. It is thought that it was possible to estimate with high accuracy because it has a unique feature of including

  In the present embodiment, as described above, the feature point is calculated using the feature amount that is not rotation-invariant by calculating the feature amount using DoG that is the rotation-invariant feature amount with respect to the extracted feature point. Unlike the case where the amount is extracted, there is no need to learn a learning image rotated at every predetermined angle. In other words, an ensemble classification tree can be created from fewer learning images compared to creating an ensemble classification tree for determining feature point attributes based on feature quantities that are not rotation-invariant. When the feature point attribute of the estimated scene is determined using, the number of feature points of the learning image matched with the feature point of the estimated image can be reduced, and the amount of calculation can be reduced accordingly. . As a result, it can be suppressed that it takes time to recognize (estimate) an image (estimation target).

  Further, in the present embodiment, as described above, a plurality of smoothed images are generated by convolving a Gaussian function with respect to the feature points of the learning image as the rotation-invariant feature quantity, and the generated plurality of smoothing features. A DoG value that is a difference between two smoothed images of the digitized image is used. Thereby, it is possible to easily calculate a rotation-invariant feature amount with respect to the feature point extracted from the learning image.

  In the present embodiment, as described above, the DoG value is calculated as a difference between the sum values of the luminance values of the two concentric regions having different ranges at the feature points. Thereby, since the luminance values of the two concentric regions are rotation invariant values, it is possible to calculate a rotation invariant feature quantity with respect to the feature points extracted from the learning image.

  In the present embodiment, as described above, a DoG value having a low correlation among a plurality of DoG values is selected, and a discriminator is created based on the selected DoG value. Thereby, unlike the case where the classifier is created using all the DoG values, the amount of calculation when determining the attribute of the feature point of the estimated scene using the classifier can be further reduced. As a result, it is possible to further suppress the time taken for image recognition.

  In the present embodiment, as described above, a plurality of vectors f having DoG values as elements are generated, and based on the Hamming distances between the plurality of vectors f, DoG values having low correlation with each other are selected. An ensemble classification tree is created based on the selected DoG value. As a result, DoG values having a low correlation with each other are selected, so that the number of dimensions of the feature vector V can be effectively reduced.

  In the present embodiment, as described above, an ensemble classification tree having a plurality of classification trees for determining the feature point attributes from the feature quantities of the feature points of the learning image calculated using the rotation-invariant feature quantities. Create As a result, even when the discrimination performance (accuracy) of one classification tree is relatively low, the ensemble classification tree having a plurality of classification trees can enhance the attribute judgment performance of feature points.

  Further, in the present embodiment, as described above, the position of the estimation target is estimated by voting in a circle on the voting surface based on the attribute of the feature point determined by the ensemble classification tree. As a result, even when the feature amount of the feature point is calculated using DoG, which is a rotation-invariant feature amount, the voting is performed circularly on the voting surface, and the estimation target exists where the voting is gathered on the voting surface By determining that it is being performed, it is possible to easily estimate the estimation target (the center position of the workpieces 200 and 201).

  Further, in the present embodiment, as described above, the estimation scene is an image of a plurality of works 200 and 201 stacked in bulk, and the voting plane is based on the feature point attributes determined using the ensemble classification tree. The center positions of the plurality of workpieces 200 and 201 stacked in bulk are estimated by voting in a circle. Thereby, the center position of the several workpiece | work 200 and 201 piled up can be rapidly estimated with the ensemble classification | category tree produced based on the rotation invariable feature-value (DoG).

  In the present embodiment, as described above, the learning image is configured from the image of one workpiece 200. This reduces the time required for estimation (recognition) of the estimation target, unlike when preparing multiple learning images of a work rotated at a predetermined angle and creating an ensemble classification tree from multiple learning images. can do.

  In the present embodiment, as described above, the learning image and the estimated scene are configured from two-dimensional images. Thereby, unlike the case where the learning image and the estimation scene are configured from three-dimensional images, it is possible to quickly create an ensemble classification tree and estimate an estimation target.

  In the present embodiment, as described above, the feature points of the estimated scene are configured from the local images of the estimated scene. Thereby, unlike the case where the feature amount of the feature point is calculated at all points (pixels) of the estimation scene, the estimation target can be quickly recognized.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is shown not by the above description of the embodiments but by the scope of claims for patent, and further includes all modifications within the meaning and scope equivalent to the scope of claims for patent.

  For example, in the above-described embodiment, an example in which DoG (DoG value) is used as a rotation-invariant feature amount has been described. However, a feature amount other than DoG may be used as a rotation-invariant feature amount. For example, the distance from the center of the estimation target to the edge (contour) of the estimation target is calculated at a predetermined angular interval around the optical axis, and the feature amount obtained by frequency analysis of the obtained result is a rotation invariant. It may be used as a feature amount.

