JP2011002882A - Imaging apparatus, image processing program, and imaging method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an imaging apparatus which quickly and correctly detects an object.SOLUTION: The imaging apparatus obtains an image, obtains a low resolution image based on the image, and detects an object candidate to the low resolution image. When the low resolution image is the object candidate, the imaging apparatus detects the object to a high resolution image of high resolution rather than the low resolution image.

Description

  The present invention relates to an imaging apparatus, an image processing program, and an imaging method.

  Patent Document 1 discloses a method of preparing low-resolution and high-resolution master image patterns and detecting an inspection object using these master image patterns. In Patent Document 1, first, a low-resolution master image pattern is applied to an entire low-resolution image to narrow the area. After narrowing down the area, a high-resolution whole image is acquired, and a high-resolution master image pattern is applied to the high-resolution whole image to detect the inspection object.

  Patent Document 2 discloses a technique for acquiring a low-resolution whole image to detect a face area, detecting a face area, acquiring a high-resolution image for the face area, and identifying a subject individually. Has been.

  Patent Document 3 discloses a technique for performing character recognition using higher-order local autocorrelation.

  In Non-Patent Document 1, a weighted sum of a plurality of weak classifiers based on rectangular information is integrated by AdaBoost to create a strong classifier, and a strong classifier is cascaded to detect a face as a target of interest in an image. Techniques to do this are disclosed.

  Non-Patent Document 2 discloses a technique for performing gesture recognition by applying AdaBoost.

JP 05-288520 A JP 2007-334802 A Japanese Examined Patent Publication No. 58-047064 P. Viola and M. Jones. "Rapid Object Detection Using a Boosted Cascade of Simple Features," in Proc. Of CVPR, vol.1, ppp.511-518, December, 2001 Takeshi Mita, Toshimitsu Kaneko, Bjo¨ rn Stenger, and Osamu Hori, "Discriminative Feature Co-Occurrence Selection for Object Detection" in JOURNAL OF LATEX CLASS FILES, VOL.1, NO.8, AUGUST 2002

  However, in the invention of the above-mentioned Patent Document 1, it is necessary to acquire a low-resolution whole image and a high-resolution whole image. For example, even when there is no inspection object, the low-resolution whole image and the high-resolution whole image These two images are acquired, and even when there is no inspection object, detection processing of the detection object is performed using these two acquired images. Therefore, there is a problem that the detection time of the inspection object becomes long. Further, the invention of Patent Document 2 has a problem that personal identification cannot be performed unless a face area can be detected from the entire low-resolution image. The invention of Patent Document 3 has a problem that characters cannot be recognized when the image is a low-resolution image. The techniques of Non-Patent Literature 1 and Non-Patent Literature 2 have a problem that the subject cannot be recognized when the subject is small.

  The present invention has been invented to solve such problems, and aims to quickly and accurately detect a subject in an image.

  An imaging apparatus according to an aspect of the present invention is an imaging apparatus that images a subject, and an image acquisition unit that acquires an image, and a low-resolution image acquisition that acquires a low-resolution image based on the image acquired by the image acquisition unit A first detection unit that detects a subject candidate for the low resolution image, a high resolution image acquisition unit that acquires a high resolution image that is higher in resolution than the low resolution image from the image acquired by the image acquisition unit, A second detection unit configured to detect a subject from the high-resolution image when a subject candidate is detected from the low-resolution image by the first detection unit.

  An image processing program according to another aspect of the present invention is an image processing program for processing a captured image by a computer, the image acquisition procedure for acquiring an image in the computer, and the image acquired by the image acquisition procedure. A low-resolution image acquisition procedure for acquiring a low-resolution image based on the first detection procedure for detecting a subject candidate for the low-resolution image, and a subject candidate for the low-resolution image detected by the first detection procedure And a second detection procedure for detecting a subject with respect to a high-resolution image acquired from the image acquired by the image acquisition procedure and having a higher resolution than the low-resolution image.

  An imaging method according to still another aspect of the present invention is an imaging method for imaging a subject, acquires an image, acquires a low-resolution image based on the acquired image, and selects a subject candidate for the low-resolution image. When a subject candidate is detected for a low-resolution image, the subject is detected for a high-resolution image that is acquired from the acquired image and has a higher resolution than the low-resolution image.

  According to these aspects, it is determined whether or not a subject candidate exists in the image using a low-resolution image having a relatively high processing speed. Only when it is determined that there is a subject candidate in the image, the subject is detected using the high-resolution image. Therefore, for example, when there is no subject in the image, the subject detection process can be quickly ended. Even when the subject is small, the subject candidate can be detected by using the low-resolution image. When a subject candidate is detected, subject detection is performed using a high-resolution image, so that the subject can be accurately detected from the image.

  According to the present invention, a subject in an image can be detected quickly and accurately.

It is a schematic block diagram which shows the imaging device of 1st Embodiment. It is a schematic block diagram which shows the 1st detection part of 1st Embodiment. It is a figure which illustrates a rectangle feature. It is a schematic block diagram which shows the 2nd detection part of 1st Embodiment. It is a figure which shows the relationship between the detection rate and overdetection rate in a 1st detection part and a 2nd detection part. It is a figure which illustrates a teacher image. FIG. 7 is a diagram illustrating a rectangular feature of the teacher image in FIG. 6. It is a flowchart which shows the face detection control of 1st Embodiment. It is a figure which illustrates the local pattern of a high-order local autocorrelation. It is a figure explaining the rectangular feature of the 1st extraction filter and the 2nd extraction filter. It is a schematic block diagram which shows the 1st detection part of 4th Embodiment. It is a figure explaining the 1st extraction filter of 4th Embodiment. It is a schematic block diagram which shows the imaging device of 5th Embodiment. It is a schematic block diagram which shows the detection part of 5th Embodiment. It is a flowchart which shows the face detection control of 5th Embodiment. It is a schematic block diagram which shows the imaging device of 6th Embodiment.

  An imaging apparatus according to a first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a schematic block diagram of the imaging apparatus according to the first embodiment.

  The imaging apparatus according to the first embodiment includes an image acquisition unit 1, a first reduction unit 2, a low resolution image acquisition unit 3, a first detection unit 4, a second reduction unit 5, and a high resolution image acquisition unit 6. And a second detection unit 7. Each of the above parts (or all of them) may be composed of a logic circuit. Alternatively, each of the above-described units (or all of them) may be composed of a memory for storing data, a memory for storing an arithmetic program, a CPU (central processing unit) for executing the arithmetic program, an input / output interface, and the like.

  The image acquisition unit 1 acquires an image from the outside.

  The first reduction unit 2 acquires the first image by reducing the image acquired by the image acquisition unit 1 at a set reduction rate. The size of the subject in the input image is unknown. Therefore, the first reduction unit 2 can acquire the first image by reducing at a plurality of reduction ratios.

  The subject is an object that the photographer pays attention to when photographing, and is, for example, a human face, a human hand, or an animal face, but is not limited thereto. Note that the imaging apparatus can also detect a human hand that is shot relatively small, and can recognize a human gesture. In the present embodiment, a case where the subject is a human face will be described.

