JP6616906B1 - Detection device and detection system for defective photographing data - Google Patents

Abstract

An object of the present invention is to determine whether or not shooting data to be determined is defective shooting data without being affected by the installation state of the camera. In the dirt detection device 2, the feature amount extraction unit 23 that extracts the feature amount of the shooting data, the shooting data with the defect shot with the dirt attached to the lens, and the entire shooting range are normally shot. A model generation unit 25 that generates an identification model for identifying the presence / absence of a defect in the photographic data by learning based on the characteristic amount of the photographic data in which the defect-free photographic data is determined, and the feature amount of the photographic data to be determined A determination unit 26 that determines whether there is a defect in the shooting data to be determined by collating with the identification model, and a notification unit 27 that notifies the monitoring person who cleans the camera 5 that it has been determined that there is a defect. Have.

Description

  The present invention relates to a detection device and a detection system for defective shooting data, and more particularly to detection of shooting data having a defect as shooting data.

  Many cameras are installed in buildings and other facilities for surveillance purposes. The monitor in the building monitoring system views the data captured by the camera. However, if the camera lens is dirty or cloudy, the entire predetermined imaging range cannot be monitored. For such monitoring purposes, we would like to resolve as soon as possible the situation in which defective shooting data continues to be shot as shooting data.

  Therefore, Patent Document 1 proposes a technique for determining the degree of contamination and the degree of contamination by calculating the degree of contamination from the amount of change with time of a subject within a fixed imaging range. Patent Document 2 proposes a technique for identifying dirt from a difference from a normal image by learning a neural network.

Japanese Patent No. 5402131 Japanese Patent Laid-Open No. 11-007536 JP 2003-174644 A JP 2014-13449 A JP 2007-318355 A JP 2009-141906 A Japanese Patent Laid-Open No. 11-234543 Japanese Patent No. 3757789 Japanese Unexamined Patent Publication No. 2006-1000087 JP 2007-318355 A Japanese Patent No. 4644283 JP 2001-119614 A

  However, Patent Document 1 is applicable only when the shooting range of the camera is fixed. Also in Patent Document 2, in order to obtain a difference from a normal image, the photographing range must be fixed. In addition, since normal images in each camera must be taken in advance, when the number of cameras is enormous, a load and a large storage capacity are required for preparing normal images.

  An object of the present invention is to determine whether or not imaging data to be determined is defective imaging data without being affected by an installation state of imaging means for generating imaging data.

According to the present invention, there is provided a device for detecting imaging data with defects, which is feature amount extraction means for extracting feature amounts of imaging data, and imaging data without defects or imaging data without defects as imaging data. Model generation means for generating an identification model for detecting defective shooting data based on the feature amount extracted by the feature amount extraction means for each determined shooting data, and the feature amount extraction means And determining means for determining the presence or absence of a defect in the photographing data by inputting the characteristic amount of the photographing data to be judged extracted by the imaging model, and the feature quantity extracting means captures data relating to chromatic aberration. This is extracted as a feature amount of data.

According to the present invention, there is provided an apparatus for detecting imaging data having a defect, an acquisition means for acquiring imaging data in which defect-free imaging data having no defect and imaging data having a defect other than that are determined, and the acquisition a preprocessing means excludes the imaging data by the feature extraction is estimated that difficult from the obtained with defect imaging data and defect free photography data by means defect there shooting data and the defect obtained by the obtaining means Feature quantity extraction means for extracting the feature quantity of the photographic data not excluded by the preprocessing means from the no photographic data, and photographic data without defects or photographic data with defects and defects. Shooting data with defects based on the feature amount extracted by the feature amount extraction means A model generation unit that generates an identification model for detection, and a determination that determines whether there is a defect in the photographic data by inputting the feature amount of the photographic data to be determined extracted by the feature amount extraction unit into the identification model And the model generation unit learns based on the feature amount extracted by the feature amount extraction unit from each of the plurality of defect-captured shooting data and the plurality of defect-free shooting data captured in a plurality of environments. To generate the identification model .

