Image processing system

Abstract

The present invention provides an image processing system capable of improving robustness, setting operability, and detection accuracy. The image processing system detects an image of an object from an image obtained by photographing the object. The image processing system comprises: a first detection device that detects the image of the object from the image based on a model pattern that represents the characteristics of the image of the object; a learning device that uses the image used in detection by the first detection device as input data and the detection results of the first detection device as teacher data to learn the learning model; and a second detection device that detects the image of the object from the image based on the learning model learned by the learning device.

Description

Technical Field

[0001] The present invention relates to an image processing system. Background Art

[0002] As an image processing system for detecting an image of an object from an image obtained by capturing the object, there is known a system that detects the image of the object based on a model pattern that expresses the characteristics of the image of the object (see, for example, Patent Document 1).

[0003] Also, as an image processing system for detecting an image of an object from an image obtained by photographing the object, there is known a system that uses machine learning to learn appropriate features corresponding to the object and detects the image of the object (see, for example, Patent Document 2).

[0004] Prior art literature

[0005] Patent Literature

[0006] Patent Document 1: Japanese Patent Application Laid-Open No. 2017-91079

[0007] Patent Document 2: Japanese Patent Application Publication No. 2018-200531 Summary of the Invention

[0008] Problems to be solved by the invention

[0009] In an image processing system using a model pattern representing the characteristics of an object's image, detection is performed focusing on one of the object's characteristics (e.g., outline). Therefore, this characteristic may not be visible due to changes in brightness, making it impossible to detect the object.

[0010] In contrast, an image processing system using machine learning can improve robustness compared to an image processing system using a model pattern that represents the characteristics of the image of an object. However, in the case of deep learning, for example, a large number of images with different positions and postures of the object need to be annotated (e.g., labeled) to generate teacher data. Such setup operations have become the main reason why users avoid using deep learning. In addition, labeling is generally performed by users through manual operations. Such setup operations have become the main reason for the decline in accuracy.

[0011] As described above, in the field of image processing for detecting an image of an object from an image obtained by capturing the object, improvements in robustness, setting operability, and detection accuracy are desired.

[0012] Solutions for solving problems

[0013] The image processing system disclosed herein is an image processing system for detecting an image of an object from an image obtained by photographing the object, and the image processing system comprises: a first detection device for detecting the image of the object from the image based on a model pattern that represents the characteristics of the image of the object; a learning device for learning a learning model using the image used in the detection performed by the first detection device as input data and the detection result of the first detection device as teacher data; and a second detection device for detecting the image of the object from the image based on the learning model learned by the learning device.

[0014] Effects of the Invention

[0015] According to the present disclosure, in the field of image processing for detecting an image of an object from an image obtained by capturing the object, it is possible to improve robustness, setting operability, and detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a diagram showing an example of an image processing system according to this embodiment.

[0017] Figure 2 This is a diagram showing another example of the image processing system according to this embodiment.

[0018] Figure 3 This is a diagram showing the configuration of an image processing device (first detection device) and a learning device (second detection device) in the image processing system according to the present embodiment.

[0019] Figure 4 It is a flowchart showing the process of making a model pattern.

[0020] Figure 5 FIG. 1 is a diagram showing a state where a model pattern designation area is designated in an image.

[0021] Figure 6 A diagram showing a model pattern including a plurality of feature points.

[0022] Figure 7 This is a flowchart showing the process of learning a learning model.

[0023] Figure 8 Schematically shows an example of an image processing system 201 in which a plurality of visual sensors are connected, according to this embodiment.

[0024] Figure 9 Schematically shows an example of an image processing system 301 in which a plurality of image processing apparatuses are connected according to the present embodiment.

