JP5565869B2 - Identification device and identification program - Google Patents

Description

  The present invention relates to an identification device and an identification program for identifying whether or not a detection target is included in image data.

  In fields where online processing is required, such as face detection in digital cameras and pedestrian detection in driving support for automobiles, object detection from images requires high-speed identification processing as well as detection accuracy. However, in general, an identification device with higher identification performance tends to have a larger amount of calculation, so it is difficult to satisfy these two requirements at the same time.

  As a method for detecting a detection target from an image at high speed using pattern recognition, a technique for configuring a strong classifier with high discrimination performance by combining a plurality of weak classifiers with low discrimination performance has been proposed.

  Non-Patent Document 1 discloses a strong classifier configured by arranging a plurality of weak classifiers in series. This strong classifier discriminates a face and a non-face for each weak classifier by comparing the output with a threshold value, and the input data of an area identified as a non-face by any weak classifier Processing of the later weak classifier is omitted. As a result, since most of the region of the image is a non-face region, it is possible to speed up the detection process by using such a cascade structure classifier. Patent Document 1 discloses a configuration that speeds up detection processing by setting the lowest similarity among the weak classifiers as the threshold value of the weak classifier.

Paul Viola and Michael Jones, “Rapid Object Detection using a Boosted Cascade of Simple Features,” Proc. Of CVPR2001, vol.1, pp. 511-518, 2001. JP 2009-251962 A

  By the way, in the technical field in which a plurality of weak classifiers are combined and an image is identified by a strong classifier, it is extremely important to speed up the detection process. Therefore, there is a demand for an identification device that can speed up the detection process compared to the conventional configuration.

  It is an object of the present invention to provide an identification device and an identification program that can speed up detection processing.

  The present invention is an identification device, wherein a feature value generation unit that generates a feature value of input data, and whether or not the feature value exists within a predetermined range sandwiched between a first threshold value and a second threshold value. A weak discriminator including a determination unit, a weight setting unit set so that the weight increases as the discrimination performance between the detection target and the non-detection target increases, and each of the weak discriminators, A strong discriminator for discriminating whether or not the input data is a detection target based on a linear sum of the integrated value of the feature value determination result and the weight in the weak discriminator.

  According to the above configuration, when comparing the probability distribution of the feature quantity of the detection target and the probability distribution of the feature quantity of the non-detection target, the two-sided probability of the feature quantity of the non-detection target is more than the two-sided probability of the feature quantity of the detection target. Is generally large. Therefore, when the presence / absence of the feature amount is determined based on a predetermined range sandwiched between the first threshold value and the second threshold value, a large amount of input data to be detected is included within the predetermined range, Thus, a large amount of non-detection target input data is included in the lower region below the first threshold and the upper region exceeding the second threshold. As a result, the identification can be performed more effectively than the conventional case where it is determined whether or not the input data is a detection target using the one side region and the other side region separated by one threshold value. Since a large area can be secured, the discrimination performance of each weak discriminator can be improved. As a result, a strong classifier having high performance can be obtained with a small number of weak classifiers, and thus the detection process can be speeded up.

  Further, the strong classifier in the present invention, the weak classifier set of the weak classifier is arranged in series in a plurality of stages, only the input data identified as the detection target in the weak classifier set of the previous stage, It may be used for identifying the detection target in the weak classifier set part.

  According to the above configuration, since the weak classifiers with improved discrimination performance are provided in the weak classifier aggregation units at each stage, the input of non-detection targets that are not detection targets, particularly in the early stage of identification of detection targets Since a large amount of data can be eliminated, the identification process for identifying whether the input data is a detection target can be completed in a short time.

  Further, the weight setting unit in the weak classifier according to the present invention is an Adaboost using a plurality of positive learning data to be detected and a plurality of negative learning data to be non-detected as input data. The weight may be set by a learning algorithm.

  According to said structure, the weight in each weak discriminator can be set easily and in a short time.

  In addition, the determination unit in the weak classifier according to the present invention uses, as the first threshold, an intersection at which a lower probability of a feature amount of a non-detection target is greater than a lower probability of the feature amount of the detection target, An intersection at which the upper probability of the feature quantity of the detection target is larger than the upper probability of the feature quantity of the detection target may be set as the second threshold value.

  According to said structure, a 1st threshold value and a 2nd threshold value can be set easily.

  Further, the present invention is an identification program, comprising: a computer having an input unit to which input data is input, a storage unit, and an output unit; a feature amount generating unit that generates a feature amount of the input data; A determination means for determining whether or not a predetermined range sandwiched between the first threshold value and the second threshold value is set, and the weight is set to be larger as the discrimination performance between the detection target and the non-detection target is higher. A plurality of weak discriminating means comprising weight setting means; based on a linear sum of integrated values of the determination result of the feature quantity and the weight in each of the weak discriminating means, whether or not the input data is the detection target It functions as a strong identification means for identifying.

  According to said structure, an identification apparatus can be constructed | assembled by the simple operation | work of installing an identification program with respect to information processing apparatus provided with the computer.

  The present invention can speed up the detection process.

