JP2013058162A - Feature quantity extraction device, target object detection system, computer program, and feature quantity extraction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a feature quantity extraction device, a target object detection system, a computer program, and a feature quantity extraction method which can accurately extract feature quantity which presents essential difference between a target object and a non-target object.SOLUTION: A feature quantity candidate determination unit 102 determines a plurality of feature quantity candidates related to displacement of spectral intensity at a plurality of arbitrary wavelengths for each sample pixel. An identifier generation unit 103 generates an identifier for identifying a target object and a non-target object by using boosting for each of a part of feature quantity candidates among the determined feature quantity candidates. A feature quantity extraction unit 105 extracts a feature quantity candidate used for generating an identifier as feature quantity.

Description

  The present invention relates to a feature amount extraction device for extracting a feature amount for identifying an object based on image data including spectral information of pixels, an object detection system including the feature amount extraction device, and the feature amount extraction device. The present invention relates to a computer program and a feature amount extraction method for realizing it.

  As one of the techniques used in various inspections such as food inspection and biological inspection, in recent years, an analysis technique using near-infrared spectrum information has attracted attention. Near-infrared light has a wavelength close to the visible light range and enables non-destructive and non-invasive compositional analysis of analytes, so it is possible to use near-infrared light in various fields such as food, medicine, and industry. Research on an object detection system using light has been made (for example, see Non-Patent Document 1).

  On the other hand, pattern recognition techniques such as multivariate analysis or data mining are often used as techniques for analyzing near-infrared spectrum information. These methods generate an identification standard or a calibration curve for an object through learning of a large amount of sample data.

  One of the methods for expressing near-infrared spectrum information is a near-infrared multispectral image (hereinafter also referred to as “NIRMS image”). An NIRMS image is an image that includes near-infrared spectrum information (intensity or absorbance, etc.) of a plurality of wavelengths for each pixel. By using an NIRMS image, an analysis applying an image processing technique based on composition information of an object is performed. It can be carried out.

Kikuo Takemoto, Akira Seki, Masao Makiuchi, Hideo Takahashi, “Near-infrared sensing technology for biological and environmental measurements”, Science Forum, 1999

  In order to realize a high-performance object detection system, it is important to accurately extract a feature amount that shows an essential difference between an object (object) and a non-object (non-object). When the feature amount to be used is insufficient, information necessary for identification is insufficient, and a desired identification accuracy cannot be obtained. On the other hand, when an unnecessarily large number of feature quantities are used, not only the calculation amount of the identification process is increased, but also the accuracy of the classifier may be reduced due to overfitting to the learning sample. And, in order to accurately extract characteristic quantities that are significant for the identification of the object without any excess or deficiency, specialized knowledge on the object or a large amount of steady analysis work is required. This indicates that it is not easy to construct a highly accurate object detection system.

  The present invention has been made in view of such circumstances, and is capable of accurately extracting a feature amount that shows an essential difference between an object (object) and a non-object (non-object). It is an object of the present invention to provide a quantity extraction device, an object detection system including the feature quantity extraction device, a computer program for realizing the feature quantity extraction device, and a feature quantity extraction method.

  A feature amount extraction apparatus according to a first aspect of the present invention extracts a feature amount for identifying an object based on image data including spectrum information of a plurality of wavelengths for each pixel included in an image composed of a plurality of pixels. Determining means for determining a plurality of feature amount candidates related to the displacement of the spectrum intensity at any of a plurality of wavelengths of each sample pixel from a plurality of sample pixels extracted from an image obtained by imaging the object and the non-object; Generating means for generating a discriminator for discriminating an object and a non-object using boosting for each of some of the feature quantity candidates determined by the determining means, and generating the discriminator And extracting means for extracting the feature quantity candidates used for the process as feature quantities.

  The feature amount extraction apparatus according to a second aspect is characterized in that, in the first aspect, the determining means is configured to determine a feature amount candidate related to a change in spectral intensity in a near-infrared wavelength range. And

  According to a third aspect of the present invention, in the first or second aspect of the invention, the feature amount extraction apparatus further includes a provision unit that imparts an arbitrary numerical value to each sample pixel, and the generation unit is correct when the target unit or the non-target unit. The discriminator is generated so that the sum of the numerical values of the identified sample pixels is maximized.