[0025] Description of Reference Numerals

[0026] 1, 201, 301: image processing system; 2: object; 4: workbench; 5: input image; 10: image processing device (first detection device); 11: visual sensor; 12: image processing unit; 13: storage unit; 20: robot; 21: robot hand; 25: robot control device; 30: learning device (second detection device); 50: model pattern; 51: model pattern coordinate system; 60: model pattern designated area; 70: image coordinate system; 100: unit controller; 110: network bus. DETAILED DESCRIPTION

[0027] Hereinafter, an example of an embodiment of the present invention will be described with reference to the accompanying drawings. In addition, in each of the drawings, the same or corresponding parts are denoted by the same reference numerals.

[0028] Figure 1 FIG. 1 is a diagram showing an example of an image processing system according to the present embodiment. Figure 2 This is a diagram showing another example of the image processing system according to the present embodiment. The image processing system 1 is a system having the following two detection functions: a model pattern 50 (see FIG. 1 ) representing the characteristics of the image of the object 2; Figure 6 ) from the input image 5 (refer to Figure 5 ) in the image of the object 2; and detecting the image of the object 2 from the input image 5 based on the learning model.

[0029] Figure 1 The image processing system 1 shown includes a visual sensor 11, an image processing device (first detection device) 10, and a learning device (second detection device) 30. The visual sensor 11 is fixed in position. An object 2 is placed on a workbench 4. The visual sensor 11 is an imaging device such as a camera for imaging the object 2. The visual sensor 11 is fixed by a support unit (not shown) in a position where it can image the object 2. Image information acquired by the visual sensor 11 is transmitted to the image processing device 10.

[0030] The image processing device (first detection device) 10 uses the image processing described later to obtain the input image 5 (see Figure 5 ) detects the image of the object 2. The learning device (second detection device) 30 learns the learning model based on the detection result of the image processing device (first detection device) 10, and extracts the image 5 (refer to the image 5) received by the visual sensor 11 based on the learning model. Figure 5 ) in which the image of the object 2 is detected.

[0031] on the other hand, Figure 2 The image processing system 1 shown in FIG. Figure 1The image processing system 1 shown further includes a robot 20 and a robot control device 25, and the position of the visual sensor 11 is movable. The robot 20 is an arm-type robot with a robot hand 21 mounted at its front end. The visual sensor 11 is fixed to the robot hand 21, which serves as a finger of the robot 20. The robot hand 21 is a movable part that is moved by the robot 20 and its own mechanisms. Therefore, the position of the visual sensor 11 also moves. Alternatively, the robot hand 21 of the robot 20 can grasp the object 2 and move the object 2 into the field of view of the fixed visual sensor 11.

[0032] The image processing device 10 for performing image processing is configured to be able to communicate with the robot control device 25 that controls the movement of the robot 20, so that information can be exchanged between the image processing device 10 and the robot 20. The image processing device (first detection device) 10 takes into account the movement and state of the robot 20 and the robot hand 21, and obtains the input image 5 (see FIG. 1 ) from the visual sensor 11 by image processing described later. Figure 5 ) in the image of the detection object 2. Figure 1 and Figure 2 The image processing system 1 shown includes an image processing device (first detection device) 10 and a learning device (second detection device) 30.

[0033] Figure 3 1 is a diagram showing the configuration of an image processing device (first detection device) 10 and a learning device (second detection device) 30 in an image processing system 1 according to the present embodiment.

[0034] First, the image processing device (first detection device) 10 includes an image processing unit 12 and a storage unit 13. The image processing unit 12 generates a model pattern by modeling the image of the object 2, for example, a model pattern that represents the characteristics of the image of the object 2. The storage unit 13 stores the model pattern. Next, an example of creating a model pattern 50 will be described.

[0035] Figure 4 1 is a flowchart showing the process of creating the model pattern 50 . Figure 5 This is a diagram showing a state where the model pattern designation area 60 is designated in an image. Figure 6 is a diagram showing a model pattern 50 including a plurality of feature points P_i.