It is explanatory drawing which shows the learning mode and identification mode in an identification device. It is explanatory drawing of a rectangle pair, (A) is the state in which the 1st rectangle was included in the 2nd rectangle, (B) is the state in which the 1st rectangle and the 2nd rectangle partly overlapped, (C) is The first rectangle and the second rectangle are completely separated from each other. It is explanatory drawing of the rectangular pair selected as a weak discriminator. It is explanatory drawing which shows a 1st threshold value and a 2nd threshold value. It is explanatory drawing which shows the relationship between a hyperplane and a weak discriminator. It is a block diagram of a cascade structure. It is explanatory drawing which shows the relationship between the image in a window, and a weak classifier. It is a schematic diagram of a face image area. It is a schematic diagram of a face image area. It is a flowchart of a learning mode process routine. It is a flowchart of a weak classifier candidate preparation processing routine. It is a flowchart of a neighborhood search processing routine. It is a flowchart of a strong discriminator construction processing routine. It is a flowchart of a strong identification process routine. It is a front view which shows the pachinko machine which added the game media rental apparatus (sand).

(Identification device 10)
Embodiments of the present invention will be described with reference to the drawings.
As shown in FIG. 1, the identification device 10 according to the present embodiment identifies a learning mode in which a plurality of weak classifiers 2 are learned from learning data to generate a strong classifier 1, and input data is identified by the strong classifier 1. It is possible to switch to the identification mode. The identification device 10 includes a strong classifier 1 and a plurality of weak classifiers 2 that constitute the strong classifier 1. Each weak discriminator 2 determines whether or not the feature amount exists in a predetermined range sandwiched between the first threshold value and the second threshold value, and the feature amount generation unit 3 that generates the feature amount of the input data 4 and a weight setting unit 5 set so that the weight increases as the discrimination performance between the detection target and the non-detection target increases. The strong classifier 1 identifies whether or not the input data is a detection target based on the linear sum of the integrated value of the determination result of the feature amount and the weight in each weak classifier 2. .

  Here, “learning data” is sample data in which a plurality of positive learning data to be detected and a plurality of negative learning data to be non-detected are prepared in advance. For example, if the detection target is a person's face, the image data of the person's face becomes positive learning data, and image data other than the person's face such as the whole body, arms, other animals, buildings, etc. becomes negative learning data. .

(Learning algorithm)
“Learning” is a process of weighting the weak classifier 2 using positive learning data and negative learning data. The learning algorithm for weighting is not particularly limited, but the Adaboost algorithm is exemplified. The outline of the Adaboost algorithm will be described. First, a weak classifier candidate (weak classifier 2 before learning) is created. Thereafter, the learning data (features and classes) is initialized with 1 / N of the number of samples N, so that all the learning data is set to the same weight. In each learning data, the error rate of each weak classifier candidate is obtained by integrating the weights of the learning data in which the class of the learning data and the output class of the weak classifier candidate do not match. The weight of the learning data is updated so that the reliability increases as the error rate decreases. That is, the weight of the learning data is updated so that the weight is reduced for the learning data in which the weak classifier candidate is correctly identified, while the weight is increased for the learning data that the weak classifier candidate has incorrectly identified. Updated to Thereafter, normalization is performed so that the sum of the weights of the learning data becomes 1, so that one weak classifier candidate is created. After such weak classifier candidates are repeatedly created, all or a predetermined number of weak classifier candidates are selected as the weak classifiers 2.

(Feature value)
The “feature amount” is obtained by digitizing image data, and is not particularly limited as long as the similarity can be obtained by comparing with other feature amounts. For example, in the case of a Haar-like feature, in order to determine whether it is a detection target or a non-detection target, the feature amount uses the difference between the average luminance value in the white rectangular region and the average luminance value in the black rectangular region. In this case, high-speed computation is possible by using Integral Image.

(Weak classifier 2, weak classifier candidate)
The weak classifier 2 and the weak classifier candidate are a relationship in which the weak classifier 2 before learning is a weak classifier candidate, and are the same except for the state of weight. Similar to the configuration proposed by Viola et al., The weak classifier 2 and the weak classifier candidate may be a feature consisting of two rectangles, a feature consisting of three rectangles, a feature consisting of four rectangles, or in a pixel high-dimensional space. In order to maximize the degree of freedom and to speed up the determination, it may be configured to handle only the lightness difference between the two rectangular regions.

  For example, when a window with a predetermined number of pixels is set and a task is given whether a face image is contained in this window, when considering a high-dimensional space in which the luminance value of each pixel is an independent variable It can be seen that the hyperplane set that delimits the half-space defined by the Haar-like weak classifier is only a small part of the total hyperplane set. The hyperplane set will be described later. Therefore, the restriction of the Haar-like weak classifier for setting the boundary that more effectively separates the negative region from the positive region expected to occupy a finite region in the high-dimensional space is that the search space is It is expected to be necessary if a method that does not become enormous can be prepared separately.

  Viola et al. Used a feature consisting of two rectangles, a feature consisting of three rectangles, and a feature consisting of four rectangles. From the above prediction, it is recognized that the necessity of Haar-like is unnecessary. Accordingly, only the following conditions may be imposed on the combination of two rectangular areas. Each of the two rectangular regions can be used as a weak classifier candidate by combining all the rectangles satisfying two conditions of (1) being included in the detection window and (2) not completely overlapping each other. Examples of possible rectangular pairs are shown in FIGS. FIG. 2A shows a state in which the first rectangle 31 is included in the second rectangle 32. FIG. 2B shows a state where the first rectangle 31 and the second rectangle 32 partially overlap each other. FIG. 2C shows a state where the first rectangle 31 and the second rectangle 32 are completely separated from each other.