  In the feature amount extraction apparatus according to a fourth aspect of the present invention, in the third aspect of the invention, the assigning unit mistakenly identifies the target or candidate non-target by the classifier generated by the generation unit for one feature amount candidate. The numerical value of the sampled pixel is increased, and the generating means is configured to generate a discriminator using the sample pixel whose numerical value is increased when generating a discriminator for the next feature quantity candidate. It is characterized by being.

  An object detection system according to a fifth aspect of the present invention is an object having a feature quantity extraction device according to any one of the aforementioned inventions and a support vector machine using the feature quantity extracted by the feature quantity extraction device as a feature vector. And a detection device.

  An object detection system according to a sixth aspect of the present invention is an object having a feature classifier according to any one of the foregoing inventions, and a strong classifier that combines the classifier generated by the feature quantity extractor as a weak classifier. And an object detection device.

  A computer program according to a seventh aspect of the present invention is a computer for causing a computer to extract a feature amount for identifying an object based on image data including spectral information of a plurality of wavelengths for each pixel included in an image composed of a plurality of pixels. In the program, the computer determines a plurality of feature quantity candidates related to the displacement of the spectrum intensity at a plurality of arbitrary wavelengths of each sample pixel from a plurality of sample pixels extracted from images obtained by imaging the object and the non-object. Generating a discriminator for identifying an object and a non-object using boosting for each of some of the feature quantity candidates in the determined feature quantity candidates; and generating the discriminator And a step of extracting the feature quantity candidates used in the above as feature quantities.

  A feature amount extraction method according to an eighth aspect of the present invention is a feature amount extraction apparatus that extracts a feature amount for identifying an object based on image data including spectrum information of a plurality of wavelengths for each pixel included in an image composed of a plurality of pixels. In the feature amount extraction method according to the above, a plurality of feature amount candidates related to the displacement of the spectrum intensity at any of a plurality of wavelengths of each sample pixel from a plurality of sample pixels extracted from images obtained by capturing an object and a non-object A step of determining, a step of generating a discriminator for discriminating an object and a non-object using boosting for each part of the feature amount candidates in the determined feature amount candidates, and And a step of extracting the feature amount candidates used for generation as feature amounts.

  In the first invention, the seventh invention, and the eighth invention, the determining means is configured to determine each sample pixel from a plurality of sample pixels extracted from an image obtained by imaging the object (object) and the non-object (non-object). A plurality of feature amount candidates related to the displacement of the spectral intensity at a plurality of arbitrary wavelengths are determined. The image includes spectral information indicating spectral intensities of a plurality of wavelengths for each pixel. Absorbance may be used instead of intensity. For example, if the spectral intensities at two different wavelengths λ1 and λ2 of an arbitrary sample pixel are P (λ1) and P (λ2), the feature amount candidate is x1 = P (Λ1) −P (λ2). For example, when different wavelengths are divided into 256 such as λ1, λ2,..., Λ256, there are 32640 combinations of selecting two from 256, and therefore feature quantity candidates are for each sample pixel. There are 32640 pieces. The feature quantity candidates to be determined may include all existing feature quantity candidates, or some feature quantity candidates may be determined according to the object.

  The generating means generates a discriminator for discriminating the object and the non-object using boosting for each of some of the feature quantity candidates determined, and the extracting means The feature quantity candidates used for generation are extracted as feature quantities. Boosting is one method for generating a discriminator that distinguishes between an object and a non-object through, for example, sample learning of two types of objects (object and non-object). One feature amount candidate is selected from the determined feature amount candidates, and a discriminator (a boundary for distinguishing the sample pixels) for identifying a sample pixel that is previously known to be an object or a non-object A classifier (a boundary for distinguishing sample pixels) that can be configured and identified most correctly is generated. A discriminator is similarly generated for other feature quantity candidates. Then, the feature quantity candidate when the discriminator is generated is extracted as a feature quantity.

  By extracting the feature quantity candidates used when generating the classifier in the sample learning process as feature quantities, the feature quantities that can distinguish between objects and non-objects, that is, the essence between objects and non-objects It is possible to accurately extract a feature quantity in which a specific difference appears from a large number of determined feature quantity candidates.