According to the present invention, there is provided a device for detecting imaging data with defects, which is feature amount extraction means for extracting feature amounts of imaging data, and imaging data without defects or imaging data without defects as imaging data. Model generation means for generating an identification model for detecting defective shooting data based on the feature amount extracted by the feature amount extraction means for each determined shooting data, and the feature amount extraction means Determining means for inputting the feature quantity of the imaging data to be determined extracted by the above into the identification model and determining the presence / absence of a defect in the imaging data, and the model generation means performs imaging under a plurality of environments By learning based on the feature quantity extracted by the feature quantity extraction means from each of the plurality of photographed data with defects and the plurality of photographed data without defects. Generating the identification model Te, is intended to be re-learning to the identification model based on the correct answer data of the shooting data and the captured data is determined the determining means false.

According to the present invention, there is provided a device for detecting imaging data with defects, which is feature amount extraction means for extracting feature amounts of imaging data, and imaging data without defects or imaging data without defects as imaging data. Model generation means for generating an identification model for detecting defective shooting data based on the feature amount extracted by the feature amount extraction means for each determined shooting data, and the feature amount extraction means Determining means for inputting the feature quantity of the shooting data to be determined extracted by the above into the identification model and determining the presence or absence of defects in the shooting data, wherein the determination means are arranged in a time series order. Is continuously input to the identification model as shooting data to be determined, a predetermined number of times after determining that one of the plurality of shooting data is defective. The shadow data are continuously determined that there is a defect in which it is determined that there is a defect in the first determination target of the shooting data.

  Further, the feature amount extraction means extracts the feature amount of the shooting data based on the color characteristics of the shooting data.
  The feature amount extraction unit divides the photographing data into a predetermined number and extracts a feature amount for each divided image.

The defective photographing data detection system according to the present invention is photographing data photographed by a plurality of surveillance cameras included in a facility surveillance system, and there is no defect-free photographing data as a photographing data and a defect that is not so. An imaging unit that acquires imaging data that has been determined, and imaging data that is estimated to be difficult to extract a feature amount from the imaging data with defects and the imaging data without defects acquired by the acquisition unit. preprocessing means for sending the exclusion to whether we feature extraction means, a feature amount of photographic data that are not excluded by the pre-processing means from the defect there shooting data and defect free imaging data acquired by the acquisition unit said feature extraction means for extracting, feature amounts and without defect imaging data of the extracted defect there photographing data by the feature quantity extracting means Based on each of the collected amounts, a model generation unit that generates an identification model for detecting imaging data with a defect, and a feature amount of the determination target imaging data extracted by the feature amount extraction unit is input to the identification model. And determining means for determining the presence or absence of defects in the photographing data, and notifying means for notifying detection of defective photographing data when the determining means determines that there is a defect.

  Further, the imaging data with a defect is imaging data in which an adhering matter to the lens of the monitoring camera is captured.

  According to the present invention, it is determined whether or not the shooting data to be determined is defective shooting data without being affected by the installation state of the shooting means for generating shooting data.

It is a whole lineblock diagram of a dirt detection system in this embodiment. It is a hardware block diagram of the computer which forms the stain | pollution | contamination detection apparatus in this Embodiment. It is a flowchart which shows the identification model production | generation process in this Embodiment. It is a flowchart which shows the determination process in this Embodiment.

  Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is an overall configuration diagram of a dirt detection system in the present embodiment. FIG. 1 shows a configuration in which a monitoring system 1, a dirt detection device 2, and a management terminal 3 are connected to a network 4 such as a LAN.

  The monitoring system 1 is a system for monitoring facilities. A plurality of cameras 5 are scattered in the facility. Although the monitoring system 1 is assumed to be installed in the same facility as the camera 5, it may be installed outside the facility. The dirt detection device 2 determines whether or not the shooting data is defective shooting data as shooting data by analyzing the shooting data generated by shooting with the camera 5. The management terminal 3 is a terminal device used by a monitor of the monitoring system 1 or the like. Since the contamination of the camera 5 is removed by the worker, the management terminal 3 may be used by the worker.

