JP2011232845A - Feature point extracting device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect feature points stably even if a concealed region is present.SOLUTION: A feature point extracting device includes: an object detection unit 101 that detects an object from input images; a weight setting unit 102A that sets a weight for each feature point by using a likelihood A supplied from the object detection unit; a learning result storing unit 104 that includes a shape model showing a statistical positional relationship of the feature points and a profile showing a predetermined range of feature amount with respect to each of the positions of the feature points; a candidate point searching unit 103 that searches best candidate position of each feature point; and a shaping processing unit 105 that projects candidate points on the shape model by using a shape factor. The weight with respect to each feature point set by the weight setting unit 102A is used for weighting a difference between the candidate positions of the feature points and the positions of the feature points in the corresponding shape model to compute the shape factor.

Description

  The present invention relates to a technique applied to an apparatus, a method, a program, and the like for detecting feature points of an object from an image.

  When extracting facial feature points (eye, mouth, nose end points, etc.) from an image, generally face detection is first performed to limit the processing range. As a method for obtaining a point position for a face detected from an image (hereinafter referred to as “feature point extraction method”), there is a technique called ASM (Active Shape Model) (Non-patent Document 1). The procedure of the feature point extraction method will be described below.

FIG. 10 is a configuration diagram of a feature point extraction apparatus 1000 for implementing ASM.
The object detection unit 1001 detects an object (for example, a face) from the input image and calculates a face detection position.

  A shape model representing a statistical positional relationship between the feature point positions acquired for the learning face image and a profile representing a feature amount in a range defined in advance for each of the feature point positions are a learning result storage unit 1004. Retained. The shape model includes an average position vector indicating an average position of feature points and an eigenvector matrix indicating statistical variations of the feature point positions.

  The candidate point search unit 1003 uses the average position vector held in the learning result storage unit 1004 and the initial position calculated from the face detection position detected by the object detection unit 1001 as the base point and the profile of the candidate position of each feature point and the The best candidate position of each feature point is searched from the difference from the profile held in the learning storage unit.

  The shaping processing unit 1005 projects the candidate point onto the face shape model using the difference between the candidate point and the corresponding point in the face shape model. The candidate point search and shaping process are repeatedly executed until the difference between the projected candidate point position and the new candidate point position satisfies the constraint condition. The final candidate point position is determined as the feature point position.

S. Milborrow and F. Nicolls, “Locating Facial Features with an Extended Active Shape Model,” In Proceedings of ECCV, pp.504-513, 2008. Jie Chen et al., "Modification of the AdaBoost-based Detector for Partially Occluded Faces," In Proceedings of IEEE ICPR, pp.516-519, 2006. Kiyoto Ichikawa et al., "Component-based robust face detection using AdaBoost and decision tree," In Proceedings of IEEE FGR, pp.413-420, 2006.

  However, the conventional techniques have the following problems. In general, in the man-machine interface, etc., part of the face is hidden by other objects or enlargement due to the user's movement or surrounding environment, or part of the face is strongly illuminated by lighting fluctuations, etc. In some cases, the background is significantly different from the learning image.

  On the other hand, in the conventional technique, the candidate point search unit 1003 in FIG. 10 performs a candidate position search on the feature points of the entire face, and the acquired candidate positions of the feature points are within the stored profile and search range. It is only a position with a relatively high degree of coincidence. In other words, any point could be a candidate point if the coincidence was high by chance. In addition, the shaping processing unit 1005 directly uses the difference between the candidate point and the corresponding point in the face shape model, and projects the candidate point onto the face shape model. The deviation of the candidate points has caused a deviation in the shaping process of all candidate points.

  For example, as shown in FIG. 11, in the case of a complex background 1102 with edges and patterns such as walls and pillars near the face contour, or when there is a collar similar to the jaw contour, the candidate point search unit 1003 displays the face region. The position that deviates from the position becomes the candidate position, and the final feature point extraction result converges to the deviated position 1107. Further, when the face is partially concealed by a human hand or the like 1101, candidate point positions are acquired in a region that is not a face in the first place, so that candidate points that deviate greatly have a negative effect on feature point extraction in regions that are not partially concealed (1105).

  Therefore, the present invention aims to solve these problems and provide an apparatus, a program, and the like that can more accurately detect the position of a feature point from a face image according to the user's operation and the surrounding environment. To do.