  In the second invention, the determining means determines a feature amount candidate related to the displacement of the spectral intensity in the near infrared wavelength range. For example, sample pixels in which objects and non-objects are known in advance are collected from near-infrared multispectral images (NIRMS images) obtained by imaging objects and non-objects with a camera that has sensitivity to the near infrared. To do. Then, feature amount candidates are determined from the collected sample pixels. In the case of analysis of an analysis object using near infrared light spectrum information (for example, discrimination between an object and a non-object), the spectrum intensity distribution becomes a complicated shape depending on the analysis object, such as multivariate analysis or data mining. When the recognition technique is used, conventionally, it has been necessary to extract the feature amount by trial and error. Even when the wavelength distribution is complex, it is possible to extract optimal feature quantities for identifying objects and non-objects from a large number of feature quantity candidates without any shortage of information necessary for identification. In addition, it is possible to prevent a situation in which a larger number of features are used than necessary.

  In the third invention, the assigning means assigns an arbitrary numerical value (coefficient, weight, etc.) to each sample pixel, and the generating means sets the numerical value of the sample pixel correctly identified as an object or non-object. The discriminator is generated so that the sum is maximized. As a result, it is possible to generate a discriminator (boundary at a feature amount candidate) that can most accurately identify the sample pixel.

  In the fourth invention, the assigning means increases the numerical value of the sample pixel that is erroneously identified as an object or non-object by the classifier generated by the generating means for one feature quantity candidate. When generating the discriminator for the next feature quantity candidate, the generating unit generates the discriminator using the sample pixel having a larger numerical value. When generating a discriminator for a certain feature quantity candidate, the discriminator is generated for the next feature quantity candidate by increasing the numerical value of the erroneously identified sample pixel as a result of discrimination by the discriminator. For example, a sample pixel that could not be correctly recognized by the classifier generated at the t-th time will be intensively learned because the numerical value is large when the next t + 1-th classifier is generated, It is possible to reduce the number of combinations of features that are significant for identifying objects and non-objects and have different spectral intensity distributions as much as possible, and automatically extract the features required for target analysis without excess or deficiency. it can.

  The fifth aspect of the invention includes a feature quantity extraction device and an object detection device having a support vector machine using the feature quantity extracted by the feature quantity extraction device as a feature vector. This makes it possible to discriminate between objects and non-objects with high accuracy, and in the past, acquisition of specialized knowledge for analysis objects or analysis work, which required much labor, Simplification and analysis can be expanded.

  In a sixth aspect of the invention, the apparatus includes a feature quantity extraction device and an object detection device having a strong classifier in which a classifier generated by the feature quantity extraction device is combined as a weak classifier. This makes it possible to discriminate between objects and non-objects with high accuracy, and in the past, acquisition of specialized knowledge for analysis objects or analysis work, which required much labor, Simplification and analysis can be expanded.

  According to the present invention, it is possible to accurately extract a feature quantity that shows an essential difference between an object and a non-object from among a large number of determined feature quantity candidates.

It is a block diagram which shows an example of a structure of the target object detection system of this Embodiment. It is a schematic diagram which shows an example of a NIRMS image. It is a schematic diagram which shows an example of a feature-value candidate. It is a schematic diagram which shows the other example of a feature-value candidate. It is a schematic diagram which shows an example of the production | generation of a discriminator. It is a schematic diagram which shows an example of the coefficient of the sample pixel at the time of producing | generating a discriminator. It is a flowchart which shows the procedure of the feature-value extraction process by the feature-value extraction apparatus of this Embodiment. It is explanatory drawing which shows the detection result of this Embodiment and a comparative example. It is a schematic diagram which shows an example of the process result image of this Embodiment and a comparative example.

  Hereinafter, the present invention will be described with reference to the drawings illustrating embodiments thereof. FIG. 1 is a block diagram illustrating an example of a configuration of an object detection system 1 according to the present embodiment. The object detection system 1 according to the present embodiment includes a feature quantity extraction device 10, an object detection device 20, and the like. The object detection system 1 may be configured to include the NIRMS image (near-infrared multispectral image) acquisition unit 2 or may not include it.