  Here, “defect photographing data as photographing data” will be described. The shooting data is shot and generated for use for some purpose. In the case of the present embodiment, the shooting data is generated by the camera 5 shooting a predetermined range for the purpose of monitoring the facility. Therefore, shooting data that cannot be used for the purpose can be referred to as defective shooting data as shooting data. “Shooting data with defects as shooting data” should not be interpreted as being equivalent to the fact that shooting data was not generated, and should be interpreted as shooting data that cannot be used for the purpose.

  In the present embodiment, imaging data having defects as imaging data is referred to as “imaging imaging data with defects”. Note that imaging data that is not defective imaging data, that is, imaging data that is not defective as imaging data is referred to as “defect-free imaging data”. The shooting data can be classified into either shooting data with a defect or shooting data without a defect.

  “Defective shooting data” can be referred to as shooting data in which the entire shooting range of the camera 5 is not shot normally, for example. For example, when the dirt of the lens of the camera 5 adheres, or when the lens of the camera 5 becomes cloudy or has water drops. Even if the lens is cloudy, the entire shooting range may be shot, but it cannot be said that shooting is normal unless it can be used for surveillance purposes. There are two types of dirt: the dirt that has adhered to the lens if it has not been removed, and the dirt that peels off for some reason. In addition, shooting data with defects is also available when a flying object, such as a plastic bag, is entangled with the camera 5 or an insect or the like stops on the lens and closes the lens. Is generated.

  Further, as a factor that the entire photographing range of the camera 5 is not photographed normally (a factor that causes photographing data with a defect), it is not limited to the case where dirt or the like is directly attached to the lens of the camera 5. For example, even when the camera 5 is mounted in a box and the transparent plate (glass) on the front surface (surface facing the lens) of the box becomes dirty or cloudy, defective shooting data is generated.

  In the present embodiment, unless otherwise specified, a case where dirt is attached to the lens will be described as a situation where shooting data with a defect is generated. In other words, it may be assumed that “dirt adhesion” includes everything that causes generation of imaging data with a defect such as cloudiness of a lens, adhesion of water droplets, and further entanglement of flying objects. The dirt detection apparatus 2 in the present embodiment is an apparatus that detects imaging data with a defect, and strictly speaking, is not an apparatus that detects only dirt attached to the lens of the camera 5.

  FIG. 2 is a hardware configuration diagram of a computer forming the dirt detection device 2 according to the present embodiment. In the present embodiment, the computer that forms the dirt detection device 2 can be realized by a general-purpose hardware configuration such as a personal computer (PC). That is, as shown in FIG. 2, the computer has a CPU 31, ROM 32, RAM 33, a hard disk drive (HDD) 34 as storage means, a mouse 35 and keyboard 36 provided as input means, and a display 37 provided as display means. Are connected to an internal bus 40 through an input / output controller 38 and a network interface (IF) 39 provided as communication means.

  The management terminal 3 may also be composed of a general-purpose computer such as a PC, like the dirt detection device 2. The monitoring system 1 is the same, but may be constructed by a plurality of computers depending on the scale of the facility.

  Returning to FIG. 1, the monitoring system 1 includes a data collection unit 11 and a photographing data storage unit 12. The monitoring system 1 may be an existing system. Components not used in the description of the present embodiment are omitted from the drawing. The data collection unit 11 collects shooting data from the camera 5. The shooting data storage unit 12 stores shooting data collected by the data collection unit 11. Since one camera 5 normally shoots continuously, the shooting data storage unit 12 requires a huge storage capacity. However, generally, a storage area that can be used is allocated to each camera 5, and when shooting data is recorded in the shooting data storage unit 12, old data is overwritten and saved.