  The feature point extraction apparatus of the present invention includes a learning result holding unit that holds a shape model representing a statistical positional relationship between feature points of an object, an input unit that inputs an image, and the feature points according to the input image. A weight setting unit for setting a weight on the input image, a candidate point search unit for searching for a candidate position for the feature point of the input image, and projecting the candidate position for the feature point to the shape model according to the weight for the feature point A shaping processing unit.

  With this configuration, it is possible to give locality to the shaping process, which is a global process of projecting the candidate position onto the shape model, by using the weight of the feature point set according to the input image. Therefore, feature point extraction that is resistant to local fluctuations can be performed.

  In addition, the feature point extraction device of the present invention includes an object detection unit that detects an area occupied by the object in the input image based on the object feature amount, and the weight setting unit is based on the object feature amount. A weight for the feature point is set, and the learning result holding unit further holds a learning profile representing a feature amount in a predefined range with respect to the feature point position, an average position of the feature point, and the candidate A point search unit searches for a candidate position of the feature point from a candidate profile that represents a feature amount in a predefined range from a search base point of the input image with respect to the feature point position and the learning profile, and the shaping processing unit includes: A shaping coefficient is obtained from the weight for the feature point, the candidate position, and the average position, and the candidate position of each feature point obtained by the candidate point search unit is converted into a shape model by using the shaping coefficient. Trying to characterized in that shadow.

  With this configuration, the weighting unit can set a weight as a candidate point search and shaping process for the local situation near each feature point in the input image by using the object feature amount. Feature point extraction that is not easily affected by concealment or the like can be performed.

  Furthermore, in the feature point extracting apparatus of the present invention, the weight setting unit sets weights based on likelihoods calculated by statistics of object feature amounts used in the object detection unit around each feature point position. To do.

With this configuration, feature points belonging to a region that seems to be an object are set with high weights, and regions that are not like objects such as partial concealment regions are set with low weights, thereby extracting feature points that focus on regions that are likely to be objects. Can be implemented.

  Furthermore, in the feature point extraction apparatus of the present invention, the object detection unit detects an object for each component of the object, and the weight setting unit belongs to a component that is not detected by the object detection unit. The weight for the feature point is set small, and the weight for the feature point belonging to the detected component is set large.

  With this configuration, it is possible to perform feature point extraction in accordance with changes in the location and size of the constituent elements of the object due to the user's movement and the camera zoom level.

  Further, in the feature point extraction apparatus of the present invention, the profile at the candidate position of the feature point and the profile held in the learning result storage unit are composed of a plurality of components calculated from images around the feature point position, The candidate point search unit searches for the best candidate position based on the degree of coincidence between the components of the two profiles, and the weight setting unit determines the background among the components of the profile held in the learning result storage unit. The weight for the corresponding part is set small.

  With this configuration, it is possible to prevent erroneous extraction of a background edge or the like near the contour of the object as a feature point.

  Furthermore, in the feature point extracting apparatus of the present invention, the weight setting unit sets a weight according to statistical variation of each feature point in the learning image.

  With this configuration, by setting a large weight for feature points having relatively large variations, it is possible to perform shaping processing so as to follow the variations.

  Furthermore, the feature point extraction apparatus according to the present invention determines the number of feature points to be processed by the candidate point search unit and the shaping processing unit according to the size of the target region occupied in the image detected by the target detection unit. Increase or decrease.

  With this configuration, when a constituent element of an object is detected, feature points can be extracted with high accuracy by arranging the feature points at a higher density than when the entire object is detected.

  Furthermore, in the feature point extraction apparatus of the present invention, the learning result storage unit holds a plurality of shape models for each combination of feature points to be used.

  With this configuration, since the shape model can be changed according to the size and part of the detected component of the target object, a more accurate shaping process can be performed.

  Furthermore, in the feature point extraction apparatus of the present invention, the candidate point search unit selectively performs a candidate point search according to the weight of each feature point position set by the weight setting unit.

  With this configuration, it is possible to reduce the amount of calculation required for the candidate point search.

  Furthermore, in the feature point extracting apparatus of the present invention, the weight setting unit sets a weight based on a partially concealed region or zoom state of the object set by the user of the feature point extracting apparatus.

  With this configuration, feature points can be extracted based on the weight set by the user.

  Furthermore, when the feature point extraction apparatus of the present invention repeatedly performs the processes in the candidate point search unit and the shaping processing unit, the weight setting unit updates the weight based on the previous shaping processing result.