  In the above-described embodiment, an example in which the dimension of the feature vector V is reduced by selecting DoGs (DoG values) having a low correlation with each other is shown. However, the dimension of the feature vector V may not be reduced.

  Moreover, in the said embodiment, although the example which made the difference of the luminance value of each two concentric area | regions where the range in a feature point differs in the DoG value was shown, the difference of the average value of a luminance value is shown, for example It may be a DoG value.

  In the above embodiment, an example using an ensemble classification tree as a classifier has been shown. However, a classifier other than the ensemble classification tree (for example, one classification tree or a support vector machine (SVM)) may be used. .

  Moreover, in the said embodiment, although the example which estimates the center position of a workpiece | work from the image of the several workpiece | work piled up separately was shown, the estimation object (person in a photograph, It is also possible to estimate a predetermined building in an aerial photograph.

  Moreover, although the example which uses the image of one workpiece | work (image of the surface of one workpiece | work and a back surface) as a learning image was shown in the said embodiment, for example, it is mounted on the plane. In addition to the image of the workpiece, an image of the workpiece inclined with respect to the placement surface or an image of the side surface of the workpiece may be used as a learning image. This makes it possible to accurately estimate a plurality of workpieces stacked in various postures.

  In the above embodiment, an example is shown in which the approximate position of the workpiece is estimated by the image processing unit of the sensor unit. However, a part other than the image processing unit of the sensor unit (for example, a robot controller or a separate unit) The approximate position of the workpiece may be estimated by a personal computer (PC).

  Moreover, although the example which uses the above-mentioned image recognition method for a robot system was shown in the said embodiment, you may use the above-mentioned image recognition method for systems other than a robot system.

3 Sensor unit (image recognition device)
34 Image processing unit (first feature point extraction means, feature amount calculation means, classifier creation means, second feature point extraction means, recognition means)
200, 201 Workpiece (estimation target)

Claims (12)

Extracting a plurality of feature points from the learning image;
Calculating a feature quantity using a rotation-invariant feature quantity for the extracted feature points;
Creating a discriminator for determining an attribute of the feature point based on the calculated feature amount of the feature point of the learning image;
Extracting a plurality of feature points from the estimated image;
A step of recognizing the estimated image by aggregating attributes of a plurality of feature points of the extracted estimated image using the classifier to determine a position of an estimation target .   The rotation-invariant feature amount generates a plurality of smoothed images by convolving a Gaussian function with respect to a feature point of the learning image, and smoothes two of the generated smoothed images. The image recognition method according to claim 1, comprising a DoG (Difference-of-Gaussian) value that is a difference between images.   The image recognition method according to claim 2, wherein the DoG value is a difference between the sum of luminance values of two concentric regions having different ranges in the feature point. The DoG value includes a plurality of DoG values;
The step of creating a discriminator for determining the attribute of the feature point selects a DoG value having a low correlation among the plurality of DoG values, and selects the discriminator based on the selected DoG value. The image recognition method according to claim 2, further comprising a creating step.   The step of creating the discriminator based on the selected DoG value generates a plurality of vectors having the DoG value as elements, and the DoG values having low correlation with each other based on the Hamming distance between the plurality of vectors. The image recognition method according to claim 4, further comprising: creating a classifier based on the selected DoG value.   The step of creating a discriminator for determining the attribute of the feature point is for determining the attribute of the feature point from the feature amount of the feature point of the learning image calculated using the rotation-invariant feature amount. The image recognition method of any one of Claims 1-5 including the step which creates the ensemble classification | category tree which has two or more classification trees.   The step of determining the attribute of the feature point of the extracted learning image using the classifier and recognizing the estimated image is performed on the voting surface based on the attribute of the feature point determined by the classifier. The image recognition method according to claim 1, comprising a step of estimating a position of an estimation target by voting in a circle. The estimated image includes images of a plurality of workpieces stacked in bulk,
The step of estimating the position of the estimated image is performed by voting in a circle on a voting surface based on the attribute of the feature point determined using the discriminator. The image recognition method according to claim 7, comprising a step of estimating a position.   The image recognition method according to claim 8, wherein the learning image is composed of one image of the workpiece.   The image recognition method according to claim 1, wherein the learning image and the estimated image are two-dimensional images.   The image recognition method according to claim 1, wherein the feature point of the estimated image is a local image of the estimated image. First feature point extracting means for extracting a plurality of feature points from the learning image;
A feature quantity calculating means for calculating a feature quantity using a rotation-invariant feature quantity for the extracted feature points;
A discriminator creating means for creating a discriminator for determining an attribute of the feature point based on the calculated feature amount of the feature point of the learning image;
Second feature point extracting means for extracting a plurality of feature points from the estimated image;
An image recognition apparatus comprising: a recognizing unit that aggregates attributes of a plurality of feature points of the extracted estimated image using the classifier to determine a position of an estimation target and recognizes the estimated image.

Images