  FIG. 3 shows an example of a rectangular pair in which a front face image is learned as a positive sample using the above method and selected as a weak classifier having the highest discriminating power. The outer rectangle is relatively dark because it includes both eyes and mouth, and the inner rectangle is relatively bright because it does not include both eyes and mouth. Both eyes and mouth are always darker than the surroundings because shadows occur in most light source environments. On the other hand, since somewhere in the rectangular area becomes bright, it does not become completely dark. Therefore, the brightness difference between the two rectangles falls within a range from a certain value to a certain value, and the accuracy can be improved by making a determination with two threshold values.

  In the Viola-Jones method, the number of weak classifier candidates exceeds 180,000 if the classifier is configured with 24 × 24 pixels, but the number of weak classifier candidates further increases in the combination method of two rectangular areas that are further relaxed. To do. For example, 24 × 24 pixels is 3.9 billion or more. Therefore, a heuristic method may be applied to shorten the calculation time. Specifically, the amount of change in the rectangular size and position of the weak classifier candidate is not set for each pixel, but for example, every three pixels. As a result, the number of weak classifier candidates is reduced to 360,000 with 24 × 24 pixels. This significant omission may cause a rectangular pair with high discriminating power to leak out of the candidates, but a rectangular pair with higher discriminating power may be found by a neighborhood search.

  After selecting the best candidate weak classifiers according to the boosting algorithm from the first prepared weak classifier candidates, a weak classifier candidate group is generated by slightly changing parameters such as the rectangle width and rectangle position of the candidate, and some local classifiers are generated. A weak classifier candidate with a performance close to optimum is found by finding the best one by a search method (for example, a hill-climbing method) and selecting it again.

  For the operation of taking the luminance value average of all the pixels in the rectangle, if the size of the rectangle is large to some extent, it can be expected that the luminance value between the rectangles with a slight change in the coefficient will not be significantly different, Since the classifier performance can also be expected to change relatively smoothly, this neighborhood search is expected to work well.

The following parameters can be changed during local search. Two threshold values d1 and d2
2. 2. White rectangle X, Y coordinates White rectangle width Xw, length Yw
4). 4. Black rectangular coordinates Black rectangle horizontal width, vertical width Here, “white rectangle” and “black rectangle” are convenient names. For example, a reference coordinate on the left is “white rectangle”, and a right is “black rectangle”. (When the X coordinates are the same, the upper one is called “white rectangle” and the lower one is called “black rectangle”). Which rectangle has the higher brightness is determined by the distribution of the actual positive samples regardless of the designation, and is reflected in the values of the threshold values d1 and d2.

(First threshold, second threshold)
As shown in FIG. 4, the “first threshold value and second threshold value” are one side (lower region) and the other side (upper region) of the normal distribution, which is the distribution of the difference in brightness sum of the two rectangular regions. And a threshold value provided for each. The first threshold is an intersection where the lower probability of the feature quantity of the non-detection target is larger than the lower probability of the feature quantity of the detection target, and the second threshold is the upper probability of the feature quantity of the non-detection target The intersection may be larger than the upper probability of the feature quantity to be detected. In this case, the first threshold value and the second threshold value can be easily set. The output h (x) of the weak classifier 2 is “1” if the feature quantity x exists in the range from the first discrimination value θ1 to the second threshold value θ2, and “0” if it does not exist.

  The first threshold value and the second threshold value may be obtained by the following method. In the following description, a threshold value determination method when a cascade structure in which the weak classifiers 2 are arranged in series is described. However, even in a structure other than the cascade structure, a determination method for the cascade first stage and the cascade second and subsequent stages is also described. May be adopted as a method for determining an initial value before adjustment. Details of the cascade structure will be described later.

In the first stage of the cascade, threshold values are set at positions (both sides of the bell curve) where most of the positive input images (within 3σ = 99.73%) of the weak classifier are positively determined. After obtaining the mean and variance, it is set assuming a normal distribution. Specifically, the first threshold value d _p1 and the second threshold value d _p2 are expressed by the following relational expressions when the arithmetic mean μ _p and variance σ _p of the brightness difference distribution of the rectangular pairs in the positive learning image are used.

d _p1 = μ _p −3σ _p
d _p2 = μ _p + 3σ _p

In the second and subsequent stages of the cascade, assuming that the lightness difference distribution of the negative learning image and the lightness difference distribution of the positive learning image are normal distributions, the intersection of the normal distribution curves is set to the first threshold value d _p1 and the first threshold value d _p1 . 2 Set to the initial value of the threshold value _dp2 . However, the value is not used as it is, and thereafter, the threshold is adjusted by changing to a threshold that reduces the weighted loss of adaBoost by the hill-climbing method in the adaBoost learning loop. If the arithmetic mean μ _p and variance σ _p of the lightness difference distribution of the rectangular pair in the positive learning image and the arithmetic mean μ _n and variance σ _n of the lightness difference distribution of the rectangular pair in the negative learning image are two normal distribution curves. The intersection of can be obtained by the following equation.

When x is expanded, it becomes a quadratic equation as shown below, which is solved.