  The NIRMS image acquisition unit 2 is a camera having sensitivity to the near infrared, for example. The NIRMS image has spectral information (for example, intensity, absorbance, etc.) for a plurality of wavelengths for each pixel constituting the image.

  FIG. 2 is a schematic diagram showing an example of a NIRMS image. As shown in FIG. 2, the NIRMS image is obtained by recording spectral information (for example, intensity, absorbance, etc.) discretely plotted within a specific wavelength range in each pixel.

  In this embodiment, an object (also referred to as an object) is an object to be detected (detected) in various inspections such as food inspection or biological inspection, and a non-object (also referred to as a non-object) is other than the object. And is distinguished (identified) from the object.

  In the learning process, the NIRMS image acquisition unit 2 captures an imaging target in which an object and a non-object are known in advance, generates an NIRMS image as a learning sample, and outputs the generated NIRMS image to the feature amount extraction apparatus 10 To do. The learning process can be performed offline.

  In the detection process, the NIRMS image acquisition unit 2 captures an object to be inspected to generate an NIRMS image, and outputs the generated NIRMS image to the object detection device 20. The detection process is performed online.

  Next, a configuration related to the learning process will be described. As shown in FIG. 1, the feature amount extraction apparatus 10 of the present embodiment includes a sample pixel collection unit 101, a feature amount candidate determination unit 102, a discriminator generation unit 103, a coefficient addition unit 104, a feature amount extraction unit 105, and the like. Prepare.

  The sample pixel collection unit 101 collects sample pixels used for learning processing (including feature amount extraction processing) from the NIRMS image. For example, a total of N sample pixels of the object (object) and non-object (non-object) sample pixels are collected.

  The feature amount candidate determination unit 102 has a function as a determination unit that determines a plurality of feature amount candidates related to the displacement of the spectrum intensity at arbitrary wavelengths of each sample pixel. Instead of the spectrum intensity, an absorbance spectrum may be used.

  FIG. 3 is a schematic diagram illustrating an example of feature quantity candidates. In FIG. 3, the horizontal axis indicates the wavelength, and the vertical axis indicates the spectral intensity. The graph of FIG. 3 shows the spectral intensity for each wavelength of any sample pixel (eg, object pixel or non-object pixel) in the NIRMS image. For example, assuming that the spectrum intensity at two different wavelengths λ1 and λ2 is P (λ1) and P (λ2), the feature quantity candidate is x1 = P (λ1) −P (Λ2).

  For example, when different wavelengths are divided into 256 such as λ1, λ2,..., Λ256, there are M = 32640 combinations M of which two are selected from 256, and each of the feature quantity candidates is There are 32640 sample pixels. Therefore, the feature quantity candidate can be represented by an M-dimensional vector X = (x1, x2,... XM) for each sample pixel. Note that the feature quantity candidates to be determined may include all existing feature quantity candidates (M feature quantities in the above example), or some feature quantity candidates (M pieces of feature quantities) depending on the object. Part) can also be determined.

  FIG. 4 is a schematic diagram showing another example of feature quantity candidates. In the example of FIG. 3, the difference between the spectral intensities at two different wavelengths is used as the feature quantity candidate as the feature quantity candidate related to the displacement of the spectrum intensity at any of a plurality of wavelengths. It is not limited to the example.

  In FIG. 4, the horizontal axis indicates the wavelength, and the vertical axis indicates the spectral intensity. As shown in FIG. 4, when attention is paid to three wavelengths λ1, λ2, and λ3 as different wavelengths, and spectrum intensities at the respective wavelengths are P1, P2, and P3, x = (P3 −P2) − (P2−P1) = P3−2 × P2 + P1 is used. The feature quantity candidates illustrated in FIG. 4 capture the degree of intensity change (the degree of curve on the curve) in the vicinity of the maximum point (or minimum point) of the intensity in the spectrum intensity distribution as a feature. In the following description, the feature quantity candidates illustrated in FIG. 3 are employed.