[0036] like Figure 5As shown, the object 2 to be taught as the model pattern 50 is placed within the field of view of the visual sensor 11, an image of the object 2 is captured, and an input image 5 including the object 2 is acquired (S11). At this time, it is preferable that the positional relationship between the visual sensor 11 and the object 2 be the same as the positional relationship when the object 2 is detected during actual use.

[0037] In the captured image, the area where the object 2 is captured is designated as the area of ​​the model pattern 50 (S12). Hereinafter, the area designated in step S12 is referred to as the model pattern designated area 60. The model pattern designated area 60 of this embodiment is designated using a rectangle or a circle to surround the object 2.

[0038] Next, feature points are extracted (S13). Feature points are points that constitute the model pattern 50. A plurality of feature points P_i (i=1 to NP) are extracted from the model pattern designated area 60. Various methods can be used to extract feature points P_i. In this embodiment, edge points in the image with a large brightness gradient and that can be used to obtain the contour shape of the object are used as feature points P_i.

[0039] Physical quantities of edge points include their position, brightness gradient direction, and brightness gradient magnitude. When the brightness gradient direction of an edge point is defined as the posture of a feature point, this can be combined with the position to define the position and posture of the feature point. Physical quantities of edge points—namely, their position, posture (brightness gradient direction), and brightness gradient magnitude—are stored as physical quantities of the feature point.

[0040] A model pattern coordinate system 51 is defined, and based on the model pattern coordinate system 51 and the origin O, the posture vector v_Pi and position vector t_Pi of the feature point P_i are expressed. Regarding the origin O set in the model pattern coordinate system 51, for example, the center of gravity of all the feature points P_i that constitute the model pattern 50 is defined as the origin O. Furthermore, the origin O can be defined using any appropriate method, such as selecting an arbitrary point from the feature points P_i. Furthermore, the method using the model pattern coordinate system 51 is merely an example, and other methods can also be used to represent the position and posture of the feature point P_i. Furthermore, regarding the axial direction (posture) of the model pattern coordinate system 51, for example, two arbitrary points can be selected from the feature points P_i that constitute the model pattern 50, and the direction from one toward the other can be defined as the X-axis direction, and the direction orthogonal to the X-axis direction can be defined as the Y-axis direction. Alternatively, it can be defined so that, in an image in which the model pattern 50 is created, the image coordinate system is parallel to the model pattern coordinate system 51. In this way, the settings of the model pattern coordinate system 51 and the origin O can be appropriately modified depending on the situation. Note that the method of extracting edge points as feature points is a well-known technique, and further detailed description thereof will be omitted.

[0041] Next, the model pattern 50 is generated based on the physical quantities of the extracted feature points P_i (S14). The physical quantities of the extracted feature points P_i are stored in the storage unit 13 as the feature points P_i constituting the model pattern 50. In this embodiment, a model pattern coordinate system 51 is defined in the model pattern designation area 60, and the position and posture of the feature points P_i are converted from the image coordinate system 70 (refer to FIG. Figure 5 ) is stored as the value expressed by the model pattern coordinate system 51 (refer to Figure 6 ) represents the value.

[0042] return Figure 3 The image processing unit (first detection unit) 12 detects the image of the object 2 from the input image 5 based on the model pattern that represents the characteristics of the image of the object 2. First, the image processing unit 12 extracts feature points from the input image 5. Feature points can be extracted using the same method as the method used to extract feature points when creating the model pattern. In this embodiment, edge points are extracted from the input image to serve as feature points.

[0043] Next, the image processing unit (first detection unit) 12 matches the feature points extracted from the input image 5 with the feature points constituting the model pattern 50 to detect the object 2. There are many methods for detecting the object, and for example, well-known methods such as generalized Hough transform, RANSAC, and ICP algorithm can be used.

[0044] The storage unit 13 stores the detection result of the image of the object 2 by the image processing device (first detection device) 10 and the data of the input image 5 corresponding thereto.