  The low resolution image acquisition unit 3 acquires a low resolution image from the first image. The low-resolution image is an image of an area divided by a predetermined size in the first image. In the present embodiment, the low resolution image is an image of an area divided by 10 × 10 pixels. In this embodiment, the low resolution image is an image of 15 × 15 pixels or less, and the high resolution image is an image of 16 × 16 pixels or more. However, it is not limited to this.

  When the low-resolution image acquisition unit 3 performs face candidate determination by the first detection unit 4 described later on the acquired low-resolution image and determines that the acquired low-resolution image does not include a face candidate, A new low-resolution image is acquired from one image. The low resolution image is sequentially acquired in all areas in the first image. The low resolution image acquisition unit 3 sequentially acquires low resolution images while shifting the first image vertically and horizontally by a predetermined amount. The low resolution image acquisition unit 3 sets the low resolution image, for example, first in the upper left of the first image, and then shifts it to the left by a predetermined pixel, for example, 1 pixel, and acquires the low resolution image. When the low-resolution image reaches the right end of the first image, the low-resolution image acquisition unit 3 acquires the low-resolution image by shifting it downward by a predetermined pixel, for example, 1 pixel.

  The first detection unit 4 will be described with reference to FIG. FIG. 2 is a schematic block diagram of the first detection unit 4. The first detection unit 4 includes a first feature extraction unit (first feature amount calculation unit) T1i-j (i = 1 to L, j = 1 to M) and a first determination unit H1i (i = 1 to L). With. The first detection unit 4 detects subject candidates for the low-resolution image acquired by the low-resolution image acquisition unit 3.

  The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) is an identifier that determines whether the low-resolution image acquired by the low-resolution image acquisition unit 3 includes the feature of the face candidate. is there. The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) is, for example, a first feature configured by a rectangular feature shown in FIG. 3 or a rectangular feature obtained by combining the rectangular features shown in FIG. One extraction filter (first feature) is provided. The first extraction filter is a filter that determines whether or not a low-resolution image has a feature of a face candidate. That is, the first extraction filter represents the feature of the face candidate. The first extraction filter has the same size as the low-resolution image, and is 10 × 10 pixels in this embodiment. Note that the rectangular feature is not limited to that shown in FIG. 3, and any feature can be used as long as the face candidate can be identified.

The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) calculates the feature amount of the low resolution image. The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) superimposes the low-resolution image and the first extraction filter, and the low-resolution image corresponding to the rectangular feature of the first extraction filter. The feature amount is calculated based on the pixel value. For example, when the rectangular features of the first extraction filter are FIGS. 3A, 3B, and 3D, the feature amount is
Feature amount = (sum of white rectangular pixel values of low resolution image) − (sum of black rectangular pixel values of low resolution image)
Is calculated by For example, when the rectangular feature of the first extraction filter is FIG.
Feature amount = (sum of white rectangular pixel values of low resolution image) − (sum of black rectangular pixel values of low resolution image) × 2
Is calculated by

  The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) compares the calculated feature quantity with the first threshold value, and if the feature quantity is larger than the first threshold value, the low resolution “1” indicating that the image includes the feature of the face candidate is output. On the other hand, when the feature amount is equal to or smaller than the first threshold, “0” indicating that the low-resolution image does not include the feature of the face candidate is output. The first threshold is a value set for each first feature extraction unit T1i-j (i = 1 to L, j = 1 to M).

  The first determination unit H1i (i = 1 to L) is a discriminator configured by one or more first feature extraction units H1i-j (i = 1 to L, j = 1 to M). The first determination unit H1i (i = 1 to L) is a value (“1” or “0”) output from each first feature extraction unit T1i-j (i = 1 to L, j = 1 to M). Is multiplied by the weight (reliability) of the first extraction filter of each first feature extraction unit T1i-j (i = 1 to L, j = 1 to M), and each first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) are calculated. Then, the first determination unit H1i (i = 1 to L) calculates a total evaluation value obtained by adding the evaluation values of the first feature extraction units T1i-j (i = 1 to L, j = 1 to M). . The first determination unit H1i (i = 1 to L) compares the calculated total evaluation value with the second threshold value, and when the total evaluation value is larger than the second threshold value, the first determination is performed on the low-resolution image next. Part H1i (i = 1 to L). For example, the first determination unit H11 calculates a total evaluation value obtained by adding the evaluation values of the first feature extraction units T11-1 to T11-M. When the total evaluation value is larger than the second threshold value of the first determination unit H11, the low resolution image is input to the next first determination unit H12. On the other hand, when the total evaluation value is equal to or smaller than the second threshold value, the first determination unit H1i (i = 1 to L) determines that the low resolution image does not include a face candidate, and the currently acquired low resolution The face candidate detection determination in the image ends. The second threshold is a value set for each first determination unit H1i (i = 1 to L).

  The first detection unit 4 is configured by cascading a plurality of first determination units H1i (i = 1 to L). When the last first determination unit H1n determines that the low-resolution image includes a face candidate, the low-resolution image is finally determined to be an image including a face candidate. That is, face candidates exist in the acquired image. Thus, the 1st detection part 4 detects a face candidate from a low resolution image. On the other hand, with respect to the low-resolution image in the entire range of the first image acquired by changing the reduction ratio, the total evaluation value is less than or equal to the second threshold value in any of the first determination units H1i (i = 1 to L). If determined, no face candidate is detected from the first image. That is, no face candidate exists in the acquired image.

  The second reduction unit 5 acquires a second image from the image acquired by the image acquisition unit 1. The second image is a higher resolution image than the first image acquired by the first reduction unit 2. For example, the second reduction unit 5 acquires the input image as it is, or acquires the second image by reducing it at a reduction rate smaller than that of the first reduction unit 2. The second reduction unit 5 acquires the second image only when a face candidate is detected in the first image by the first detection unit 4.

  The high resolution image acquisition unit 6 acquires a high resolution image from the second image. The high-resolution image is an image of an area divided by a predetermined size in the second image. In the present embodiment, the high resolution image is an image of an area divided by 24 × 24 pixels. The high-resolution image acquisition unit 6 acquires a new high-resolution image from the second image when face detection determination is performed by the second detection unit 7 described later in the acquired high-resolution image. The new high-resolution image acquisition method is the same as the low-resolution image acquisition method in the low-resolution image acquisition unit 3.

  The 2nd detection part 7 is demonstrated using FIG. FIG. 4 is a schematic block diagram of the second detection unit 7. The second detection unit 7 includes a second feature extraction unit (second feature amount calculation unit) T2i-j (i = 1 to L, j = 1 to M) and a second determination unit H2i (i = 1 to L). With. The second detection unit 7 detects a face from the second image.

  The second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) is an identifier that determines whether the high resolution image acquired by the high resolution image acquisition unit 6 includes facial features. is there. The second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) is, for example, a first feature configured by a rectangular feature shown in FIG. 3 or a rectangular feature obtained by combining the rectangular features shown in FIG. Two extraction filters (second feature) are provided. The second extraction filter is a filter that determines whether or not the high-resolution image has facial features. The rectangular feature of the second extraction filter is, for example, a rectangular feature as shown in FIG. The second extraction filter has the same size as the high resolution image, and is 24 × 24 pixels in the present embodiment. Note that the rectangular feature is not limited to that shown in FIG.