  The dirt detection apparatus 2 includes a photographing data acquisition unit 21, a preprocessing unit 22, a feature amount extraction unit 23, a normalization unit 24, a model generation unit 25, a determination unit 26, a notification unit 27, and an identification model storage unit 28. Yes. Components not used in the description of the present embodiment are omitted from the drawing. The shooting data acquisition unit 21 acquires shooting data from the monitoring system 1. The feature amount extraction unit 23 extracts the feature amount of the shooting data acquired by the shooting data acquisition unit 21, while the preprocessing unit 22 performs the pre-processing for selecting the shooting data to be processed by the feature amount extraction unit 23. Process. Specifically, after excluding the imaging data that is difficult to extract the feature amount in the feature amount extraction unit 23 from the imaging data with the defect and the imaging data without the defect acquired by the imaging data acquisition unit 21. It is sent to the feature quantity extraction unit 23. The model generation unit 25 generates defect-captured photographic data based on the feature amount extracted by the feature amount extraction unit 23 for each photographic data that is determined to be defect-free photographic data or defect-captured photographic data. An identification model for detection is generated. The determination unit 26 inputs the feature amount of the photographic data to be determined extracted by the feature amount extraction unit 23 into the identification model, and determines the presence or absence of a defect in the photographic data. That is, it is determined whether or not the shooting data to be determined corresponds to shooting data with a defect. When the determination unit 26 determines that there is a defect, the notification unit 27 notifies the management terminal 3 of the detection of the imaging data with the defect. In the identification model storage unit 28, the arbitrary model generated by the model generation unit 25 is stored.

  Each component 21 to 27 in the dirt detection device 2 is realized by a cooperative operation of a computer that forms the dirt detection device 2 and a program that operates on the CPU 31 mounted on the computer. The identification model storage unit 28 is realized by the HDD 34 mounted on the dirt detection device 2. Alternatively, the RAM 33 or an external storage means may be used via a network.

  Further, the program used in this embodiment can be provided not only by communication means but also by storing it in a computer-readable recording medium such as a CD-ROM or USB memory. The program provided from the communication means or the recording medium is installed in the computer, and various processes are realized by the CPU of the computer sequentially executing the program.

  Next, the operation in this embodiment will be described.

  The dirt detection device 2 in the present embodiment detects the dirt on the lens of the camera 5 by determining whether or not the shooting data generated by shooting by the camera 5 is shooting data with a defect. The camera 5 generates, for example, 30 frame images per second according to the specifications. Continuously generated frame images are reproduced as moving images by being reproduced continuously. The shooting data in this embodiment corresponds to a frame image. The camera 5 always shoots a predetermined shooting range. Although the monitoring system 1 may change the shooting range of the camera 5, the dirt detection device 2 in the present embodiment can detect the dirt of the lens without being affected by the change of the shooting range of the camera 5. Further, it is not necessary for each camera 5 to take a normal image in advance.

  However, in the present embodiment, it is necessary to generate an identification model of imaging data with a defect in advance. Hereinafter, the generation process of the identification model in the present embodiment will be described with reference to the flowchart shown in FIG.

  In this embodiment, it is necessary to prepare imaging data that has been determined to be imaging data with a defect or imaging data without a defect. In order to facilitate comparison, the imaging data with defect and the imaging data without defect are ideally generated by imaging the same imaging range with the same camera 5. However, this does not mean that the image data must be generated under such conditions. It is only necessary to know whether the imaging data is imaging data with defects or imaging data without defects. Further, the imaging data with defect and the imaging data without defect need not be limited to data captured by the camera 5 in the monitoring system 1. However, in the present embodiment, for the sake of convenience, description will be made on the assumption that shooting data shot by the camera 5 in the monitoring system 1 is used.