[0045] Next, refer to Figure 3 The following describes the learning device (second detection device) 30. The learning device 30 performs machine learning using the input image used in detection by the image processing device (first detection device) 10 as input data and the detection results (e.g., position, posture, and size) of the image processing device (first detection device) 10 as training data. Using the learning model constructed through this machine learning, the learning device (second detection device) 30 detects the image of the object 2 from the input image 5 containing the object 2 from the visual sensor 11.

[0046] The learning device (second detection device) 30 includes a state observation unit 31 , a label acquisition unit 32 , a learning unit 33 , a storage unit 34 , and an output presentation unit (output utilization unit) 35 to construct such a learning model.

[0047] The state observation unit 31 obtains input data from the image processing device (first detection device) 10 and outputs the obtained input data to the learning unit 33. As described above, the input data is the data of the input image 5 including the object 2 used in the detection performed by the image processing device (first detection device) 10.

[0048] The label acquisition unit 32 acquires labels from the image processing device (first detection device) 10 and outputs the acquired labels to the learning unit 33. Here, the labels are the above-mentioned teacher data, that is, the detection results (such as position, posture, and size) of the image processing device (first detection device) 10.

[0049] The learning unit 33 performs supervised learning based on the input data and labels to construct a learning model. The learning unit 33 can use well-known methods such as YOLO (You Only Look Once) and SSD (Single Shot Multibox Detector).

[0050] For example, the learning unit 33 performs supervised learning using a neural network. In this case, the learning unit 33 performs forward propagation as follows: a set of input data and labels (teacher data) is provided to a neural network composed of a combination of perceptrons, and the weighting of each perceptron included in the neural network is changed so that the output of the neural network is the same as the label. For example, in this embodiment, forward propagation is performed so that the object detection results (e.g., position, posture, size) output by the neural network are the same as the object detection results (e.g., position, posture, size) of the label.

[0051] After performing forward propagation, the learning unit 33 adjusts the weights using backpropagation (also known as error backpropagation) to reduce the output error of each perceptron. More specifically, the learning unit 33 calculates the error between the neural network's output and the label and adjusts the weights to reduce the calculated error. In this way, the learning unit 33 learns the characteristics of the teacher data and inductively obtains a learning model for estimating results based on input.

[0052] The storage unit 34 stores the learning model constructed by the learning unit 33. When new teacher data is acquired after the learning model is constructed, the learning model stored in the storage unit 34 is further subjected to supervised learning, thereby appropriately updating the once constructed learning model.

[0053] Furthermore, the learning model stored in the storage unit 34 may be shared with other learning devices. Sharing the learning model among a plurality of learning devices allows for decentralized supervised learning by each learning device, thereby improving the efficiency of supervised learning.

[0054] By using the learning model constructed in this manner, the learning device (second detection device) 30 detects the image of the object 2 from the input image 5 including the object 2 captured by the visual sensor 11 .

[0055] The output presentation unit 35 outputs the detection results (e.g., position, posture, size) of the learning device (second detection device) 30, that is, the output of the learning unit 33. The output presentation unit 35 presents the output of the learning unit 33 to the user by displaying it on a screen, for example.

[0056] The image processing device (first detection device) 10 and the learning device (second detection device) 30 are configured with a processor such as a DSP (Digital Signal Processor) or an FPGA (Field-Programmable Gate Array). The various functions of the image processing device 10 and the learning device 30 are implemented, for example, by executing predetermined software (programs, applications) stored in a storage unit. The various functions of the image processing device 10 and the learning device 30 can be implemented through a combination of hardware and software or solely through hardware (circuitry).

[0057] The storage unit 13 in the image processing device 10 and the storage unit 34 in the learning device 30 are, for example, rewritable memories such as EEPROMs.

[0058] Next, a learning example in which a learning model is constructed using the image processing device 10 and the learning device 30 will be described. Figure 7 This is a flowchart showing the process of learning a learning model.

[0059] First, the image processing device 10 acquires the input image 5 including the object 2 captured by the visual sensor 11 ( S21 ).