  The second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) calculates the feature amount of the high resolution image. The feature amount calculation method in the second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) is the same as the first feature extraction unit T1i-j (i = 1 to L, j = 1 to M). The same way.

  The second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) compares the calculated feature quantity with the third threshold value, and if the calculated feature quantity is greater than the third threshold value, , “1” indicating that the high-resolution image includes facial features is output. On the other hand, when the calculated feature amount is equal to or smaller than the third threshold, “0” indicating that the high-resolution image does not include facial features is output. The third threshold value is a value set for each second feature extraction unit T2i-j (i = 1 to L, j = 1 to M).

  The second determination unit H2i (i = 1 to L) is a discriminator configured by one or a plurality of second feature extraction units H2i-j (i = 1 to L, j = 1 to M). The second determination unit H2i (i = 1 to L) is a value (“1” or “0”) output from each second feature extraction unit T2i-j (i = 1 to L, j = 1 to M). Is multiplied by the weight (reliability) of the second extraction filter of each second feature extraction unit T2i-j (i = 1 to L, j = 1 to M), and each second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) are calculated. Then, the second determination unit H2i (i = 1 to L) calculates a total evaluation value obtained by adding the evaluation values of the second feature extraction units T2i-j (i = 1 to L, j = 1 to M). . The second determination unit H2i (i = 1 to L) compares the calculated total evaluation value with the fourth threshold value, and when the total evaluation value is larger than the fourth threshold value, the second determination is performed for the next second determination. Part H2i (i = 1 to L). On the other hand, the second determination unit H2i (i = 1 to L) determines that the high-resolution image does not include a face when the total evaluation value is equal to or less than the fourth threshold, and the currently acquired high-resolution image The face detection determination at is terminated. The fourth threshold value is a value set for each second determination unit H2i (i = 1 to L).

  The second detection unit 7 is configured by cascading a plurality of second determination units H2i (i = 1 to L). If the final second determination unit H2n determines that the high-resolution image includes facial features, the high-resolution image is finally determined to be an image including a face. In this way, the second detection unit 7 detects the face from the high resolution image. On the other hand, when it is determined that any of the second determination units H2i (i = 1 to L) determines that the total evaluation value is equal to or lower than the fourth threshold for the high resolution image acquired in the entire range of the second image. This means that no face has been detected from the second image. That is, the acquired image does not include a face.

  Here, the relationship between the detection rate and the overdetection rate in the first detection unit 4 and the second detection unit 7 will be described with reference to FIG. FIG. 5 is a ROC (Receiver Operating Characteristic) curve of the detection rate and the overdetection rate in the first detection unit 4 and the second detection unit 7. The overdetection rate indicates an actual face candidate or face correct answer rate with respect to the number determined by the first detection unit 4 or the second detection unit 7 as a face candidate or a face.

  In the first detection unit 4 and the second detection unit 7, when the detection rate increases, the overdetection rate also increases. In the present embodiment, the first detection unit 4 and the second detection unit 7 are set such that the overdetection rate of face candidates in the first detection unit 4 is larger than the overdetection rate of faces in the second detection unit 7. Set. It is desirable that the face detection rate and the face candidate detection rate be substantially the same. However, since the determination conditions are different between the first detection unit 4 and the second detection unit 7 such that the resolution of the teacher image used in learning is different between the second detection unit 7 and the first detection unit 4, the detection rate is not necessarily the same. It doesn't mean.

  In the present embodiment, when the face candidate overdetection rate is made larger than the face overdetection rate, the first detection unit 4 detects the face candidate without escaping as much as possible, and when the face candidate is detected, Whether or not there is a face in the image is accurately determined by the second detection unit 7.

  Next, a construction method of the first determination unit H1i (i = 1 to L) and the second determination unit H2i (i = 1 to L) will be described. The first determination unit H1i (i = 1 to L) and the second determination unit H2i (i = 1 to L) are both a large number of teacher images (learning images) that include faces and a large number of teacher images that do not include faces ( It is constructed by a learning method of boosting using a learning image. In this embodiment, it is especially constructed by the AdaBoost learning method. Since the learning method of AdaBoost is disclosed in detail in Non-Patent Document 1, detailed explanation is omitted, but learning using AdaBoost is performed by the following procedure. (1) Prepare a teacher image. (2) The learning weight is initialized. (2) Normalize the weight of each teacher image. (3) (a) Normalize the learning weight. (B) Calculate the identification error for each feature. (C) Select the feature that minimizes the identification error. (D) Update the learning weight. (4) Build a final extraction filter.

  The second determination unit H2i (i = 1 to L) is constructed using a high-resolution teacher image (second learning image). In this embodiment, a 24 × 24 pixel image is used as the high-resolution teacher image. The first determination unit H1i (i = 1 to L) is constructed using a low-resolution teacher image (first learning image). In the present embodiment, a 10 × 10 pixel image is used as a low-resolution teacher image. The low-resolution teacher image is created by reducing the high-resolution teacher image. Alternatively, a high-resolution teacher image and a low-resolution teacher image may be captured and created.

  By the learning described above, the first extraction filter and the second extraction filter are also constructed. Here, the first extraction filter and the second extraction filter will be described in detail. During the learning, for example, the image shown in FIG. 6A is used as the high-resolution teacher image including the face, and the image shown in FIG. 6A is reduced as the low-resolution teacher image including the face, for example. Assume that the image shown in FIG. 7A is used. When the images shown in FIG. 6A and FIG. 7A are expressed using rectangular features, the rectangular features of each teacher image are, for example, FIG. 6B and FIG. 7B. In such a case, the first extraction filter and the second extraction filter are constructed based on, for example, the rectangular features shown in FIGS. 6B and 7B.

  The low-resolution teacher image has a smaller amount of information than the high-resolution teacher image because the boundaries indicating the facial features such as eyes, nose, and mouth are blurred compared to the high-resolution teacher image. For this reason, the rectangular features obtained from the low-resolution teacher image are fewer than the rectangular features obtained from the high-resolution teacher image. That is, the first extraction filter that is constructed by learning using the low-resolution teacher image has fewer variations of the rectangular features than the second extraction filter that is constructed by learning using the high-resolution teacher image.

  As described above, since the rectangular feature of the low-resolution image has fewer variations than the rectangular feature of the high-resolution image for a certain face, the low-resolution image acquired by the low-resolution image acquisition unit 3 is the face of the first detection unit 4. It cannot be accurately determined whether the candidate feature is included. However, for example, the first determination unit H1i (i = 1 to L) is also used for a face that is small in the image acquired by the image acquisition unit 1 and difficult to be accurately identified as a face when reduced to the first image. ) Can be detected as a face candidate. This is because the first determination unit H1i (i = 1 to L) is constructed by learning using a low-resolution teacher image. When it is determined that there is a face candidate in the first image, the face can be accurately detected by performing face determination using the high-resolution image by the second detection unit 7.