  That is, the imaging data acquisition unit 21 in the dirt detection apparatus 2 acquires, as learning data, imaging data that has been determined as imaging data with defects or imaging data without defects from the monitoring system 1 (step 101). In order to improve the accuracy of the identification model generated by learning, it is preferable that the number of learning data is large. In addition, it is preferable to acquire shooting data collected under various environments. For example, the range of correspondence becomes wider as learning various shooting conditions such as indoor and outdoor situations, human and human conflicts, and various shooting times and places. On the other hand, as the correspondence range becomes wider, the judgment accuracy using the identification model tends to decrease, so in order to maintain and improve the judgment accuracy, the identification method becomes complicated and the collection of shooting data in a similar environment Is required. Shooting data acquired by the shooting data acquisition unit 21 in the identification model generation process may be generated for generation of an identification model, or shooting data with a defect or shooting without a defect among shooting data shot during monitoring. Shooting data that has been determined as data may be extracted.

  By the way, in the latter case, since the brightness does not reach a predetermined threshold due to the light-off even during nighttime or daytime, the photographing data photographed by the camera 5 during monitoring is used to extract the feature amount of the photographing data. It may not be suitable. Therefore, the pre-processing unit 22 excludes, as pre-processing of the process of extracting the feature amount of the shooting data, shooting data that is difficult to extract the feature amount from the shooting data acquired by the shooting data acquisition unit 21. (Step 102). As an object to be excluded, as described above, when the predetermined luminance is not reached, there are other cases where the variation of the pixel value is equal to or larger than a predetermined value. The preprocessing unit 22 examines the photographic data according to the conditions to be excluded.

  Subsequently, the feature amount extraction unit 23 extracts the feature amount of the shooting data sent from the preprocessing unit 22 (step 103). In the present embodiment, the feature amount of the shooting data is extracted based on the color characteristics of the shooting data. For example, when the shooting data is a color image, color characteristic data is provided for each pixel. Specifically, there are saturation, brightness, hue, and the like. Further, the entire photographing data includes color scheme, variation, structure, and the like. Of course, it is not necessary to limit to the information indicating the color characteristics. By combining these color characteristics, a feature amount that causes a difference between the image data with defect and the image data without defect is extracted. In the image data with a defect, it is estimated that the image of the dirt adhering to the lens (imaging area of the dirt portion) is black, that is, the brightness is lowered. Therefore, it is considered that the feature amount is greatly different from that in the case where no dirt is attached. In this way, a feature amount specialized for imaging data with a defect is extracted.

  In the present embodiment, the feature amount is extracted based on the color characteristics of the captured data, but the feature amount may be extracted by paying attention to chromatic aberration. “Chromatic aberration” refers to aberration that occurs due to dispersion of lens material when an image is formed with lenses, and appears as blurring or deviation of the color of the image. When dirt adheres directly to the lens of the camera 5, the distance between the object to be photographed and dirt is different. In other words, it is preferable to handle chromatic aberration because a characteristic that the refractive index is different appears. For the feature amount based on chromatic aberration, an edge portion where chromatic aberration is likely to occur is extracted from the defective photographic data, and an average of the feature amounts of the edge portion is calculated.

  Chromatic aberration includes axial chromatic aberration in which the imaging position shifts back and forth depending on the color, and lateral chromatic aberration that appears as a difference in image magnification in the case of obliquely incident light. Is available.

  In this embodiment, the average value of the feature values of the entire image or the divided image is calculated in order to extract the feature value specialized for dirt. However, other values such as entropy are calculated. May be.

  The feature amount extraction unit 23 may calculate an average of the feature amounts of the entire image (shooting data with a defect), but in order to improve the identification accuracy in the identification model, the feature amount is extracted, and the feature amount is determined for each divided image. You may calculate the average of. For example, the ratio between the feature amount of the edge portion and the feature amount in the divided image is extracted as the feature amount of the divided image. For example, when an image is divided into 24 parts, 24 feature amounts are extracted from shooting data with one defect, and the 24 feature amounts are handled as one set. If the image data with defects is not divided, the feature of the image of the dirty portion is stretched over the entire image, making it difficult to obtain the feature. On the other hand, when the image is divided, the feature of the image of the dirty portion remains in the divided image (narrow range), and the feature is likely to appear. As described above, in order to effectively extract the feature amount of the partial dirt, it is preferable to divide the image and process it. This makes it easy to obtain a correct answer by learning.