[0060] Next, the image processing unit 12 of the image processing device 10 detects the image of the object 2 from the input image 5 based on the model pattern that represents the characteristics of the image of the object 2 (S22). First, feature points are extracted from the input image 5. Feature points can be extracted using the same method as the feature point extraction method when creating the model pattern. In this embodiment, edge points are extracted from the input image as feature points. Next, the feature points extracted from the input image 5 are matched with the feature points constituting the model pattern 50 to detect the object 2.

[0061] The input image used in the detection by the image processing device (first detection device) 10 and the detection results (e.g., position, posture, and size) of the image processing device (first detection device) 10 are stored in the storage unit 13 (S23). The data of the input image and the detection result combination can be recorded automatically or at a time specified by the user.

[0062] The process from step S21 to step S23 is performed multiple times. As a result, multiple sets of data of the input image and the detection result are stored in the storage unit 13.

[0063] Next, the learning device 30 uses the input image used in the detection performed by the image processing device (first detection device) 10 as input data, and uses the detection results (such as position, posture, size) of the image processing device (first detection device) 10 as teacher data to learn the learning model (S24).

[0064] The learned model is stored in the storage unit 34. Alternatively, the learned model stored in the storage unit 34 is updated (S25).

[0065] The processing of step S24 and step S25 is performed multiple times, thereby improving the accuracy of the learning model.

[0066] After learning, the learning device (second detection device) 30 learns a learning model that returns a detection result when an input image is provided. Once the learning model is learned in this manner, both detection (first detection) based on the model pattern by the image processing device (first detection device) 10 and detection (second detection) based on the learned model by the learning device (second detection device) 30 are performed.

[0067] At this time, the image processing system 1 can compare the detection results of the image processing device (first detection device) 10 with the detection results of the learning device (second detection device) 30 and select the detection result with the higher evaluation value (score). For example, the evaluation value of the detection results of the image processing device (first detection device) 10 using the model pattern can be the ratio of the number of matching model points. On the other hand, the evaluation value of the detection results of the learning device (second detection device) 30 using the learned model can be the confidence level output by the learning device.

[0068] Alternatively, the evaluation value of the detection result of the learning device (second detection device) 30 can be obtained by the same method as the evaluation value of the detection result of the image processing device (first detection device) 10. For example, the learning device (second detection device) 30 can use the same method as the detection result of the image processing device (first detection device) 10 to score the detection result output by the learning model. For example, it can be set to apply the ratio of the number of points of the matching model points to the detection result output by the learning model. Thereby, the detection results of the image processing device (first detection device) 10 and the learning device (second detection device) 30 can be compared on the same scale.

[0069] Alternatively, the image processing system 1 can also be switched to compare the statistical value of the detection result of the image processing device (first detection device) 10 during a specified period with the statistical value of the detection result of the learning device (second detection device) 30 during the specified period, and use the detection device with a higher evaluation value calculated based on the statistical value for detection. Regarding the switching, it can be performed automatically or presented to the user and performed at the timing permitted by the user.

[0070] As described above, according to the image processing system 1 of the present embodiment, the image processing device (first detection device) 10 detects the image of the object 2 from the input image 5 based on the model pattern 50 representing the characteristics of the image of the object 2, and the learning device 30 learns the learning model based on the detection result of the image processing device 10 and the input image, and the learning device (second detection device) 30 detects the image of the object 2 from the input image 5 based on the learning model. Thereby, even with learning, the teacher data can be automatically generated by the image processing device (first detection device) 10. Therefore, there is no need for the user to collect a large amount of teacher data, and the workability of setting the learning device (second detection device) 30 can be improved. In addition, if the user performs label annotation through manual work, it is difficult to accurately specify the position and posture, but according to the present embodiment, there is no need for the user to perform label annotation through manual work, and the detection accuracy of the learning device (second detection device) 30 can be improved. And, by using the learning device (second detection device) 30 that learns, the robustness can be improved.