With this configuration, the shaping process is performed while dynamically updating the weight according to the likelihood at each position according to the candidate point position that changes by repeating the processes in the candidate point search unit and the shaping processing unit. Therefore, more accurate feature point extraction can be performed.

  The present invention has an effect of more accurately detecting the position of a feature point from a face image according to the user's operation and the surrounding environment.

The block diagram of the feature point extraction apparatus in the 1st Embodiment of this invention The flowchart for operation | movement description of the feature point extraction apparatus in the 1st Embodiment of this invention The figure for demonstrating 102 A of weight setting parts in the 1st Embodiment of this invention The figure which shows the example of the feature point extracted from the face image The block diagram of the feature point extraction apparatus in the 2nd Embodiment of this invention The figure for demonstrating the weight setting part 102B in the 2nd Embodiment of this invention The figure for showing the condition assumed in order to demonstrate the feature point extraction apparatus in the 2nd Embodiment of this invention. Explanatory drawing which shows the 3rd Embodiment of this invention. Explanatory drawing which shows the effect of the 3rd implementation of this invention Block diagram of a feature point extraction device in the prior art The figure which shows the example of a subject in a prior art, and the effect of this invention

(Embodiment 1)
Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a feature point extracting apparatus 100 according to the present invention.

  In the first embodiment, an object detection unit 101 that detects an object from an input image, a weight setting unit 102A that sets the weight of each feature point using the likelihood A output from the object detection unit, A learning result storage unit 104 that holds a shape model that represents a statistical positional relationship between feature points, and a profile that represents a feature amount within a predefined range for each feature point position, and the best of each feature point A candidate point searching unit 103 for searching for a candidate position and a shaping processing unit 105 for projecting the candidate point onto the shape model.

The shape model includes an average position vector μ indicating an average position of feature points according to an expression shown in Equation 1, and an eigenvector indicating statistical variation of the feature point positions as shown in Equation 2. It consists of a matrix P. Here, N is the number of face images or feature point vectors sk, n is the number of feature points, j is a feature point number, m is the number of eigenvectors used for shaping processing, and pm is an eigenvector. As shown in Equation 1, the feature point vector sk is composed of x and y coordinates of each feature point.

  The operation of the feature point extraction apparatus configured as described above will be described below with reference to the flowchart shown in FIG.

  In the present embodiment, as shown in FIG. 3, it is assumed that the object is a face and is partially hidden by a shielding object.

  First, an image to be processed is read (step S10). And the target object detection part 101 detects the area | region of the target object in the same image (step S11). When the target is a face, as shown in Non-Patent Document 2, a plurality of facial feature quantities for identifying a face and a non-face belong to each partial area of the face, and the facial feature quantities calculated in advance in each partial area The likelihood Aj indicating the likelihood of a face in the partial area is calculated from the statistic, and a face or a non-face is identified in each partial area, so that a partially concealed face can also be detected. Specific examples of face feature values include Haar-like feature values and LBP feature values. Specifically, as shown in FIG. 3, the partial area 301 includes partial areas such as the left side, right side, left eye, and right eye of the face.

  In addition, the likelihood A calculated here has a part of the face in the region where the value is high, and the region where the value is low is likely to be a non-face (eg background) or a concealment object. Indicates.

  Note that the face detection method and feature amount shown in Non-Patent Document 2 are examples. The present invention is not limited to this method, and any method may be used as long as a face can be detected based on likelihood. Currently, many face detection methods using various types of face feature values learned by Adaboost have been proposed. However, any method uses face detection based on the likelihood of the face feature values. It can be used as an object detection unit of the present invention.

  Next, the weight setting unit 102A sets the weight wj for each feature point position using the likelihood Aj calculated by the object detection unit 101 (step S12). As shown in Non-Patent Document 2, the object detection unit 101 includes a plurality of facial feature quantities for identifying a face and a non-face for each partial region of the face. A likelihood Aj indicating the likelihood of a face in the partial area is calculated from the statistic of the face feature amount calculated in advance in each partial area, and a face or a non-face is identified in each partial area.

  As shown in FIG. 3, if a feature point 302 is associated with each partial region 301 in the object detection unit 101 corresponding to a portion indicated by each feature point, each feature point is determined from the likelihood Aj of each partial region. It is possible to set a weight vector w having a weight wj as an element. Here, Amin is the minimum value of the likelihood Aj of each feature point position, and Amax is the maximum value of the likelihood Aj of each feature point position.