  The first threshold value and the second threshold value are applied to all weak classifiers 2 including the latter stage of the cascade structure. Thereby, oblique face identification can be performed effectively. The reason for this is that when considering a high-dimensional vector space having pixels in the rectangular detection window as elements, a rectangular weak classifier having the constraints of Viora et al. It is presumed that the combination of hyperplanes that can eliminate this is exhausted. That is, as shown in FIG. 5, if the allowed hyperplane pair is a region, a two-threshold weak discriminator having a discriminating power more than a one-threshold weak discriminator can be found at a later stage by relaxing the constraint. It is presumed that it has become (region b in FIG. 5).

  More specifically, in the case where an image with relaxed constraints is shown in the schematic diagram of FIG. 5, a high-dimensional space is considered in which the luminance values of all the pixels in the detection window of this image are independent variables. At this time, it was not found when there was a restriction that only a boundary surface with a multiple of 45 ° inclination could be set, but a new 2-threshold weak discriminator for the b region was found by allowing other angles. . In the Lienhart et al. Paper, when calculating the brightness difference between the white rectangular area and the black rectangular area, the total brightness of each rectangular area is corrected by the area ratio, so the inclination of the boundary surface in the high-dimensional space is fixed by the area ratio. (A boundary pair in the figure). On the other hand, in this embodiment, since the correction coefficient of the white rectangular area and the black rectangular area is changed by the hill-climbing method, the b boundary pair in the figure is also adopted as a two-threshold weak discriminator.

  As described above, the identification device 10 includes the weak classifier 2 that determines whether or not the feature amount is within a predetermined range sandwiched between the first threshold value and the second threshold value. When comparing the probability distribution of the feature quantity of the detection target and the probability distribution of the feature quantity of the non-detection target, a general relationship is that the two-sided probability of the feature quantity of the non-detection target is larger than the two-sided probability of the feature quantity of the detection target It is. Therefore, when the presence / absence of the feature amount is determined based on a predetermined range sandwiched between the first threshold value and the second threshold value, a large amount of input data to be detected is included within the predetermined range, Thus, a large amount of non-detection target input data is included in the lower region below the first threshold and the upper region exceeding the second threshold. As a result, the identification can be performed more effectively than the conventional case where it is determined whether or not the input data is a detection target using the one side region and the other side region separated by one threshold value. Since a large area can be secured, the discrimination performance of each weak discriminator 2 can be enhanced. As a result, the strong classifier 1 having high performance can be obtained with a small number of weak classifiers 2, and the detection processing of the detection target can be speeded up.

(Cascade structure)
Moreover, as shown in FIG. 6, the strong discriminator 1 in the discriminating apparatus 10 has a cascade structure. Specifically, the strong classifier 1 includes one or more weak classifier aggregation units 21 of the weak classifier 2 arranged in series in a plurality of stages, and an input identified as a detection target in the weak classifier aggregation unit 21 in the previous stage. Only the data is configured to be used for identifying the detection target in the weak classifier set 21 in the subsequent stage. Note that the identification device 1 may have a configuration other than the cascade structure.

  As a result, the strong classifier 1 includes the weak classifier 2 whose discrimination performance is enhanced by the first threshold value and the second threshold value in the weak classifier set unit 21 at each stage, and thus particularly starts identifying the detection target. In the initial stage (first stage side), it is possible to eliminate a large amount of non-detection target input data that is not a detection target, so that identification of whether or not the input data is a detection target can be completed in a short time.

(Explanation of weak classifier 2 with two thresholds in high-dimensional space)
The role of the Haar-like weak classifier 2 is equivalent to partitioning a high-dimensional vector space having each pixel as a direct component with a hyperplane.
Positive image Pk = (Pk ₁ , Pk ₂ , Pk ₃ ,... Pk _n ) composed of n pixels
It consists of n pixels negative image _{Nk = (Nk 1, Nk 2} , Nk 3, ··· Nk n)
A positive image set is expected to occupy a finite region connected in this high-dimensional space. If the positive image set is surrounded by a finite connection region, the Haar-like weak classifier 2 can be sandwiched between two parallel hyperplanes. Therefore, it is reasonable to set two threshold values for the Haar-like weak classifier 2. By combining a plurality of Haar-like weak classifiers 2, the positive image connection area is surrounded by a convex hull.

  In the latter stage of the cascade, the negative images that enter the convex hull surrounding the positive image connection region increase, and it becomes difficult to enclose the two parallel hyperplanes. Therefore, when the discrimination capability of the weak discriminator 2 deteriorates to a certain level or less, the threshold value may be switched from two threshold values.

Here, “can be sandwiched between two parallel hyperplanes” means that, for example, as shown in FIG. 7, a window of 16 horizontal pixels × 16 vertical pixels = 256 pixels is set, and a face image is set in this window. □ ■ □ type Haar-like weak discriminator made of 3 pixels of X ₆₉ , X ₇₀ , and X ₇₁ that will be around the right eye if there is a face in this window The determination formula of 2 is as follows.

_{(X 69 + X 71) -2X} 70> d (1)
Here, consider a 256-dimensional space in which the luminance value of each of the 256 pixels is an independent variable.
V = (X ₁ , X ₂ ,... X ₂₅₆ )
The general formula of the half space when this space V is divided into two by the hyperplane is as follows.