  Next, the face detection control of this embodiment will be described with reference to the flowchart of FIG.

  In step S100, a first image is acquired from the input image. When the step returns from step S104 described later, a first image reduced at a new reduction rate is acquired.

  In step S101, a low resolution image is acquired from the first image. When the step returns from step S103 described later, a new low resolution image is acquired.

  In step S102, it is determined whether the acquired low-resolution image includes a face candidate. That is, it is determined whether a face candidate is detected from the low resolution image. If it is determined that the low-resolution image includes a face candidate, the process proceeds to step S105. On the other hand, if it is determined that the low-resolution image does not include a face candidate, the process proceeds to step S103.

  In step S103, it is determined whether or not there is a portion in the first image from which a low resolution image has not been acquired. If there is a portion in the first image where the low-resolution image is not acquired, the process returns to step S101, a new low-resolution image is acquired from the first image, and the above control is repeated. On the other hand, if a low resolution image is acquired in the entire range of the first image, the process proceeds to step S104.

  In step S104, it is determined whether or not the first image has been acquired with all the set reduction ratios. When the first image is acquired with all the set reduction ratios, it is determined that there is no face candidate in the input image, and this control is terminated. On the other hand, if there is a reduction ratio that is not selected among the set reduction ratios, the process returns to step 100 to acquire the first image reduced at the new reduction ratio, and the above control is repeated.

  In step S105, a second image is acquired from the input image. When the step returns from step S104 described later, a second image reduced at a new reduction rate is acquired. Although a case where the input image is reduced to acquire the second image will be described here, the input image may be used as it is as the second image.

  In step S106, a high resolution image is acquired from the second image. When the step returns from step S108 described later, a high-resolution image is newly acquired.

  In step S107, it is determined whether the acquired high resolution image includes a face. That is, it is determined whether a face is detected from the high resolution image. When it is determined that the high resolution image includes a face, a signal indicating that the acquired high resolution image is a face is output.

  In step S108, it is determined whether or not there is a portion in the second image from which a high resolution image has not been acquired. If there is a portion in the second image where the high-resolution image is not acquired, the process returns to step S106, a new high-resolution image is acquired, and the above control is repeated. On the other hand, if a high-resolution image is acquired in the entire range of the second image, the process proceeds to step S109.

  In step S109, it is determined whether the second image has been acquired with all the set reduction ratios. When the second image is acquired with all the set reduction ratios, this control is terminated. On the other hand, if there is a reduction ratio that is not selected among the set reduction ratios, the process returns to step S105, a second image reduced at the new reduction ratio is acquired, and the above control is repeated.

  In the present embodiment, when a face candidate is detected from the first image, the second image is acquired immediately. However, the face candidate may be detected in the entire range of the first image. Further, the face candidates may be detected by acquiring the first image with all the set reduction ratios. And when acquiring a high resolution image from a 2nd image, you may acquire a high resolution image only in the vicinity of the location corresponding to the location where the face candidate was detected in the 1st image, and may perform face detection. Face detection control can be performed quickly by acquiring a high-resolution image and performing face detection only in the vicinity of a location where a face candidate is detected.

  The first determination unit H1i (i = 1 to L) and the second determination unit H2i (i = 1 to L) may use a support vector machine as a learning method. The support vector machine constitutes a separation plane that separates learning data into two classes. The separation plane is set at a position where the distance between the learning data and the separation plane becomes the maximum with reference to the learning data located near the class boundary. Learning methods using support vector machines are described in, for example, David Meyer, a, Friedrich Leischa and Kurt Hornikb. "The support vector machine under test" in Neurocomputing,] Volume 55, Issues 1-2, September 2003, Pages 169-186 Has been.

  In the present embodiment, the first image and the second image are acquired by reducing the image acquired by the image acquisition unit 1. However, an image acquisition unit that acquires a low-resolution image, and a high-resolution image are acquired. An image acquisition unit for acquiring the image may be provided separately.

  The effect of 1st Embodiment of this invention is demonstrated.

  Based on the image acquired by the image acquisition unit 1, a low resolution image is acquired, and face candidates are detected for the low resolution image. When a face candidate is detected for the low-resolution image, face detection is performed on the high-resolution image acquired from the image acquired by the image acquisition unit 1. This makes it possible to accurately detect a face when the acquired image includes a face. Further, when there is no face candidate in the image acquired by the image acquisition unit 1, the face detection control can be quickly terminated because the face detection by the high resolution image is not performed.

  By making the overdetection rate of face candidates in the first detection unit 4 greater than the overdetection rate of faces in the second detection unit 7, face candidates in the first image can be detected without omission. Accordingly, it is possible to prevent the first detection unit 4 from determining that there is no face candidate in the first image even when there is a face in the image, and to accurately detect the face.

  Each feature amount of the low resolution image is calculated, and it is determined whether the low resolution image is a face candidate based on the calculated feature amount. Similarly, for a high-resolution image, it is determined whether the high-resolution image is a face. As a result, face candidate detection for a low resolution image and face detection for a high resolution image can be accurately performed.

  By learning a face candidate from a low-resolution teacher image, face candidate detection in the first determination unit H1i (i = 1 to L) can be accurately performed. In addition, by learning a face with a high-resolution teacher image, face detection in the second determination unit H2i (i = 1 to L) can be accurately performed.

  Face candidates can be detected accurately by learning and creating the first extraction filter with a low-resolution teacher image. Similarly, the face detection can be accurately performed by learning and creating the second extraction filter with a high-resolution teacher image.

  By creating a first extraction filter using a low-resolution teacher image obtained by reducing a high-resolution teacher image, the first detection unit 4 can accurately detect a face appearing small in the image as a face candidate. Can do. Also, by creating the first extraction filter and the second extraction filter based on the same teacher image, the results of face candidate detection in the first detection unit 4 and face detection in the second detection unit 7 can be brought closer to each other. The face detection can be performed accurately.

  By configuring the first extraction filter and the second extraction filter with rectangular features, face detection control can be performed quickly.

  By quickly and accurately performing face candidate detection and face detection by learning the first determination unit H1i (i = 1 to L) and the second determination unit H2i (i = 1 to L) by a learning method using boosting. Can do.

  Next, a second embodiment of the present invention will be described.

  In the present embodiment, the features in the first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) and the second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) The amount calculation method is different from that of the first embodiment. The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) and the second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) are higher-order local autocorrelation functions. The feature amount is calculated using. The higher-order local autocorrelation function is an extension of the autocorrelation, and if the reference point is r and the target image is f (r), the Nth-order autocorrelation function is the displacement (a1, a2, ... , aN) is defined by

  At this time, the region in which the displacement region is limited to the local region of r is called high-order local autocorrelation. FIG. 9 shows an example of a local pattern when the order N is limited to the second order and the area is limited to 3 × 3 pixels. FIG. 9A is a local pattern when the order N is 0th order, FIG. 9B is an example of a local pattern when the order N is 1st order, and FIG. It is an example of a local pattern when the order N is secondary. The feature amount at the reference point r is calculated by multiplying the pixel value of the black circle and the pixel value of the white circle in FIG.