  The discriminator generating unit 103 has a function as a generating unit that generates a discriminator for discriminating an object and a non-object using boosting for each of some feature amount candidates in the determined feature amount candidates. Have. Boosting is one method for generating a discriminator that distinguishes between an object and a non-object through, for example, sample learning of two types of objects (object and non-object). For boosting, for example, AdaBoost, Real-AdaBoost, LogitBoost, Gentle-AdaBoost, etc. can be used.

  FIG. 5 is a schematic diagram illustrating an example of generation of a classifier. In the above example, the determined feature amount candidates can be represented by x1, x2,..., XM (M = 32640), and can be represented as an M-dimensional vector. FIG. 5 shows a two-dimensional space represented by two feature quantity candidates xi and xj for convenience, instead of the M-dimensional space. N sample pixels exist in this two-dimensional space (actually an M-dimensional space). Of the sample pixels, a sample pixel indicated by a symbol O is a pixel of an object (object), and a sample pixel indicated by a symbol x is a pixel of a non-object (non-object).

  The discriminator generation unit 103 selects any one feature amount candidate (for example, xi) from the determined M feature amount candidates, and determines whether it is an object or a non-object in advance. The discriminator (border for discriminating sample pixels) is configured to identify the sample pixel that has been identified, and the discriminator (border for discriminating the sample pixel) ht that can be identified most correctly among them is generated. To do.

  That is, as shown in FIG. 5, each sample in the M-dimensional space (represented in the two-dimensional space for simplification in the example of FIG. 5) is mapped to the feature quantity candidate xi (corresponding to one axis in the M-dimensional space). Then, a boundary on the axis xi when the number of sample pixels that can be correctly identified on the feature quantity candidate xi (axis xi) is maximized is generated as a discriminator ht.

  The feature quantity extraction unit 105 has a function as an extraction unit that extracts feature quantity candidates used for generating the classifier as feature quantities. In the example of FIG. 5, the feature amount extraction unit 105 extracts the feature amount candidate xi used when generating the discriminator Ht as a feature amount.

  The discriminator generation unit 103 similarly generates discriminators for other feature amount candidates, and the feature amount extraction unit 105 extracts the feature amount candidates used when the discriminator is generated as feature amounts. Since the number T of discriminators to be generated is, for example, about 10 to 100, the number of feature quantities to be extracted can be extremely reduced as compared with the determined number M of feature quantity candidates.

  By extracting the feature quantity candidates used when generating the classifier in the sample learning process as feature quantities, the feature quantities that can distinguish between objects and non-objects, that is, the essence between objects and non-objects It is possible to accurately extract a feature quantity in which a specific difference appears from a large number of determined feature quantity candidates.

  In the present embodiment, a NIRMS image is acquired, and sample pixels whose objects and non-objects are known in advance are collected. Then, from the collected sample pixels, a feature amount candidate related to the spectral intensity displacement is determined in the near-infrared wavelength range. In the case of analysis of an analysis object using near infrared light spectrum information (for example, discrimination between an object and a non-object), the spectrum intensity distribution becomes a complicated shape depending on the analysis object, such as multivariate analysis or data mining. When the recognition technique is used, conventionally, it has been necessary to extract the feature amount by trial and error. In this embodiment, even when the wavelength distribution is complicated, it is possible to extract a feature quantity optimal for discrimination between an object and a non-object from a large number of feature quantity candidates, and information necessary for identification can be obtained. In addition to preventing the shortage, it is possible to prevent a situation in which more feature quantities are used than necessary.

  When a classifier (boundary) is generated for one feature amount candidate, a coefficient can be given to each sample pixel. In the present embodiment, the coefficient means a numerical value, a weighting number, or the like. The coefficient assigning unit 104 assigns an arbitrary coefficient to each sample pixel. Then, the discriminator generation unit 103 generates the discriminator so that the sum of the coefficients of the sample pixels correctly identified as an object or a non-object is maximized. As a result, it is possible to generate a discriminator (boundary at a feature amount candidate) that can most accurately identify the sample pixel.

  FIG. 6 is a schematic diagram showing an example of coefficients of sample pixels when generating a discriminator. The coefficient assigning unit 104 assigns coefficients wi (i = 1, 2,..., N) to N sample pixels (learning samples). The coefficient wi can be set to, for example, 1 / N as an initial value. Assume that a discriminator ht is generated for one feature quantity candidate xi. The coefficient is reduced for sample pixels that can be correctly identified by the classifier ht. Further, the coefficient is increased for the sample pixel erroneously identified by the classifier ht.