[0071] In addition, according to the image processing system 1 of the present embodiment, it is possible to start online operation (Japanese: ライン稼働) by the image processing device (first detection device) 10 before learning the learning model. Moreover, if sufficient teacher data is collected during the online operation, it can be switched to the learning device (second detection device) 30 based on the learning model.

[0072] In addition, even if learning is performed using learning data detected by the image processing device (first detection device) 10, the features acquired by the learning device (second detection device) 30 to detect the learning data are different from the features used by the image processing device (first detection device) 10, and a learning model can be formed that uses better features for finding learning data.

[0073] As mentioned above, although embodiment of this invention was described, this invention is not limited to the said embodiment, Various changes and deformation|transformation are possible.

[0074] For example, in the above embodiment, an example of using edge points as feature points constituting the model pattern 50 is described, but the present invention is not limited to this configuration. Next, a case where a method different from the above embodiment is used as a method for generating the model pattern 50 will be described.

[0075] First, we'll explain how to extract feature points using methods other than edge points. Feature points can be detected using various methods other than edge points. For example, feature points using SIFT (Scale-Invariant Feature Transform) can be used. The SIFT feature point extraction method itself is well-known, so a detailed description will be omitted.

[0076] Alternatively, the model pattern 50 can be created by arranging geometric shapes such as line segments, rectangles, and circles to match the outline of the object 2 captured in the image. In this case, the model pattern 50 can be created by placing feature points at appropriate intervals on the geometric shapes that constitute the outline. Alternatively, an image composed of individual pixels can be used as the model pattern.

[0077] The model pattern 50 is not limited to being composed of feature points. For example, the model pattern may be formed under the condition that a certain number of pixels or more are present in an area with a specific brightness value or higher.

[0078] In the above embodiment, the image detected by the visual sensor (camera) 11 is used as the input image 5. However, an image acquired by other means may also be used. For example, CAD data can also be used as the input image 5. In the case of two-dimensional CAD data, the model pattern can be created using the same method as the method using geometric figures described above. Alternatively, in the case of three-dimensional CAD data, the shape of the object 2 represented by the CAD data is projected onto the image, and feature points are extracted from the projected image.

[0079] The model pattern 50 is produced using CAD data as follows.

[0080] (1) A local coordinate system is defined, the origin of which is located on the image (imaging plane) captured by the visual sensor (camera) 11 .

[0081] (2) By calibrating the visual sensor (camera) 11 in advance, it is possible to convert three-dimensional points expressed by the local coordinate system into two-dimensional points on the captured image.

[0082] (3) Virtually arrange the object 2 represented by CAD data in the local coordinate system. The arranged CAD data is represented by the local coordinate system. The relative relationship between the visual sensor (camera) 11 and the object 2 is set to be substantially the same as the relative relationship when the object is actually detected.

[0083] (4) Acquire a three-dimensional point group on the contour line at a predetermined interval. If necessary, specify the contour line used as the model pattern from the CAD data.

[0084] (5) The three-dimensional point group is projected onto the image captured by the visual sensor (camera) 11 to obtain a two-dimensional point group in the image coordinate system. If the direction of light and dark is specified in the CAD data, the direction of the brightness gradient can also be specified. Here, the direction of light and dark indicates which of the two areas bounded by the contour line is brighter.

[0085] (6) The two-dimensional point group on the image coordinate system obtained is converted so as to be expressed by the model coordinate system and stored in the storage unit 13 as feature points.

[0086] As described above, the input image 5 may be image information generated based on CAD data. Various methods can be used for the input image 5. For example, a range image or three-dimensional point group data can also be used as the input image 5.

[0087] In the above embodiment, an image processing system 1 consisting of an image processing device 10 connected to a separate visual sensor 11 and a learning device 30 is described as an example, but the present invention is not limited to this configuration. Next, an image processing system having a configuration different from that of the above embodiment will be described. In the following example, components identical to those of the above embodiment are denoted by the same reference numerals, and detailed descriptions thereof will be omitted.