  The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) includes the first extraction filter, for example, a filter having a local pattern shown in FIG. The first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) superimposes the first extraction filter and the low resolution image, and corresponds to the local pattern of the first extraction filter. The feature amount is calculated based on the pixel value. When the first extraction filter is composed of a plurality of local patterns, the feature amount is calculated by adding the products of the pixel values corresponding to the local patterns.

  The second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) is characterized by the same method as the first feature extraction unit T1i-j (i = 1 to L, j = 1 to M). Calculate the amount.

  The effect of 2nd Embodiment of this invention is demonstrated.

  By configuring the first extraction filter and the second extraction filter with higher-order local autocorrelation, face detection control can be performed quickly.

  Next, a third embodiment of the present invention will be described.

  In the present embodiment, the feature amount calculation in the second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) is reduced as compared with the first embodiment.

  In the present embodiment, the size of the low resolution image and the first extraction filter is 10 × 10 pixels, and the size of the high resolution image and the second extraction filter is 20 × 20 pixels. That is, the size of the high resolution image and the second extraction filter is twice the size of the low resolution image and the first extraction filter.

  The second extraction filter of this embodiment will be described. Assume that the rectangular features of the first extraction filter and the second extraction filter are the features shown in FIG. For the sake of explanation, it is assumed that the first extraction filter and the second extraction filter have three rectangular features. FIG. 10A shows the first extraction filter, and FIG. 10B shows the second extraction filter.

  The rectangular features in FIG. 10A are α, β, and γ, and the rectangular features in FIG. 10B are α ′, β ′, and γ ′. Comparing the positions of the rectangular features α, β, and γ with the positions of the rectangular features α ′, β ′, and γ ′, the relative positions of the rectangular feature α and the rectangular feature α ′ with respect to each filter are the same. However, the relative positions of the rectangular features β and γ and the rectangular features β ′ and γ ′ with respect to the filters do not match.

  Further, the size of the rectangular feature α ′ is twice the size of the rectangular feature α, and matches the size ratio of the first extraction filter and the second extraction filter. However, the size ratio of the rectangular feature γ and the rectangular feature γ ′ does not match the size ratio of the first extraction filter and the second extraction filter.

  That is, the position and size of the rectangular feature α with respect to the first extraction filter and the position and size of the rectangular feature α ′ with respect to the second feature extraction filter are relatively coincident.

  In such a case, the first extraction filter having the rectangular feature α and the second extraction filter having the rectangular feature α ′ show the same detection result in face candidate detection and face detection. For example, a first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) including a first extraction filter having a rectangular feature α converts a low-resolution image into an image including face candidate features. If it is determined that there is a second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) having a second extraction filter having a rectangular feature α ′, the high-resolution image also includes facial features. Is determined to be an image. Therefore, even if the determination in the second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) is omitted, the face detection result is not affected.

  In the present embodiment, the second feature extraction unit T2i-j (i = 1 to L, j including a second extraction filter having a rectangular feature whose relative position and size match the rectangular feature of the first extraction filter. = 1 to M), the feature amount calculation is omitted. Such a second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) does not calculate a feature amount, and the first feature extraction unit T1i-j (i = 1 to L, j The determination result in = 1 to M) is substituted. For example, when the first feature extraction unit T1i-j (i = 1 to L, j = 1 to M) outputs “1” indicating that the low-resolution image includes the feature of the face candidate, the first extraction filter rectangle The second feature extraction unit T2i-j (i = 1 to L, j = 1 to M) including the second extraction filter having a rectangular feature whose relative position and size match the feature also outputs “1”. To do.

  Note that the second extraction filter may be created by deleting a rectangular feature whose relative position and size match the rectangular feature of the first extraction filter.

  In the present embodiment, the case where the size of the high-resolution image and the second extraction filter is twice the size of the low-resolution image and the first extraction filter has been described, but the present invention is not limited to this. The size of the high-resolution image and the second extraction filter is an integral multiple of the size of the low-resolution image and the first extraction filter, and it is only necessary that the relative position and size match.

  The effect of the third embodiment of the present invention will be described.

  In the second feature extraction unit including the second extraction filter having a rectangular feature whose relative position and size match the rectangular feature of the first extraction filter, face detection control is performed by omitting calculation of the feature amount. Can be done quickly.

  Next, a fourth embodiment of the present invention will be described with reference to FIG. FIG. 11 is a schematic block diagram of the first detection unit 4 of the imaging apparatus according to the fourth embodiment.

  The imaging device of this embodiment differs in the 1st detection part 10 compared with the imaging device of 1st Embodiment. Since other configurations are the same as those in the first embodiment, description thereof is omitted here. In the present embodiment, the high resolution image is an image of an area divided by 20 × 20 pixels.

  The first feature extraction unit T11i-j (i = 1 to L, j = 1 to M) uses a filter obtained by reducing the second extraction filter as the first extraction filter. That is, in the present embodiment, the second extraction filter is constructed by the teacher image of the high-resolution image at the time of learning, and the first feature extraction unit T11i-j (i = 1 to L, j = 1 is used by using the second extraction filter. To M), it is determined whether the low-resolution image includes the feature of the face candidate. The first extraction filter is 10 × 10 pixels.

  A method of using the second extraction filter as the first extraction filter will be described with reference to FIG.

  Here, it is assumed that the rectangular feature of the second extraction filter is, for example, the feature shown in FIG. In FIG. 12A, rectangular features α, β, γ, and θ are shown for explanation.

  The rectangular features α, β, γ in FIG. 12A are reduced to ½ according to the size of the low-resolution image, and α ′, β ′, γ ′ are defined as α ′, β ′, γ ′, and FIG. Shown in (c). Here, a description will be given by setting coordinates (x, y) with the lower left of each figure in FIG. 12 as the origin. Coordinates are in units of 1 pixel. Therefore, a pixel grid that is a boundary between adjacent pixels corresponds to the coordinates.

  The coordinates of the vertices α1 to α6 of the rectangular feature α are α1 (4, 10), α2 (8, 10), α3 (12, 10), α4 (12, 16), α5 (8, 16) α6 (4, 16). Therefore, when the rectangular feature α is reduced, the vertices α1 ′ to α6 ′ of the rectangular feature α ′ are represented by α1 ′ (2, 5), α2 ′ (4, 5), α3 ′ ( 6, 5), α4 ′ (6, 8), α5 ′ (4, 8), α6 ′ (2, 8), and the vertices α1 ′ to α6 ′ overlap the intersections of the pixel grid.

  The coordinates of the vertices β1 to β6 of the rectangular feature β are β1 (7, 1), β2 (11, 1), β3 (15, 1), β4 (15, 7), β5 (11, 7), β6 (7 7). Here, when the rectangular feature β is reduced, the vertices corresponding to the vertices β1 to β6 of the rectangular feature β do not overlap with the intersection of the pixel grid. In this case, the vertex is moved based on a certain rule. In the present embodiment, the vertex that does not overlap the intersection of the pixel grid is moved to the right, lower, or lower right grid. According to this rule, the rectangular feature β ′ is represented by β1 ′ (4,0), β2 ′ (6,0) β3 ′ (8,0), β4 ′ (8,3), β5 in FIG. '(6, 3), β6' (4, 3), and the vertices β1 'to β6' overlap the intersections of the pixel grid.