  In the example of FIG. 6, the discriminator (boundary) ht can be correctly identified as an object (object) on the right side of the boundary and correctly identified as a non-object (non-object) on the left side of the boundary. Suppose you can. In this case, since the sample pixel O of the object on the right side of the boundary can be correctly identified, the coefficient is decreased, and the sample pixel X of the non-object is erroneously identified, so that the coefficient is increased. In addition, since the sample pixel O of the object on the left side of the boundary is erroneously identified, the coefficient is increased, and since the sample pixel X of the non-object is correctly identified, the coefficient is decreased.

  When generating a discriminator for the next feature quantity candidate, the discriminator generation unit 103 generates a discriminator by using sample pixels having a larger (or smaller) coefficient. When generating a discriminator for a certain feature quantity candidate, the coefficient of the sample pixel mistakenly identified as a result of discrimination by the discriminator is increased, and a discriminator is generated for the next feature quantity candidate. For example, sample pixels that could not be correctly recognized by the t-th discriminator are learned mainly because of the large coefficient when the next t + 1 th discriminator is generated, It is possible to reduce the number of combinations of features that are significant for identifying objects and non-objects and have different spectral intensity distributions as much as possible, and automatically extract the features required for target analysis without excess or deficiency. it can.

  FIG. 7 is a flowchart showing a procedure of feature amount extraction processing by the feature amount extraction apparatus 10 of the present embodiment. The feature quantity extraction device 10 (hereinafter referred to as “device 10”) acquires sample pixels (S11), and sets (applies) coefficients to the sample pixels (S12). The coefficient to be set is an initial value (for example, 1 / N where N is the number of sample pixels).

  The apparatus 10 sets the number of repetitions T (S13). Note that the number of iterations T is the number of feature values finally extracted. For example, the apparatus 10 determines M feature amount candidates (S14), and selects one feature amount candidate from the determined feature amount candidates (S15).

  The apparatus 10 constructs a discriminator through learning of sample pixels with respect to the selected feature amount candidate, and generates a discriminator (a boundary is also a feature amount candidate) having the maximum sum of the coefficients of the sample pixels correctly identified therein. (S16).

  The apparatus 10 reduces the coefficient of the sample pixel correctly identified by the generated classifier (S17), and increases the coefficient of the sample pixel erroneously identified by the generated classifier (S18). The apparatus 10 determines whether or not the number of iterations T has been reached (S19). If the number of iterations T has not been reached (NO in S19), the next feature amount candidate is selected (S20), and the processing after step S16 is performed. repeat.

  When the number of iterations T is reached (YES in S19), the apparatus 10 extracts the selected feature quantity candidate as a feature quantity (S21), and ends the process. By the feature amount extraction processing of the present embodiment, T (for example, about 10 to 100) feature amounts can be automatically extracted from a large amount (for example, M = 32640) of feature amount candidates. The extracted T feature quantities represent an essential difference between an object (object) and a non-object (non-object).

  The feature amount extraction apparatus 10 of the present embodiment can also be realized using a general-purpose computer including a CPU, a RAM, and the like. That is, a computer program that defines the processing procedure as shown in FIG. 7 is recorded on a computer program recording medium such as a CD, DVD, or USB memory, and the computer program is loaded into a RAM provided in the computer. By executing the above with the CPU, the feature quantity extraction device 10 can be realized on the computer.

  Next, a configuration related to the detection process will be described. As shown in FIG. 1, the object detection device 20 of the present embodiment includes an identification processing unit 201 and the like. The identification processing unit 201 includes a support vector machine that uses the feature amount extracted by the feature amount extraction apparatus 10 as a feature vector. For example, the discriminator F (Z) is configured by a T-dimensional vector space represented by a feature vector Z = (xt1, xt2,..., XtT) having extracted T feature amounts, and F (Z)> 0. Then, Z is an object (object), and if F (Z) ≦ 0, it can be determined that Z is a non-object (non-object). Note that F (Z) can be expressed as, for example, W · Z + b, where W is a vector representing a boundary (identifier) in the vector space, and W · Z is an inner product of the vectors. Further, b is a constant that can be set as appropriate.