  Subsequently, the normalizing unit 24 normalizes the feature amount of the shooting data (defected shooting data and defect-free shooting data) extracted by the feature amount extraction unit 23 as described above (step 104). After that, the model generation unit 25 generates an identification model for detecting defective shooting data by learning based on the normalized shooting data (step 105). In the present embodiment, an identification model is generated by inputting normalized imaging data into an SVM (Support Vector Machine) which is one of pattern recognition models using supervised learning. SVM is a highly accurate pattern recognition model with a relatively low processing load. Of course, other pattern recognition models may be used. The identification model generated by the model generation unit 25 is stored in the identification model storage unit 28.

  Even after the identification model is once generated, the identification model generation process may be repeatedly executed to improve the identification accuracy in the identification model.

  After the identification model is generated as described above, it is determined whether the imaging data to be determined is imaging data with a defect or imaging data without a defect using the identification model. The photographing data determination process will be described with reference to the flowchart shown in FIG.

  The shooting data acquisition unit 21 in the dirt detection device 2 acquires shooting data to be determined from the monitoring system 1 (step 121). Identification information for identifying the camera 5 that generated the photographing data, for example, a camera ID is added to the photographing data to be determined. Subsequently, as in the case of the identification model generation process, the preprocessing unit 22 performs preprocessing for excluding the shooting data when it is estimated that the shooting data acquired by the shooting data acquisition unit 21 is difficult to extract the feature amount. (Step 122). In the case where one piece of image data is to be determined, if the preprocessing unit 22 excludes it, the process ends here, but here the description is continued assuming that it is not excluded.

  Subsequently, the feature amount extraction unit 23 extracts the feature amount of the shooting data sent from the preprocessing unit 22 (step 123). For the feature quantity extraction, the method described in the identification model generation process is used as it is. Then, the normalizing unit 24 normalizes the feature amount of the photographic data extracted by the feature amount extracting unit 23 (step 124). In addition, what is necessary is just to process similarly to an identification model production | generation process so far.

  Subsequently, the determination unit 26 determines whether or not the shooting data is defective shooting data by comparing the normalized shooting data with the identification model (step 125). That is, the determination unit 26 determines whether or not the shooting data is defective shooting data by inputting the shooting data to the identification model. Here, when it is determined that the shooting data is defective shooting data (Y in step 126), the notification unit 27 adds the camera ID of the camera 5 and notifies the management terminal 3 of the determination result (step 127). ).

  When it is determined that the shooting data is not defective shooting data (N in step 126), the notification unit 27 does not need to notify the management terminal 3 to that effect, but may notify it.

  A user (for example, a supervisor) of the management terminal 3 recognizes that the lens of the camera 5 is dirty by this notification. The monitor then identifies the camera 5 with reference to the identification information (camera ID) of the camera 5 included in the notification, and cleans so as to remove dirt attached to the lens of the identified camera 5.

  According to the present embodiment, as described above, it is determined whether or not the imaging data to be determined is imaging data with a defect, and if it is imaging data with a defect, the camera 5 is notified by notifying the monitor. It is possible to immediately remove the dirt on the lens.

  By the way, in the present embodiment, the case where the lens of the camera 5 is dirty has been described as an example. However, as described above, as a factor that hinders shooting by the camera 5 (factor that generates defective shooting data), It can be caused by flying objects. Further, a case where reflected light from the moving body is incident on the lens can be assumed. In this case, unlike a case where dirt is adhered to the lens, after a certain amount of time has passed, the flying object may disappear from the front of the lens and return to a normal state, that is, a state in which defect-free imaging data is generated.