  The coordinates of the vertices γ1 to γ6 of the rectangular feature γ are γ1 (14, 15), γ2 (17, 15), γ3 (20, 15), γ4 (20, 18), γ5 (17, 18), γ6 (14 18). In this case, some of the vertices corresponding to the vertices γ1 to γ6 of the rectangular feature γ do not overlap the intersections of the pixel grid. On the other hand, when the same rule as the rectangular feature β is applied, in FIG. 12B, γ1 ′ (7, 7), γ2 ′ (9, 7), γ3 ′ (10, 7), γ4 ′ (10 , 9), γ5 ′ (9, 9), γ6 ′ (7, 9). When the above rule is applied, the vertex of the rectangular feature γ ′ overlaps the intersection of the pixel grid, but the area of the black rectangle (first feature region) and the area of the white rectangle (second feature region) The ratio is different from the ratio of the rectangular feature γ before the reduction. In such a case, the large area is reduced by 1 pixel to obtain γ1 ″ (8, 7), γ6 ″ (8, 9), so that a black rectangular shape as shown in FIG. The ratio of the area and the area of the white rectangle is made the same before and after the reduction.

  In order to make the ratio of the black rectangular area and the white rectangular area the same before and after reduction, the weights of the black rectangle and the white rectangle may be changed. Also by this, the second extraction filter can be reduced and used as the first extraction filter.

  Note that the rectangular feature θ in FIG. 12A is deleted because it cannot be composed of a 1-pixel black rectangle and a 1-pixel white rectangle, which are the smallest units when reduced.

  In the present embodiment, the features of the low-resolution image are obtained by the first feature extraction unit T11i-j (i = 1 to L, j = 1 to M) using the first extraction filter obtained by reducing the second extraction filter by the above method. Calculate the amount.

  The first determination unit H11i (i = 1 to L) uses the first feature extraction unit T11i-j (i = 1 to L, j = 1 to M) as a face candidate for the low resolution image. Is determined.

  Note that the first extraction filter may be enlarged and used as the second extraction filter.

  The effect of 4th Embodiment of this invention is demonstrated.

  By reducing the second extraction filter according to the ratio of the size of the low resolution image and the high resolution image and using the second extraction filter as the first extraction filter, the amount of data stored in the imaging apparatus can be reduced.

  When the second extraction filter is reduced, the second extraction filter can be used as the first extraction filter by reducing the vertex of the reduced rectangular feature so as to overlap the intersection of the pixel elements.

  When reducing the second extraction filter, the vertex of the reduced rectangular feature is moved so that the ratio of the black rectangular area and the white rectangular area of the rectangular feature is the same as the ratio before the reduction. Thus, the second extraction filter can be used as the first extraction filter.

  Next, an imaging device according to a fifth embodiment of the present invention will be described with reference to FIG. FIG. 13 is a schematic block diagram of an imaging apparatus according to the fifth embodiment.

  The imaging apparatus of the present embodiment includes an image acquisition unit 1, a reduction unit 21, a high resolution image acquisition unit 22, a low resolution image acquisition unit 23, and a detection unit 24.

  The reduction unit 21 has the same configuration as the second reduction unit 5 of the first embodiment, and acquires the second image from the image acquired by the image acquisition unit 20.

  The high resolution image acquisition unit 22 acquires a high resolution image from the second image acquired by the reduction unit 21. The high resolution image is an image of an area divided by 20 × 20 pixels in the second image.

  The low resolution image acquisition unit 23 acquires the low resolution image by reducing the high resolution image at a predetermined reduction rate. In the present embodiment, the low resolution image is a 10 × 10 pixel image obtained by reducing the high resolution image to ½.

  The detection unit 24 will be described with reference to FIG. FIG. 14 is a schematic block diagram of the detection unit 24. The detection unit 24 is configured by cascading a plurality of determination units Hk (k = 1 to n, n: integer of 2 or more). The detection unit 24 includes a first detection unit 25 and a second detection unit 26.

  The first detection unit 25 includes m determination units Hm (m: an integer of 1 or more, m <n, hereinafter, this determination unit Hk will be described as the first determination unit Hm) among the determination units Hk. Is done.

  The first determination unit Hm is a discriminator that determines whether the low-resolution image acquired by the low-resolution image acquisition unit 23 includes the feature of the face candidate. Similar to the first feature extraction unit T11i-j (i = 1 to L, j = 1 to M) of the fourth embodiment, the first determination unit Hm uses a filter obtained by reducing the second extraction filter as the first extraction filter. Is used to calculate the feature amount of the low-resolution image. Then, the first determination unit Hm determines that the low-resolution image includes the feature of the face candidate when the feature amount is larger than the first threshold value.

  If the m-th first determination unit Hm determines that the low resolution image includes the feature of the face candidate, the first detection unit 25 determines that the low resolution image is an image including the face candidate.

  The second detection unit 26 includes m + 1 to nth determination units Hn (hereinafter, this determination unit will be described as the second determination unit Hn) in the determination unit Hk.

  The second determination unit Hn is a discriminator that determines whether the high-resolution image acquired by the high-resolution image acquisition unit 22 includes facial features. When the first detection unit 25 determines that the low resolution image includes a face candidate, the second determination unit Hn calculates the feature amount of the high resolution image that created the low resolution image. Then, the second determination unit Hn determines that the high-resolution image includes facial features when the feature amount is larger than the second threshold.

  When the n-th second determination unit Hn determines that the high-resolution image includes facial features, the second detection unit 26 determines that the high-resolution image includes a face.

  Next, face detection control of this embodiment will be described with reference to the flowchart of FIG.

  In step S200, a second image is acquired from the input image. When the step returns from step S206 described later, a second image reduced at a new reduction rate is acquired.

  In step S201, a high resolution image is acquired from the input second image. When the step returns from step S205 described later, a new high-resolution image is acquired.

  In step S202, the acquired high resolution image is reduced at a predetermined reduction rate to acquire a low resolution image.

  In step S203, it is determined whether the low resolution image includes a face candidate. If it is determined that the low-resolution image includes a face candidate, the process proceeds to step S204. On the other hand, if it is determined that the low-resolution image does not include a face candidate, the process proceeds to step S205.

  In step S204, it is determined whether or not the high-resolution image acquired in step S201 includes a face. If it is determined that the high-resolution image includes a face, a signal indicating that the input second image includes a face is output.

  In step S205, it is determined whether or not there is a portion in the second image from which a high resolution image has not been acquired. If there is a portion in the second image where the high-resolution image is not acquired, the process returns to step S201, a new high-resolution image is acquired from the second image, and the above control is repeated. On the other hand, when a high-resolution image is acquired in the entire range of the second image, the process proceeds to step S206.