A ₁ X ₁ + A ₂ X ₂ +... A ₂₅₆ X ₂₅₆ ≧ d (2)
Equation (1) is a special form of Equation (2). (A ₁ = A ₂ = ... A ₆₈ = 0, A ₆₉ = 1, A ₇₀ = -2, A ₇₁ = 1, A ₇₂ = A ₇₃ = ... A ₂₅₆ = 0, d = d)
That is, when considering a 256-dimensional space V in which the luminance value of a pixel is an independent variable, the Haar-like weak classifier 2 is given as a half space that partitions V by a hyperplane.
The hyperplane at this time is shown by the following equation.

A ₁ X ₁ + A ₂ X ₂ +... A ₂₅₆ X ₂₅₆ = d (3)
A hyperplane parallel to this hyperplane is given by
A ₁ X ₁ + A ₂ X ₂ +... A ₂₅₆ X ₂₅₆ = d ′ (d ′ ≠ d) (4)
The weak classifier 2 having two thresholds is obtained by extending the Haar-like weak classifier 2 given by, for example, the expression (1) as the following expression.

d ₁ ≧ (X ₆₉ + X ₇₁ ) −2X ₇₀ ≧ d ₂ (5)
Equation (5) represents an inner region separated by two hyperplanes parallel to each other. FIG. 8 shows a schematic diagram of a face image area in a partial space (X ₆₉ , X ₇₀ , X ₇₁ ). The black area shown in a cloud shape is an area that is assumed to be a true face image, and the two planes shown by stripes are the boundary faces of the face image area determined by the weak classifier 2 defined by equation (5). .

Given extend the 8 across 256-dimensional _space, X _69, since the coefficient for the coordinate axes other than X _{70, X 71} is _zero, parallel to the axes other than the _{X 69,} X 70, X ₇₁ The hyperplane becomes the boundary surface of the weak classifier 2 defined by the equation (5). Since the other weak classifiers 2 constituting the ensemble classifier are inequalities having coefficients in other pixels, that is, other coordinate axes, hyperplanes having inclinations in those coordinate axes are defined.

  When considered in the n-dimensional Euclidean space, it cannot be said that the above condition alone is a finite region, but since the luminance value itself is actually finite (for example, 0 to 255), the luminance value space is also finite, The subset of positive images, which is a subset, is finite.

  When it is reasonable that the positive image set is expected to occupy a finite region connected in a high-dimensional space, the face region is surrounded by a weak classifier 2 having a series of two thresholds as shown in FIG. Will be. On the other hand, if the region is not a connected region, the face region is surrounded by the weak classifier 2 having two threshold values for categorization and boosting.

(Identification program)
In the above description, the case where the identification device 10 is configured by hardware has been described. However, the present invention is not limited to this, and the identification device 10 may be installed by installing an identification program in a computer.

  The identification program includes a computer having an input unit, a storage unit, and an output unit to which input data such as learning data and detection data is input, a feature amount generation unit that generates a feature amount of the input data, and the feature amount is a first threshold value Determining means for determining whether or not the object exists within a predetermined range sandwiched between the second threshold and a weight setting means for setting the weight to be larger as the discrimination performance between the detection target and the non-detection target is higher As a strong identification means for identifying whether or not the input data is a detection target based on a linear sum of the integrated value of the determination result of the feature amount and the weight in each weak identification means It is to function. Further, the identification program may cause the computer to function as means for making the weak identification means a cascade structure.

  The identification program includes information processing including a target computer by transmitting it through a communication system capable of bidirectional data communication such as the Internet or a transmission system capable of transmitting data in one direction such as television broadcasting and radio broadcasting. Installed on the device. The identification program may be recorded on a recording medium such as a compact disk or a memory stick, and installed in the information processing apparatus using the recording medium.

(Operation of the identification device 10)
In the above configuration, the identification program is transmitted from the software house or the store to the information processing device and downloaded to the storage device of the information processing device. Then, by installing the identification program in the information processing apparatus, an unlearned identification apparatus 10 is constructed as shown in FIG. That is, the weak classifier 2 (information processing apparatus) determines whether or not the feature quantity generation means for generating the feature quantity of the input data and the feature quantity are within a predetermined range sandwiched between the first threshold value and the second threshold value. A plurality of weak discriminating means comprising: a judging means for determining whether or not a weight setting means is set such that the weight increases as the discrimination performance between the detection target and the non-detection target increases. Based on the linear sum of the integrated value of the determination result of the quantity and the weight, the computer is constructed to function as strong identification means for identifying whether or not the input data is a detection target. Furthermore, the weak discriminator 2 (information processing apparatus) is constructed so that the computer functions as a means for making the weak discriminating means a cascade structure.

(Learning mode processing)
Next, the identification device 10 is set to the learning mode of FIG. Specifically, as shown in FIG. 10, a learning mode processing routine is executed. When this routine is executed, a weak classifier candidate preparation process is executed (S1). Thereby, as shown in FIG. 11, weak classifier candidates consisting of all the two rectangular regions in the detection window are formed (S21). Then, the first weak classifier candidates are narrowed down to the first weak classifier candidates for each predetermined pixel from all the weak classifier candidates (S22), and the first weak classifier candidates whose number is reduced are assigned to the storage device (not shown). It is stored in the storage unit (S23).