  Therefore, in the present embodiment, in the determination process, a plurality of pieces of shooting data arranged in chronological order are acquired as shooting data to be determined, and preprocessing, feature amount extraction, and normalization are performed on each acquired shooting data. I do. And the determination part 26 inputs imaging | photography data into an identification model continuously in the order arranged in time series. Then, the determination unit 26 does not immediately determine that there is a defect even if any of the shooting data is determined to be defective, but first determines that a predetermined number of shooting data is continuously defective from the shooting data. It was determined that there was a defect in the shooting data to be determined. That is, even if it is determined that there is a defect in the shooting data, after the flying object is detached from the lens, the shooting data following the shooting data determined as having a defect is determined as having no defect. For example, if the determination unit 26 determines that there is a defect in one piece of shooting data and then determines that there is a defect for a predetermined time, for example, 5 seconds, it is determined that there is a defect and the notification unit 27 is notified. . For example, when the camera 5 generates a 30-frame image (photographing data) per second, for example, the determination unit 26 determines that 5 × 30 = 150 pieces of photographing data are continuously defective for five seconds. Only then is it determined that there is a defect in the shooting data to be determined.

  Note that the plurality of pieces of shooting data arranged in chronological order do not necessarily need to be shot data shot continuously. For example, every n (n is a natural number) pieces of continuous shooting data may be extracted. Even if every n frames are extracted, the extracted photographing data is arranged in time series.

  Moreover, when the learning with respect to an identification model is inadequate, the case where the determination part 26 makes a misjudgment may be assumed. In consideration of this case, the determination accuracy of the determination unit 26 may be verified using test data. For example, in the determination process, the shooting data acquired by the shooting data acquisition unit 21 is used as test data, and the determination unit 26 is determined to determine whether there is a defect in the shooting data as described above. Here, it is assumed that it is determined in advance whether the test data is imaging data with defects or imaging data without defects. That is, it can be seen whether the determination result in the determination unit 26 is an erroneous determination. Here, when the determination unit 26 makes an erroneous determination, the model generation unit 25 causes the identification model to relearn based on the imaging data erroneously determined by the determination unit 26 and the correct answer data of the imaging data. In this way, the identification model identification accuracy (determination accuracy by the determination unit 26) may be improved.

  In the present embodiment, the dirt detection device 2 is provided to detect a defect in the photographing data as described above. However, the entire function of the dirt detection device 2 is mounted on another device such as the monitoring system 1. Or may be distributed and installed in a plurality of devices.

  Further, in the present embodiment, the contamination is detected using the image data of the camera in the monitoring system, but the image data collected for other purposes may be the target of the contamination detection.

  In the present embodiment, the case where dirt is attached to the lens of the camera 5 has been described as a representative. However, it is also possible to use image data captured by an imaging unit having no lens as a determination target.

  In addition, the camera 5 in the present embodiment assumes shooting data shot in an environment using air as a medium. However, the shooting data shot in an environment of other media such as water is not limited to air. It is good also as a judgment object.

DESCRIPTION OF SYMBOLS 1 Monitoring system, 2 Dirt detection apparatus, 3 Management terminal, 4 Network, 5 Camera, 11 Data collection part, 12 Photographing data storage part, 21 Photographing data acquisition part, 22 Preprocessing part, 23 Feature-value extraction part, 24 Normalization Unit, 25 model generation unit, 26 determination unit, 27 notification unit, 28 identification model storage unit, 31 CPU, 32 ROM, 33 RAM, 34 hard disk drive (HDD), 35 mouse, 36 keyboard, 37 display, 38 I / O controller 40 Internal bus.

Claims (8)