  This makes it possible to discriminate between objects and non-objects with high accuracy, and in the past, acquisition of specialized knowledge for analysis objects or analysis work, which required much labor, Simplification and analysis can be expanded. As a configuration procedure of the discriminator, a pattern recognition technique such as discriminant analysis (LDA) can be used instead of the support vector machine (SVM).

  In addition, the identification processing unit 201 may include a strong classifier in which the classifier generated by the feature quantity extraction device 10 is combined as a weak classifier, instead of the SVM or the like. For example, if the classifiers generated by the feature quantity extraction device 10 are h1, h2,..., HT, the strong classifier H can be H = h1 + h2 +. Note that the discriminator h may be multiplied by a desired coefficient.

  This makes it possible to discriminate between objects and non-objects with high accuracy, and in the past, acquisition of specialized knowledge for analysis objects or analysis work, which required much labor, Simplification and analysis can be expanded.

  Next, the detection result which detected the hair which mixed the target object detection system of this Embodiment in the food processing line as a foreign material (target object, object) is demonstrated. In the following description, feature amounts are extracted using boosting according to the present embodiment, and detection method A using the extracted feature amounts is selected by trial and error (based on human knowledge) as a comparative example. The detection method B using the feature amount will be described.

  As a basic setting, 187 NIRMS images obtained by capturing food processing lines with a near-infrared camera are prepared. Among them, there are 46 images showing the hair as an object, and the remaining 141 images do not show the hair. The NIRMS image stores 256 near-infrared spectrum intensities obtained by discretely plotting a wavelength range of 1.0 μm to 2.5 μm at intervals of about 6 nm for each pixel.

  Let the object be “pixels with hair reflected” and the non-object “pixels with no hair reflected”. As sample pixels, 8706 objects were collected from 32 NIRMS images, and 32690 non-objects were collected. That is, the total number of sample pixels is 41396. Further, the feature quantity candidates are considered as a difference value of the spectrum intensity between two different wavelengths, and there are 32640 possible feature quantity candidates (that is, feature quantity candidates to be determined).

  In the detection method A of the present embodiment, 44 feature amounts are extracted from 32640 feature amount candidates using boosting. The sample pixels used when extracting the feature amount are 8706 objects and 32690 non-objects. In the identification process, first, all pixels of the NIRMS image are identified by linear SVM, and then identification is performed by nonlinear SVM only for pixels determined to be objects (hair pixels) by linear SVM. .

  On the other hand, the detection method B of the comparative example (conventional) extracted 44 feature amounts selected based on human knowledge from 32640 feature amount candidates. As in the detection method A, the learning sample pixels are 8706 objects and 32690 non-objects. Similarly to the detection method A, the identification processing is performed by first identifying all the pixels of the NIRMS image by the linear SVM, and then only for the pixels determined as objects (hair pixels) by the linear SVM. Identification was performed with non-linear SVM.

  In 34 images that are not used as sample pixels (learning samples), 304 hairs are present, and the detection rate is determined depending on how many of the 304 hairs have been detected. However, when the condition that there is a pixel identified as an object by connecting three or more pixels at the position where the hair exists is determined to be that the hair has been detected. Moreover, the false detection rate was calculated | required by how many pixels were identified as hair with respect to all the pixels of 141 NIRMS images in which the hair was not reflected. The processing speed is an average value of the time required to identify one pixel.

  FIG. 8 is an explanatory diagram showing detection results of the present embodiment and the comparative example. As shown in FIG. 8, it can be seen that the present embodiment (detection method A) is superior to the comparative example (detection method B) in all aspects of detection rate, false detection rate, and processing speed.

  FIG. 9 is a schematic diagram illustrating an example of a processing result image between the present embodiment and the comparative example. In FIG. 9, thin line segments represent hair, and dots represent erroneously detected noise. As shown in FIG. 9, it can be seen that the hair shape (trajectory) of the present embodiment (detection method A) is clearer and clearer than that of the comparative example (detection method B). It can also be seen that erroneously detected noise is reduced in the present embodiment (detection method A) compared to the comparative example (detection method B).