  When the first weak classifier candidates are prepared as described above, the weights D of all the first weak classifier candidates are initialized as shown in FIG. 10 (S2). Thereafter, one first weak classifier candidate is fetched (S3). Then, one piece of learning data is captured (S4), and a class d (1, -1) corresponding to this learning data is set. For example, a class d of “1” is set if the learning data is positive learning data, and a class d of “−1” is set if the learning data is negative learning data. Then, a feature amount x using the learning data and the first weak classifier candidate is calculated (S6).

  Next, it is determined whether or not a value dθ1 obtained by multiplying the class d by the first threshold θ1 is smaller than a value dx obtained by multiplying the class d by the feature amount x (S7). When the value dθ1 is not smaller than the value dx, that is, when dθ1 <dx is not satisfied (S7: NO), the output h (x) is set to “0” (S9), and the processing of S12 is performed. Executed. On the other hand, if dθ1 <dx is satisfied (S7: YES), the value dθ2 obtained by multiplying the class d by the second threshold θ2 is the value dx obtained by multiplying the class d by the feature amount x. It is determined whether it is larger than (S8).

  When the value dθ2 is not larger than the value dx, that is, when dx <dθ2 is not satisfied (S8: NO), after the output h (x) is set to “0” (S9), the processing of S12 is performed. Executed. On the other hand, if dx <dθ2 is satisfied (S8: YES), the output h (x) is set to “1” (S10). Then, after a predetermined weight is added to the first weak classifier candidate whose output h (x) is set to “1” (S11), the process of S12 is executed.

  When the setting of the output h (x) and the addition of the weight are performed as described above, it is subsequently determined whether or not the acquisition of all the learning data has been completed (S12). When the capturing is not completed (S12: NO), the process is re-executed from S4, and the setting of the output h (x) and the addition of the weight are performed for the next learning data. On the other hand, when the acquisition of all the learning data is completed (S12: YES), normalization is performed such that the sum of the weights of all the learning data becomes 1, thereby learning one learned first weak identification. A container candidate is created and stored (S13). Thereafter, it is determined whether learning of all the first weak classifier candidates is completed (S14). If not completed (S14: NO), the process is re-executed from S3, and the learning for the next unlearned first weak classifier candidate is repeated. Then, when learning of all the first weak classifier candidates is completed (S14: YES), this routine ends.

(Learning mode processing: Neighborhood search processing)
When learning of all the first weak classifier candidates is completed as described above, the neighborhood search processing routine of FIG. 12 is executed. First, the learned first weak classifier candidates are rearranged in order of weight (S31). One first weak classifier candidate is selected in order from the largest weight (S32). All the second weak classifier candidates existing around the selected first weak classifier candidate are acquired. The vicinity of the first weak classifier candidate is performed by changing the weak classifier parameter (S33).

  Thereafter, all the second weak classifier candidates are learned using the learning data. As the learning method, the same method as the learning mode process of FIG. 10 is adopted (S34). Among the selected first weak classifier candidates and all the second weak classifier candidates, the weakest classifier candidate with the maximum weight is registered as the weak classifier 2 (S35). It is determined whether or not selection of all the first weak classifier candidates is completed (S36), and if not completed (S36: NO), the process is re-executed from S32. Then, when completed (S36: YES), this routine is ended. In the present embodiment, neighborhood search processing is performed for all the first weak classifier candidates. However, if the classification performance is high within the range of the number of weak classifiers 2 included in the strong classifier 1, The neighborhood search process may be performed only for the expected first weak classifier candidate.

(Strong classifier construction process)
As described above, when the weak classifier 2 is acquired based on the first weak classifier candidate and the second weak classifier candidate, a strong classifier construction process is then executed. Specifically, as shown in FIG. 13, a plurality of weak classifier aggregation units 21 are created in a predetermined storage area (S41). Thereafter, the learned weak classifiers 2 are rearranged in order of weight (S42). The weak classifiers 2 are extracted in units of a predetermined number in order from the largest weight order (S43). The extracted weak classifier 2 is registered in the unregistered weak classifier set unit 21 (S44).

  It is determined whether or not the registration to all weak classifier sets 21 has been completed (S45), and if not completed (S45: NO), the process is re-executed from S43. Then, when the registration to all weak classifier assembly units 21 is completed (S45: YES), this routine ends. As a result, as shown in FIG. 6, a strong classifier 1 having a cascade structure in which a plurality of weak classifier aggregates 21 are connected in series is formed. In addition, it is preferable to arrange the weak classifiers 2 having high discrimination performance in the weak classifier assembly unit 21 on the front stage side of the cascade structure in that many non-detection targets can be eliminated at the initial stage of discrimination.

(Strong identification processing)
When the strong discriminator 1 is constructed as described above, the discrimination mode shown in FIG. 1 is set. Specifically, as shown in FIG. 14, a strong identification processing routine is executed. First, moving images and still images are taken in as inspection data (input data) from imaging devices installed in passages, doors, and various devices, and this inspection data is input to the first-stage weak classifier aggregation unit 21 (S51). ).

  In the weak classifier set unit 21, the feature amount x using the inspection data and one weak classifier 2 is calculated. Then, it is determined whether or not the condition of first threshold θ1 <feature value x <second threshold θ2 is satisfied (S53). If the condition is not satisfied (S54: NO), the output h (x) = 0 is set (S55). On the other hand, if the condition is satisfied (S54: YES), the output h (x) = 1 is set (S54). Thereafter, the total output value is updated by adding the output h (x) to the total output value (S56).