As a result, as shown by 302 in FIG. 3, the likelihood in the region with partial concealment is low, and the likelihood is high in the region that is not. That is, the weight is set to be low in an area with partial concealment, and the weight is set to be high for an area that is not. Note that the weight wj for each feature point may be determined from statistical variation of each feature point. For example, the feature points near the eyebrows and the nose are relatively little changed compared to the feature points located near the eyes and mouth, that is, the variation is small. In such a case, the weight may be set high. As a result, feature points can be extracted according to the likelihood of each face region, so that feature point extraction can be performed preferentially in regions that are not partially concealed, and candidates based on partial concealment can be obtained. It is possible to suppress the influence of a large shift of the point position.

Next, the number of iterations i is set to 1, and a vector ss ⁽ⁱ⁾ = [x1,..., Xn, y1,... Yn] having the base point position (xj, yj) of the candidate point search at each feature point j as an element, Initialization is performed with elements of the average vector μ (step S13).

Next, as shown in Equation 4 for each feature point, the candidate point search unit 103 obtains the profile g acquired for each position (x, y) from the search base point (xj, yj) of the input image in the input image. Based on the dissimilarity f (x, y) between (x, y) and the profile g held in the learning result storage unit 104, the best candidate point position (xj ^′ , yj ^′ ) having the lowest dissimilarity ) (Step S14), and a candidate point position vector having the best candidate point position (xj ^′ , yj ^′ ) as an element is denoted by sc. Here, Σg is a covariance matrix of the profile g held by the learning result storage unit 104.

  The profile is composed of a plurality of components (for example, pixel values and differential images) calculated from images around the feature point position, and contributes to the portion corresponding to the background among the components of the profile held in the learning result storage unit. The rate may be set small. Thereby, it is possible to prevent a confusing edge or the like included in a complicated background in the real environment from being a candidate point. Note that as the feature amount used as the profile, luminance around the feature point, its differential value, SIFT feature amount, or the like is used.

  In addition, for feature point positions whose weights set by the weight setting unit 102A are equal to or less than a predefined threshold value or 0, it is not necessary to search for candidate point positions. Thereby, it is possible to reduce the processing amount of the candidate point position search.

  Next, the shaping processing unit 105 projects the candidate point position (x ′, y ′) onto the shape model (step S15). The shaping processing unit 105 calculates a shaping coefficient b for projecting the candidate point position vector sc onto the shape model using the equation shown in Equation 5. The shaping coefficient b is calculated by the product of a vector d obtained by multiplying each element of the difference vector between the candidate point position vector sc and the average vector μ by each element wj of the weight vector w and the eigenvector matrix P.

  Here, the weight determined based on the likelihood of face detection around the feature point is multiplied by the difference vector at each feature point position. For this reason, the difference with respect to the feature point position set to a value with low likelihood and low weight has a small influence on the shaping coefficient b calculation. As a result, feature points can be extracted with emphasis on feature points located in regions with high object-likeness, and the influence on shaping processing due to partial concealment can be suppressed.

  Next, in order to give a constraint condition to the shape model, it is determined whether or not the shaping coefficient b is smaller than a predefined threshold value bMAX (step S16). When the shaping coefficient b is equal to or greater than bMAX, b = bMAX is set, and variation in the shape model is suppressed, and the process proceeds to step S18 (step S17). If the constraint condition is satisfied, the process proceeds to step S18.

In step S18, after updating the number of repetitions i, the candidate point position vector sc is projected onto the shape model by the shaping coefficient b held within the number constraint condition as shown in Equation 6, and the candidate point A vector ss ⁽ⁱ⁾ that becomes a new starting position of the search is calculated.

The number of iterations i is below the maximum number of iterations iMAX or the difference between the candidate point position vector sc and the search base point position vector ss ^(i-1) || ss ^(i-1) −sc || It is determined whether or not it is smaller. If the condition is not satisfied, the process proceeds to step S14. If the condition is satisfied, the iterative process is terminated, and ss ⁽ⁱ⁾ is output as a feature point extraction result (step S19).

  As described above, steps S14 to S19 are repeatedly performed until the constraint condition in step S19 is satisfied. FIG. 4 shows an example of feature points extracted from the face image.