Feature amount extraction means for extracting feature amounts of the shooting data;
For each shooting data that has been determined to be shooting data without defects or shooting data with defects as shooting data, there is a defect based on the feature amount extracted by the feature amount extraction means Model generation means for generating an identification model for detecting imaging data;
A determination unit that inputs the feature amount of the shooting data to be determined extracted by the feature amount extraction unit to the identification model and determines the presence or absence of a defect in the shooting data;
Have
The defect-capturing data detection apparatus, wherein the feature amount extraction unit extracts data relating to chromatic aberration as a feature amount of the photographing data. An acquisition means for acquiring imaging data in which defect-free imaging data without defects as imaging data and imaging data with defects not so are determined,
Pre-processing means for excluding imaging data that is estimated to be difficult to extract feature values from imaging data with defects and imaging data without defects acquired by the acquisition means;
Feature quantity extraction means for extracting the feature quantity of the imaging data not excluded by the preprocessing means from the imaging data with defects acquired by the acquisition means and the imaging data without defects;
For each shooting data that has been determined to be shooting data without defects or shooting data with defects as shooting data, there is a defect based on the feature amount extracted by the feature amount extraction means Model generation means for generating an identification model for detecting imaging data;
A determination unit that inputs the feature amount of the shooting data to be determined extracted by the feature amount extraction unit to the identification model and determines the presence or absence of a defect in the shooting data;
Have
The model generation unit learns the identification model by learning based on the feature amount extracted by the feature amount extraction unit from each of a plurality of defect-captured shooting data and a plurality of defect-free shooting data captured in a plurality of environments. An apparatus for detecting imaging data having a defect, characterized in that it is generated. Feature amount extraction means for extracting feature amounts of the shooting data;
For each shooting data that has been determined to be shooting data without defects or shooting data with defects as shooting data, there is a defect based on the feature amount extracted by the feature amount extraction means Model generation means for generating an identification model for detecting imaging data;
A determination unit that inputs the feature amount of the shooting data to be determined extracted by the feature amount extraction unit to the identification model and determines the presence or absence of a defect in the shooting data;
Have
The model generation unit learns the identification model by learning based on the feature amount extracted by the feature amount extraction unit from each of a plurality of defect-captured shooting data and a plurality of defect-free shooting data captured in a plurality of environments. An apparatus for detecting defective photographic data, wherein the identification model is relearned based on photographic data generated and erroneously determined by the determination means and correct answer data of the photographic data. Feature amount extraction means for extracting feature amounts of the shooting data;
For each shooting data that has been determined to be shooting data without defects or shooting data with defects as shooting data, there is a defect based on the feature amount extracted by the feature amount extraction means Model generation means for generating an identification model for detecting imaging data;
A determination unit that inputs the feature amount of the shooting data to be determined extracted by the feature amount extraction unit to the identification model and determines the presence or absence of a defect in the shooting data;
Have
In the case where a plurality of pieces of shooting data arranged in chronological order are continuously input to the identification model as determination target shooting data, the determination unit determines that one of the plurality of pieces of shooting data is defective and is predetermined. An apparatus for detecting imaging data having a defect, wherein the imaging data to be determined is determined to be defective only after a plurality of imaging data is determined to be defective continuously.   5. The apparatus according to claim 1, wherein the feature amount extraction unit extracts a feature amount of the shooting data based on a color characteristic of the shooting data.   5. The defective shooting data according to claim 1, wherein the feature amount extraction unit divides the shooting data into a predetermined number and extracts a feature amount for each of the divided images. Detection device. Capture data captured by a plurality of surveillance cameras in a facility monitoring system, in which there is no defect-free imaging data and no defective imaging data is acquired as imaging data. Acquisition means;
Preprocessing means for sending the defect there shooting data and excluded to whether we feature amount extracting means for capturing data by the feature extraction is estimated that difficult from a defect without imaging data acquired by the acquisition unit,
Said feature quantity extraction means for extracting a feature amount of photographic data that are not excluded by the pre-processing means from the obtained with defect imaging data and defect free imaging data by the acquisition unit,
A model generating unit that generates an identification model for detecting defective shooting data based on the feature amount of the defective shooting data extracted by the feature amount extraction unit and the feature amount of the shooting data without defect;
A determination unit that inputs the feature amount of the shooting data to be determined extracted by the feature amount extraction unit to the identification model and determines the presence or absence of a defect in the shooting data;
If the determination means determines that there is a defect, a notification means for notifying the detection of imaging data with a defect,
A dirt detection system for defective photographing data, comprising:   8. The detection system for defective imaging data according to claim 7, wherein the imaging data with defect is imaging data in which an object attached to a lens of the surveillance camera is imaged.

Images