  In the past, in order to accurately select the necessary number of features that are significant for target identification, specialized knowledge about the object was required, and a lot of steady analysis work was required. According to the present embodiment, it is possible to automatically and accurately extract a feature quantity that is significant for distinguishing between an object and a non-object, and to easily configure an object detection system. That is, it is possible to improve the performance (high accuracy and high speed) of the identification processing by automatically extracting feature amounts. In addition, it is possible to acquire specialized knowledge necessary for realizing the object detection system, or to reduce the load of analysis work. Further, the configuration and construction of the object detection system can be easily performed.

  In the above-described embodiment, the configuration uses near-infrared spectrum information. However, the present invention is not limited to near-infrared light, and visible spectrum information can also be used.

  The disclosed embodiments are to be considered in all respects as illustrative and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

DESCRIPTION OF SYMBOLS 2 NIRMS image acquisition part 10 Feature-value extraction apparatus 101 Sample pixel collection part 102 Feature-value candidate determination part 103 Classifier production | generation part 104 Coefficient provision part 105 Feature-value extraction part 20 Object detection apparatus 201 Identification processing part

Claims (8)

In a feature amount extraction apparatus that extracts a feature amount for identifying an object based on image data including spectrum information of a plurality of wavelengths for each pixel included in an image composed of a plurality of pixels,
Determination means for determining a plurality of feature quantity candidates related to the displacement of the spectrum intensity at any of a plurality of wavelengths of each sample pixel from a plurality of sample pixels extracted from an image obtained by imaging an object and a non-object;
Generating means for generating a discriminator for discriminating an object and a non-object by using boosting for each of some of the feature quantity candidates determined by the determination means;
A feature quantity extraction apparatus comprising: extraction means for extracting feature quantity candidates used for generating the discriminator as feature quantities. The determining means includes
The feature amount extraction apparatus according to claim 1, wherein the feature amount candidate is determined so as to determine a feature amount candidate related to a change in spectral intensity in a near-infrared wavelength range. Provided with an assigning means for assigning an arbitrary numerical value to each sample pixel,
The generating means includes
The classifier is generated so that the sum of the numerical values of the sample pixels correctly identified as the object or the non-object is maximized. Feature extraction device. The giving means is
The number of sample pixels mistakenly identified as an object or non-object by the classifier generated by the generation unit for one feature amount candidate is increased,
The generating means includes
4. The feature quantity extraction apparatus according to claim 3, wherein when generating a classifier for the next feature quantity candidate, the classifier is generated by using a sample pixel having a larger numerical value. .   5. A feature amount extraction apparatus according to claim 1, and an object detection apparatus having a support vector machine using the feature amount extracted by the feature amount extraction apparatus as a feature vector. An object detection system characterized by.   5. A feature amount extraction device according to claim 1, and an object detection device having a strong classifier in which a classifier generated by the feature amount extraction device is combined as a weak classifier. An object detection system characterized by that. In a computer program for causing a computer to extract a feature amount for identifying an object based on image data including spectral information of a plurality of wavelengths for each pixel included in an image composed of a plurality of pixels,
On the computer,
Determining a plurality of feature quantity candidates related to the displacement of the spectral intensity at a plurality of arbitrary wavelengths of each sample pixel from a plurality of sample pixels extracted from an image obtained by imaging an object and a non-object; and
Generating a discriminator for identifying an object and a non-object using boosting for each of some of the feature quantity candidates in the determined feature quantity candidates;
And a step of extracting, as a feature amount, a feature amount candidate used for generating the discriminator. In a feature amount extraction method by a feature amount extraction apparatus that extracts a feature amount for identifying an object based on image data including spectrum information of a plurality of wavelengths for each pixel included in an image composed of a plurality of pixels,
Determining a plurality of feature quantity candidates related to the displacement of the spectral intensity at a plurality of arbitrary wavelengths of each sample pixel from a plurality of sample pixels extracted from an image obtained by imaging an object and a non-object; and
Generating a discriminator for identifying an object and a non-object using boosting for each of some of the feature quantity candidates in the determined feature quantity candidates;
Extracting the feature quantity candidates used for generating the discriminator as a feature quantity.

Images