  Note that, since the base point of the candidate point search changes in each repetition of the candidate point search and the shaping process, the weight wj may be recalculated from the likelihood Aj in each partial region of the object detection unit 101. . In the process of feature point extraction, the accuracy of the feature point extraction position gradually increases each time the candidate point search and the shaping process are repeated, so that highly accurate weighting is possible and more accurate feature point extraction can be expected.

  Note that the learning result storage unit 104 may hold a plurality of shape models and select a shape model to be used as a starting point for the candidate point search according to the position and size of the face partial concealment region. This makes it possible to extract feature points with higher accuracy by using the shape model learned according to the face partial concealment region.

  According to the first embodiment, by using the weight for each feature point set by the weight setting unit 102A by the shaping processing unit 105 according to the above configuration, shaping according to the detection situation of the partially concealed face. Processing can be performed and detection can be performed more accurately according to the user's operation and the surrounding environment.

  In the first embodiment of the present invention, the operation is described using an example in which the object is a face. The object is a human or a human component including at least a limb, a torso, and a shoulder, Facial or facial components including at least eyes, nose, mouth, eyebrows, jaws and ears, animals or animal components including at least tail, eyes, nose, mouth and ears, car or bonnet, side mirror, There may be a vehicle in which a shape model can be set, such as a vehicle component including at least a license plate, or a built-in X-ray image.

(Embodiment 2)
FIG. 5 is a block diagram showing an example of the feature point extracting apparatus 200 according to the second embodiment to which the present invention is applied. The feature point extraction apparatus according to the second embodiment includes a weight setting unit 102B that receives the object part information B from the object detection unit 101 and sets a weight for each feature point. Other configurations are the same as those of the first embodiment (FIG. 1).

  The object detection unit 101 detects a face as in the first embodiment. Hereinafter, the operation of the weight setting unit 102B will be described.

In the present embodiment, a situation illustrated in FIG. 6 is assumed.
FIG. 6 shows the positions of the user 603 and the camera 604 and their temporal changes at 601 and 602, and images 611 and 612 that are input to the object detection unit at the respective times. As the user 603 approaches the camera 604, facial components occupy most of the input image. Note that the camera 604 may zoom in on the face of the user 603, resulting in the same situation as in FIG.

  As shown in Non-Patent Document 3, the object detection unit 101 performs object detection using a discriminator prepared for each face component such as eyes, mouth, and nose in addition to the face. In the situation of 602 in FIG. 6, the object detection unit 101 detects a region including the left eye and the nose as shown in FIG. 7. Non-Patent Document 3 discusses only an object detection method for each face component, and this cannot extract feature points.

  The weight setting unit 102B holds information about which part each feature point corresponds to, and based on the object part information B received from the object detection part 101, the feature belonging to the component not detected The weight for the point is set small, and the weight for the feature point belonging to the detected component is set large. In the example of FIG. 7, the weights of the feature points belonging to the left eye and the nose are high, and the weights of the feature points of other portions, for example, the portion belonging to the mouth are low.

  Note that the number of feature points processed by the candidate point search unit 103 and the shaping processing unit 105 may be increased or decreased according to the size of the area occupied by the facial component. Thus, even when only a partial region of the face can be used for feature point extraction as in 602 of FIG. 6, the accuracy of feature point extraction can be increased by increasing the number of feature points to be extracted. For example, when a part of the face is enlarged greatly, the distance on the image between the feature points used before enlargement becomes large, so the feature points are interpolated between the original feature points. By increasing the accuracy, the accuracy can be improved.

  Note that the learning result storage unit 104 may hold a plurality of shape models having different number of feature points or combinations. Thus, by using the shape model learned according to the number of feature points and the combination according to the size and position of the object detected by the object detection unit 101, the user can look into the camera or zoom in with the camera. It is possible to respond flexibly to changes and to extract feature points with higher accuracy. According to the second embodiment, with the above-described configuration, the weight for each feature point set by the weight setting unit 102B is utilized by the shaping processing unit 103, so that a part of the face is captured as the user looks into the camera. Even in such a state, by dynamically changing the weights for each feature point and performing the shaping process, it is possible to perform the feature point extraction process so as to follow the user's movement. It can be detected more accurately according to the surrounding environment.