  In step S206, it is determined whether the second image has been acquired with all the set reduction ratios. When the second image is acquired with all the set reduction ratios, this control is terminated. On the other hand, if there is a reduction ratio that is not selected among the set reduction ratios, the process returns to step 200 to acquire a second image reduced at the new reduction ratio.

  In this embodiment, the low resolution image is created by reducing the high resolution image. However, the high resolution image may be created by acquiring the low resolution image and enlarging the low resolution image.

  Similarly to the first embodiment, a low-resolution image and a high-resolution image are prepared, the first detection unit performs face candidate determination using the low-resolution image, and the second detection unit uses the high-resolution image. Face determination may be performed.

  The effect of 5th Embodiment of this invention is demonstrated.

  If the high-resolution image is reduced to obtain a low-resolution image, and it is determined that the low-resolution image is not a face candidate, face detection is not performed using the high-resolution image, and the low-resolution image is determined to be a face candidate. Only when it is determined, face detection determination is performed using a high-resolution image. Thereby, it is possible to reduce the frequency of face detection using a high-resolution image that takes time to process, and to perform face detection quickly.

  A sixth embodiment of the present invention will be described with reference to FIG. FIG. 16 is a schematic block diagram of the sixth embodiment. In the present embodiment, a digital still camera (hereinafter referred to as a camera) is used as the imaging apparatus. In the present embodiment, the face detection control described in the first embodiment is performed.

  The camera includes an image sensor 30, an A / D converter 31, a first reduction unit 32, a low resolution image acquisition unit 33, a first detection unit 34, a second reduction unit 35, and a high resolution image acquisition unit. 36, a second detection unit 37, an image display unit (display unit) 38, a memory (storage unit) 39, and a CPU (control unit) 40.

  The image sensor 30 outputs an analog signal at a predetermined timing in accordance with light incident on the light receiving surface through the optical system 41. The imaging device 30 is, for example, a format called a CCD (charge coupled device) or a CMOS (complementary metal oxide semiconductor) sensor, or other various types of imaging devices.

  The A / D converter 31 converts an analog signal output by the image sensor into a digital signal as image data.

  The first reduction unit 32 reduces the image obtained by the image sensor and acquires the first image. The first image is displayed on the image display unit 38 as a through image. The through image displayed on the image display unit 38 is an image reduced at a predetermined reduction rate. The low resolution image acquisition unit 33 acquires a low resolution image from the first image. The first detection unit 34 detects face candidates from the low resolution image.

  When the first detection unit 34 determines that there is a face candidate in the first image, the actual detection is performed, and the second reduction unit 35 acquires the second image.

  The high resolution image acquisition unit 36 acquires a high resolution image from the second image. The acquired high resolution image is stored in the memory 39. Note that a high-resolution image may be acquired by continuous shooting. The second detection unit 37 detects a face from the high resolution image. If no face is detected by the second detection unit 37, the high resolution image is deleted from the memory 39.

  When deleting a high-resolution image from the memory 39, a folder for storing a deletion candidate image such as a trash can may be provided, and the high-resolution image to be deleted may be temporarily stored therein. Alternatively, the user may be asked to confirm the deletion, and the deletion may be performed when the user instructs the deletion. This can prevent a necessary image from being deleted.

  The image display unit 38 displays a through image. The image display unit 38 may display the portion corresponding to the face detected by the second detection unit 7 in the through image, for example, surrounded by a rectangular frame. This can indicate to the user that a face has been detected.

  The CPU 40 controls the entire camera. The CPU 40 executes various programs of the camera by executing a program stored in the memory.

  The effect of the sixth embodiment of the present invention will be described.

  When a face candidate is detected from a low-resolution image, a high-resolution image is acquired, and the acquired high-resolution image is stored in the memory 39, so that face detection can be performed quickly.

  When a face is not detected by a high resolution image, the capacity of the memory 39 can be reduced by deleting the high resolution image stored in the memory 39.

  In addition, it is not restricted to the said embodiment, It is possible to combine the structure of the said embodiment.

  In addition, the imaging apparatus can be mounted on an electronic device such as a digital camera, a digital video camera, or an electronic endoscope, which is a device that depends on an electric current or an electromagnetic field in order to operate correctly.

  In the above description of the embodiment, the processing performed by the imaging apparatus is premised on processing by hardware, but is not necessarily limited to such a configuration. For example, a configuration in which processing is performed separately by software is also possible.

  In this case, the imaging apparatus includes a main storage device such as a CPU and a RAM, and a computer-readable storage medium in which a program for realizing all or part of the above processing is stored. Here, this program is called an image processing program. Then, the CPU reads out the image processing program stored in the storage medium and executes information processing / calculation processing, thereby realizing processing similar to that of the imaging apparatus.

  Here, the computer-readable recording medium refers to a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, and the like. Alternatively, the image processing program may be distributed to a computer via a communication line, and the computer that has received the distribution may execute the image processing program.

  It goes without saying that the present invention is not limited to the above-described embodiments, and includes various modifications and improvements that can be made within the scope of the technical idea.

1, 20 Image acquisition unit 2, 32 First reduction unit 3, 23, 33 Low resolution image acquisition unit 4, 10, 25, 34 First detection unit 5, 35 Second reduction unit 6, 22, 36 High resolution image acquisition Unit 7, 26, 37 Second detection unit 21 Reduction unit 24 Detection unit 30 Image sensor 31 A / D converter 38 Image display unit (display unit)
39 Memory (storage unit)
40 CPU (control unit)
T1i-j, T11i-j first feature extraction unit (first feature amount calculation unit)
H1i, H11i 1st judgment part
Hk discriminator
T2i-j second feature extraction unit (second feature amount calculation unit)
H2i second judgment part

Claims (32)