  It is determined whether or not all weak classifiers 2 have been identified (S57). If the identification is not completed (S57: NO), the process is re-executed from S52. On the other hand, when all the identifications are completed (S57: YES), it is subsequently determined whether or not the total output value is larger than a predetermined value (S58). If the total output value is not larger than the predetermined value (S58: NO), the weak discriminator set 21 in the subsequent stage assumes that the inspection data (input data) is non-detection target data, that is, negative data. Discarded without performing the identification process, and the presence or absence of the next inspection data is determined (S62).

  On the other hand, if the total output value is larger than the predetermined value (S58: YES), the weak discriminator gathering unit 21 in the subsequent stage assumes that the inspection data (input data) is detection target data, that is, positive data. Identification processing is performed. Specifically, it is determined whether or not it is the last weak classifier set unit 21 (S59). If it is not the last weak classifier set unit 21 (S59: NO), the test data is input to the next weak classifier set unit 21 (S61). Then, by re-execution from S52, the weak classifier aggregation unit 21 identifies whether or not the inspection data is a detection target.

  If it is the last weak classifier assembly unit 21 (S59: YES), the inspection data is registered as a detection target (S60). Then, the presence or absence of the next inspection data is determined (S62), and if the next inspection data exists (S62: YES), the process is re-executed from S51. On the other hand, if there is no next inspection data (S62: NO), this routine is terminated.

(Game media lending device)
A game medium lending device provided with the identification device 10 will be described. In the following description, the game medium lending device will be described, but the present invention is not limited to this, and is mounted on all types of devices used for identifying a detection target such as a face from image data. can do. For example, the identification device 10 may be mounted on a gaming machine such as a pachinko or pachislot machine, or may be mounted on an imaging device such as a portable digital camera, a monitoring camera, a door phone, or an interphone. As a mounting form, the configuration may be such that the identification device 10 is provided separately from the control unit of the device to be mounted, or the control unit of the device to be mounted causes the processing of the identification device 10 to be executed. In addition, a program or data may be incorporated.

  The game medium lending device will be specifically described. As shown in FIG. 15, a sand M20 that is a game medium lending device is provided in the pachinko machine M10. Each sand M20 is installed corresponding to the adjacent pachinko machine M10, and is connected to be communicable with the corresponding pachinko machine M10. Each sand M20 is communicably connected to a hall computer that performs system management and sales management of the sand M20 for the entire hall. In the present embodiment, the sand M20 that is communicably connected to the hall computer will be described. However, the present invention is not limited to this, and even a sand that does not communicate with the hall computer may be used. Good.

  On the front portion 21 of each sand M20, an LED (light-emitting diode) M31, a card insertion slot M32, a bill insertion slot M33 into which a bill can be inserted, an operation unit M34 constituted by a touch panel LCD (liquid crystal display), A camera unit M35, a non-contact IC card reader / writer M36, a fraction outlet M40, a counting inlet M39, and the like are provided. The card insertion slot M32 is an insertion slot that can accept an information card issued by a card issuing machine in a hall, for example. The LED unit M31 is configured by a full color LED.

  As the camera unit M35, any imaging system camera may be used as long as it captures an image of a subject (player). This is because the identification process by the strong classifier of the identification device 10 is performed, so that the presence / absence of the player can be determined with high accuracy even in an environment where the background easily changes.

  The camera of the camera unit M35 sends an imaging signal to the control board. The control board incorporates a program and data so as to have the function of the identification device 10, thereby performing identification processing by the strong classifier based on the imaging signal. That is, the control board determines whether or not a face image to be detected exists in the imaging signal (image data) taken in front of the sand M20, and the player based on the result of this identification process. And a player determination process for determining whether or not there exists.

  In the above configuration, operations of the sand M20 and the pachinko machine M10 will be described. When the player sits in front of the pachinko machine M10, the player is photographed by the camera unit M35 of the sand M20 arranged on the side surface of the pachinko machine M10. Then, the presence of the player is detected by executing the identification process and the player determination process using the imaging signal of the camera unit M35. On the condition that the player is present, the operation process of the sand M20 is permitted, and for example, an information card or the like can be received.

  The player can receive a pachinko ball as a game medium necessary for the game by inserting an information card or a predetermined amount of banknotes into the card insertion slot M32 or the banknote insertion slot M33. Also, the player can receive a pachinko ball required for the game by holding the non-contact IC card over the non-contact IC card reader / writer M36.

  When the sand M20 receives an input of a value medium such as an information card, a bill, and a non-contact IC card, the sand M20 transmits a command to pay out a number of pachinko balls corresponding to the amount of the inserted value medium to the pachinko machine M10 provided with the command. By doing so, the pachinko ball is paid out from the pachinko machine M10. The player can play a pachinko game with the pachinko balls paid out to the upper plate M12 in the pachinko machine M10.

  In the pachinko machine M10, the pachinko balls are paid out to the upper plate M12 according to the game result. When the player drops a pachinko ball on the upper plate M12 onto the lower plate M13 by performing a predetermined operation, the pachinko ball falls onto a guide plate M14 mounted below the lower plate M13. The guide tray M14 guides the pachinko balls dropped from the lower tray M13 to the counting slot M39 provided in the sand M20. The pachinko balls introduced into the counting inlet M39 are counted by a counter provided inside the sand M20.