[0088] Figure 8 Schematically shows an example of an image processing system 201 connected to a plurality of visual sensors 11 according to this embodiment. Figure 8In the embodiment, N visual sensors 11 serving as imaging devices (input image acquisition devices) are connected to a unit controller 100 via a network bus 110. The unit controller 100 has the same functions as the image processing device 10 and the learning device 30 described above, and acquires input images 5 of N objects 2 acquired by each of the N visual sensors 11.

[0089] Like this, in Figure 8 In the illustrated image processing system 201, the unit controller 100 detects images of N objects 2 from N input images 5 obtained by capturing images of the N objects 2, based on N model patterns that represent the characteristics of the images of the N objects 2. A learning device (second detection device) 30 then learns the learning model based on the N input images used in the detection by the image processing device (first detection device) 10 and the N detection results of the image processing device (first detection device) 10. In this example, the learning process can also be performed online and sequentially.

[0090] According to this, by performing learning using the learning data of various objects 2 , a general-purpose learner can be trained.

[0091] Figure 9 Schematically shows an example of an image processing system 301 in which a plurality of image processing apparatuses 10 are connected according to this embodiment. Figure 9 In the image processing system 301, m image processing devices 10, which are imaging devices (input image acquisition devices), are connected to the unit controller 100 via a network bus 110. Each of the image processing devices 10 is connected to one or more visual sensors 11. The image processing system 301 as a whole includes a total of n visual sensors 11.

[0092] Like this, in Figure 9 In the illustrated image processing system 201, the unit controller 100 of each of the plurality of image processing devices (first detection devices) 10 detects the image of the object 2 from an input image 5 obtained by capturing the image of the object 2, based on a model pattern representing the characteristics of the image of the object 2. A learning device (second detection device) 30 then learns the learning model based on the N input images 5 used in the detection by the plurality of image processing devices (first detection devices) 10 and the N detection results of the image processing devices (first detection devices) 10. In this example, the learning process can also be performed sequentially online.

[0093] According to this, by performing learning using the learning data of various objects 2 , a general-purpose learner can be trained.

Claims

1. An image processing system for detecting an image of an object from an image obtained by capturing the object, the image processing system comprising: a first detection device for detecting an image of the object from the image based on a model pattern representing a feature of the image of the object; a learning device that uses the image used in the detection by the first detection device as input data and the detection result of the first detection device as teacher data to learn a learning model; as well as The second detection device detects the image of the object from the image based on the learning model learned by the learning device.

2. The image processing system according to claim 1, wherein Of the detection result of the first detection device and the detection result of the second detection device, the detection result having a higher evaluation value is selected.

3. The image processing system according to claim 2, wherein: The evaluation value for the detection result of the second detection device is obtained by the same method as the evaluation value for the detection result of the first detection device.

4. The image processing system according to any one of claims 1 to 3, characterized in that The switch is to select a detection device having a higher evaluation value calculated from the statistical value of the detection result of the first detection device in a predetermined period and the statistical value of the detection result of the second detection device in the predetermined period.

5. The image processing system according to any one of claims 1 to 3, characterized in that The first detection device detects the images of the plurality of objects from a plurality of images obtained by photographing the plurality of objects based on a plurality of model patterns expressing characteristics of the images of the plurality of objects. The learning device learns a learning model using the plurality of images used in the detection by the first detection device as input data and the plurality of detection results of the first detection device as teacher data.

6. The image processing system according to any one of claims 1 to 3, characterized in that: A plurality of first detection devices are provided, wherein the plurality of first detection devices detect the image of the object from the image based on a model pattern representing a feature of the image of the object, The learning device learns a learning model using the plurality of images used in detection by the plurality of first detection devices as input data and the plurality of detection results of the plurality of first detection devices as teacher data.

Images