An imaging device for imaging a subject,
An image acquisition unit for acquiring images;
A low-resolution image acquisition unit that acquires a low-resolution image based on the image acquired by the image acquisition unit;
A first detection unit for detecting a subject candidate for the low-resolution image;
A high-resolution image acquisition unit that acquires a high-resolution image that is higher in resolution than the low-resolution image from the image acquired by the image acquisition unit;
An imaging apparatus comprising: a second detection unit configured to detect the subject from the high resolution image when the subject candidate is detected from the low resolution image by the first detection unit. .   The high-resolution image acquisition unit acquires the high-resolution image from the image acquired by the image acquisition unit when the subject candidate is detected for the low-resolution image by the first detection unit. The imaging apparatus according to claim 1, wherein the imaging apparatus is characterized.   The imaging apparatus according to claim 1, wherein the low-resolution image acquisition unit acquires the low-resolution image by reducing the high-resolution image.   The over-detection rate of the subject candidates, which is the correct rate of the subject candidates with respect to the number of subject candidates detected by the first detection unit, is the correct answer of the subject with respect to the number of subjects detected by the second detection unit. The imaging apparatus according to claim 2, wherein the imaging device is larger than an overdetection rate of the subject, which is a rate. The first detection unit includes:
A first feature amount calculation unit for calculating a feature amount of the low-resolution image;
A first determination unit that determines whether or not the subject candidate is included in the low-resolution image based on the feature amount calculated by the first feature amount calculation unit;
The imaging apparatus according to claim 2, wherein the first detection unit detects the subject candidate for the low-resolution image based on a determination result by the first determination unit.   Whether the first determination unit includes the subject candidate in the low-resolution image by learning using a plurality of first learning images that are an image including the subject candidate and an image not including the subject candidate image. 6. The imaging apparatus according to claim 5, wherein whether or not to determine is determined.   The feature of the low-resolution image is based on a first feature learned using a plurality of first learning images that are an image including the subject candidate and an image not including the subject candidate image. The imaging apparatus according to claim 5, wherein an amount is calculated.   The first learning image is an image obtained by reducing a plurality of second learning images that are a high-resolution image including the subject and a high-resolution image not including the subject. The imaging device described.   The imaging apparatus according to claim 7, wherein the first feature is configured by a rectangular feature or a higher-order local autocorrelation.   The imaging apparatus according to claim 6, wherein the first determination unit learns by a learning method using boosting.   The imaging apparatus according to claim 6, wherein the first determination unit learns by a learning method using a support vector machine. The second detector is
A second feature amount calculation unit for calculating a feature amount of the high-resolution image;
A second determination unit that determines whether or not the high-resolution image includes the subject based on the feature amount calculated by the second feature amount calculation unit;
The imaging apparatus according to claim 2, wherein the second detection unit detects the subject from the high-resolution image based on a determination result by the second determination unit.   The second determination unit learns using a plurality of second learning images that are a high-resolution image including the subject and a high-resolution image not including the subject, thereby including the subject in the high-resolution image. The imaging apparatus according to claim 12, wherein it is determined whether or not the image is to be captured.   The second feature amount calculation unit is configured to use the high resolution based on a second feature learned using a plurality of second learning images that are a high resolution image including the subject and a high resolution image not including the subject. The image pickup apparatus according to claim 12, wherein a feature amount of the image is calculated.   The imaging apparatus according to claim 14, wherein the second feature includes a rectangular feature or a higher-order local autocorrelation.   The imaging apparatus according to claim 13, wherein the second determination unit learns by a learning method using boosting.   The imaging apparatus according to claim 13, wherein the second determination unit learns by a learning method using a support vector machine.   The high-resolution image acquisition unit acquires the high-resolution image from the image acquired by the image acquisition unit only near a position corresponding to the low-resolution image from which the subject candidate is detected by the first detection unit. The imaging apparatus according to claim 2. The first detection unit includes:
A first feature amount calculation unit that calculates a feature amount of the low-resolution image based on a first feature learned by a plurality of first learning images that are an image including the subject candidate and an image not including the subject candidate image; ,
A first determination unit that determines whether or not the subject candidate is included in the low-resolution image based on the feature amount calculated by the first feature amount calculation unit, the second detection unit,
The resolution of the high-resolution image is higher than that of the first learning image, based on the second feature learned by a plurality of second learning images that are images including the subject candidate and images not including the subject candidate image. A second feature amount calculation unit for calculating a feature amount;
A second determination unit that determines whether or not the high-resolution image includes the subject based on the feature amount calculated by the second feature amount calculation unit;
The second feature amount calculation unit calculates a feature amount of the high-resolution image for a second feature having a relative position and size that is identical to the first feature in the second feature. The imaging apparatus according to claim 2, wherein the imaging apparatus is not. The first detection unit includes:
A first feature amount calculation unit that calculates a feature amount of the low-resolution image based on a first feature;
A first determination unit that determines whether or not the subject candidate is included in the low-resolution image based on the feature amount calculated by the first feature amount calculation unit;
The second detector is
A second feature amount calculation unit that calculates a feature amount of the high-resolution image based on a second feature;
A second determination unit that determines whether or not the high-resolution image includes the subject based on the feature amount calculated by the second feature amount calculation unit;
The second feature is learned by a plurality of learning images that are a high-resolution image including the subject and a high-resolution image not including the subject,
The first feature amount calculation unit is configured to reduce the low feature based on the first feature obtained by reducing the second feature learned by the learning image according to a ratio of the size of the high resolution image and the low resolution image. The imaging apparatus according to claim 2, wherein a feature amount of the resolution image is calculated. The first feature and the second feature include a feature composed of a first feature region and a second feature region,
When the reduction is performed, if the vertices of the first feature region and the second feature region after the reduction do not overlap with the intersection of the pixel lattice that is a boundary between adjacent pixels, the vertex is the pixel lattice. The imaging apparatus according to claim 20, wherein the vertex is moved so as to overlap the intersection of the two.   The ratio between the first feature region and the second feature region before the reduction is adjusted by adjusting the areas of the first feature region and the second feature region after the reduction. The imaging device according to claim 21, wherein a ratio between the first feature region and the second feature region after the reduction is made equal.   The ratio between the first feature region and the second feature region before the reduction is adjusted by adjusting the weight between the first feature region and the second feature region after the reduction. The imaging device according to claim 21, wherein a ratio between the first feature region and the second feature region after the reduction is made equal. The first detection unit includes m (m is an integer of 1 or more) first determination units,
The second detection unit includes n (n is an integer of m + 1 or more) second determination units,
The imaging device according to claim 3, wherein the first detection unit and the second detection unit are connected in series.   The storage unit that stores the high-resolution image acquired by the high-resolution image acquisition unit when the subject candidate is detected for the low-resolution image by the first detection unit. The imaging device described in 1.   A controller that deletes the high-resolution image stored in the storage unit when the subject is not detected by the second detection unit with respect to the high-resolution image stored in the storage unit; The imaging apparatus according to claim 25, characterized in that:   A control unit that moves the high-resolution image stored in the storage unit to an image deletion folder when the subject is not detected by the second detection unit with respect to the high-resolution image stored in the storage unit The imaging apparatus according to claim 25, comprising:   28. The imaging apparatus according to claim 27, wherein the control unit obtains permission to delete the high-resolution image after moving the high-resolution image to the image deletion folder.   26. The imaging apparatus according to claim 25, further comprising a display unit that displays the subject surrounded by a frame when the second detection unit detects the subject with respect to the high-resolution image.   An electronic apparatus comprising the imaging apparatus according to any one of claims 1 to 29. An image processing program for processing a captured image by a computer,
An image acquisition procedure for acquiring images;
A low-resolution image acquisition procedure for acquiring a low-resolution image based on the image acquired by the image acquisition procedure;
A first detection procedure for detecting subject candidates for the low-resolution image;
A high-resolution image acquired from the image acquired by the image acquisition procedure and having a higher resolution than the low-resolution image when the subject candidate is detected for the low-resolution image by the first detection procedure And a second detection procedure for detecting the subject. An imaging method for imaging a subject,
Get an image,
Acquiring a low-resolution image based on the acquired image;
Detect subject candidates for the low-resolution image,
When the subject candidate is detected for the low-resolution image, the subject is detected from a high-resolution image acquired from the acquired image and having a higher resolution than the low-resolution image. An imaging method.