(Embodiment 3)
8 and 9 show Embodiment 3 of the present invention. In the third embodiment, reference numeral 801 denotes a user, 802 denotes a camera, and 803 denotes a monitor. Feature points are extracted from an image photographed by the camera 802 by a feature point extraction device and displayed on the monitor 803.
In the third embodiment, a state is shown in which a user 801 takes a picture of his / her face and puts on makeup while displaying it on the monitor 803. At this time, if it is possible to extract feature points from the image of the eye that has been captured by the camera 802 and zoomed in, and display a guideline for makeup based on the feature points, FIG. As shown in FIG. 5, there is an excellent effect that the feature points 1002 of the eyes and eyebrows can be extracted and the guideline 1001 can be accurately drawn based on the extracted feature points.

  As described above, the feature point extraction apparatus and the feature point extraction method according to the present invention can be operated even when a face is partially concealed or even when a part of the face is captured as the user looks into the camera. In addition, by performing the shaping process by changing the weight for each feature point, it is possible to perform the feature point extraction process to follow the user's movement, which is taken by a surveillance camera, a still camera, etc. As a technique applied to AR, avatar and portrait generation for feature point extraction, moving object subject extraction, object authentication using the feature point, morphing, information presentation and makeup guideline display, etc. Useful.

DESCRIPTION OF SYMBOLS 100 Feature point extraction apparatus 101 Object detection part 102A Weight setting part 103 Candidate point search part 104 Learning result storage part 105 Shaping process part

Claims (12)

A learning result holding unit that holds a shape model representing a statistical positional relationship between feature points of an object;
An input unit for inputting an image;
A weight setting unit that sets weights on the feature points according to the input image;
A candidate point search unit for searching candidate positions for the feature points of the input image;
A feature point extraction apparatus comprising: a shaping processing unit that projects a candidate position for the feature point onto the shape model in accordance with a weight for the feature point. An object detection unit that detects an area occupied by the object in the input image based on the object feature amount;
The weight setting unit sets a weight for the feature point from the object feature amount,
The learning result holding unit further includes:
A learning profile that represents a feature amount within a predefined range with respect to the feature point position, and an average position of the feature point are retained,
The candidate point search unit searches for a candidate position of the feature point from a candidate profile that represents a feature amount in a predefined range from a search base point of the input image with respect to the feature point position and the learning profile,
The shaping processing unit obtains a shaping coefficient from a weight for the feature point, the candidate position, and the average position,
The feature point extraction apparatus according to claim 1, wherein the candidate position of each feature point obtained by the candidate point search unit is projected onto a shape model by using the shaping coefficient. The feature according to claim 1 or 2, wherein the weight setting unit sets a weight based on a likelihood calculated by statistics of an object feature amount used in the vicinity of each feature point position and the object detection unit. Point extraction device. The object detection unit detects an object for each component of the object,
The weight setting unit is configured to set a small weight for a feature point belonging to a component not detected by the object detection unit and a large weight for a feature point belonging to a detected component. The feature point extraction device described in 1. The profile at the feature point candidate position and the profile held in the learning result storage unit are composed of a plurality of components calculated from images around the feature point position.
The candidate point search unit searches for the best candidate position based on the degree of coincidence between the components of the two profiles, and the weight setting unit includes the profile components held in the learning result storage unit, The feature point extracting apparatus according to claim 1, 2, or 3, wherein a weight for a portion corresponding to a background is set to be small. 4. The feature point extraction device according to claim 1, wherein the weight setting unit sets a weight according to a statistical variation of each feature point in the learning image. The number of feature points to be processed by the candidate point search unit and the shaping processing unit is increased or decreased according to the size of the target region in the image detected by the target detection unit. The feature point extraction device described. 9. The feature point extraction device according to claim 4, wherein the learning result storage unit holds a plurality of shape models for each combination of feature points to be used. The candidate point search unit, according to the weight of each feature point position set by the weight setting unit,
The feature point extraction apparatus according to claim 1, 2, or 3, wherein the search for candidate points is selectively performed. 4. The feature point extraction apparatus according to claim 1, wherein the weight setting unit sets a weight based on a partially concealed region or zoom state of an object set by a user of the feature point extraction apparatus. When repeatedly performing the processing in the candidate point search unit and the shaping processing unit,
4. The feature point extraction apparatus according to claim 1, wherein the weight setting unit updates the weight based on a previous shaping process result. A learning result holding step for holding a shape model representing a statistical positional relationship between feature points of the object;
An input step for inputting an image;
A weight setting step for setting a weight to the feature point according to the input image;
A candidate point search step of searching for candidate positions for the feature points of the input image;
A feature point extraction method comprising: a shaping processing step of projecting candidate positions for the feature points onto the shape model in accordance with weights for the feature points.

Images