  The counted results are recorded on a card inserted into the card insertion slot M32 or stored in a storage unit provided in the hall computer. The pachinko balls counted in the counting unit are discharged from a discharge port provided on the back surface of the sand M20 and collected. In addition to the pachinko ball counting path inserted from the counting inlet M39, a fraction supply unit is provided on the back side of the sand M20 to receive the fraction pachinko ball from the outside. By this supply unit, a certain amount (for example, 24 balls) of pachinko balls for fractions is always stored, and if necessary, from the fraction outlet M40, a guide plate M14 (specifically, , And is paid out to a dedicated payout tray provided below the guide plate M14.

  When the player is away from the game while the game is being performed as described above, the absence of the player is detected by the identification process and the player determination process using the imaging signal from the camera unit M35. When the detection state in which the player is present is changed to the absence detection state, it is determined whether or not an information card or cash (amount data) remains. When the monetary amount data remains, a warning sound or warning sound is generated. Thus, when the player leaves the seat even though the balance remains, the player can be alerted.

  In the detailed description above, the characteristic portions have been mainly described so that the present invention can be understood more easily. The present invention is not limited to the embodiments described in the detailed description above, but can be applied to other embodiments, and the scope of application is various. The terms and terminology used in the present specification are used to accurately describe the present invention, and are not used to limit the interpretation of the present invention. Moreover, it would be easy for those skilled in the art to infer other configurations, systems, methods, and the like included in the concept of the present invention from the concept of the invention described in this specification. Accordingly, the description of the claims should be regarded as including an equivalent configuration without departing from the scope of the technical idea of the present invention. The purpose of the abstract is to simplify the technical contents and essence of this application by patent offices and general public institutions, and engineers belonging to this technical field who are not familiar with patents, legal terms or technical terms. It is intended to be able to make a quick determination through a simple survey. Accordingly, the abstract is not intended to limit the scope of the invention to be evaluated by the claims. Moreover, in order to fully understand the object of the present invention and the specific effects of the present invention, it is desired that the interpretation should be made with sufficient consideration of the literatures already disclosed.

DESCRIPTION OF SYMBOLS 1 Strong discriminator 2 Weak discriminator 3 Feature-value production | generation part 4 Determination part 5 Weight setting part 10 Discriminating device 21 Weak discriminator collection part

Claims (6)

A feature amount generation unit that generates a feature amount using a difference between an average luminance value of pixels in the white rectangular area of the input data and an average luminance value of pixels in the black rectangular area;
A determination unit for determining whether or not the feature amount is within a predetermined range sandwiched between a first threshold value and a second threshold value;
A weight setting unit that is set so that the weight increases as the discrimination performance between the detection target and the non-detection target increases.
From the white-block region and all of the plurality narrowed from the weak classifier candidate first weak classifier candidate in the detection by the variation of the amount of change or rectangular position of the rectangle width and a plurality of pixels each window of the black rectangular area A selected weak classifier,
A plurality of the weak classifiers, and a strong classifier for identifying whether or not the input data is a detection target based on a linear sum of an integrated value of a determination result of a feature amount and a weight in each weak classifier; An identification device comprising: The strong classifier is
The assembly of the weak classifiers is arranged in series in a plurality of stages, and only input data that is identified as a detection target in the preceding assembly is used for identification of the detection target in the subsequent assembly. 2. The identification device according to 1. The weight setting unit in the weak classifier is:
The weight is set by an Adaboost learning algorithm using the plurality of positive learning data to be detected and the plurality of negative learning data to be non-detected as the input data. The identification device according to claim 1 or 2. The determination unit in the weak classifier is:
The intersection at which the lower probability of the non-detection target feature amount is larger than the lower probability of the detection target feature amount is defined as the first threshold value.
The intersection point at which the upper probability of the feature quantity of the non-detection target is larger than the upper probability of the feature quantity of the detection target is set as the second threshold value. Identification device. Each of the weak classifiers is selected from the plurality of first weak classifier candidates, and then a plurality of second weak classifier candidate groups in which the rectangular width or the rectangular position of the first weak classifier candidates are changed. 5. The identification device according to claim 1, wherein the identification device is generated and further selected from the plurality of second weak classifier candidate groups by a local search method. 6. A computer including an input unit to which input data is input, a storage unit, and an output unit,
Feature amount generation means for generating a feature amount using a difference between an average luminance value of pixels in the white rectangular area and an average luminance value of pixels in the black rectangular area of the input data, the feature amount being a first threshold value and a second threshold value Determination means for determining whether or not a predetermined range sandwiched between, and a weight setting means set so that the weight increases as the discrimination performance between the detection target and the non-detection target increases, selected from white rectangular regions and all of the plurality narrowed from the weak classifier candidate first weak classifier candidate in the detection by the variation of the amount of change or rectangular position of the rectangle width and a plurality of pixels each window of the black rectangular area A plurality of weak identification means,
Based on the linear sum of the integrated value of the feature value determination result and the weight in each of the weak discriminating units, the weak discriminating unit functions as a strong discriminating unit that discriminates whether or not the input data is the detection target. Identification program.