WO2022153827A1

WO2022153827A1 - Article detection method, device, and electronic apparatus

Info

Publication number: WO2022153827A1
Application number: PCT/JP2021/047883
Authority: WO
Inventors: 剣俊蒋; 賢捷郭; 鵬関
Original assignee: オムロン株式会社
Priority date: 2021-01-12
Filing date: 2021-12-23
Publication date: 2022-07-21
Also published as: CN114764773A

Abstract

The present disclosure provides an article detection method, a device, and an electronic apparatus. The article detection comprises: using a deep learning positioning network model to position a detection target in an article image and acquire positioning data of the detection target; calculating a feature value of the detection target on the basis of the positioning data; and comparing the calculated feature value with a preset detection threshold value to determine from the comparison result whether or not the detection target has a defect. Thus, in a practical application, production speed requirements can be satisfied, and it is possible to quickly determine whether or not the detection target has a defect.

Description

Article detection methods, devices, and electronic devices

The present application relates to the technical field of article detection, and particularly to articles detection methods, devices, and electronic devices.

When producing or manufacturing an article, it is necessary to detect the article. By detecting the article, it is possible to determine the presence or absence of defects, and it is possible to improve the non-defective rate. With conventional vision techniques, it is difficult to detect an object under test in a complex environment (eg, when the article is in a packaging box or in a clear plastic bag), and it is also possible to inspect the appearance of the article. It is impossible. Currently, by means based on deep learning segmentation networks, it is possible to detect the appearance morphology of an article.

It should be noted here that the introduction of the background art described above is intended to make the technical solution of the present application easier to explain more clearly and completely, and is provided only for those skilled in the art to understand. .. It should not be found that the above technical solutions have been made known to those skilled in the art solely by the fact that those solutions are described in the background art section of the present application.

However, according to the present inventor, the article detection means based on the deep learning segmentation network requires a large amount of calculation and takes a considerable amount of time, so that the real-time property is not good, and in practical applications, the production speed requirement is satisfied. We found that it was difficult to do so, and it was not possible to quickly determine whether or not the detection target was defective.

In order to solve at least one of the above problems, an article detection method, an apparatus, and an electronic device are proposed in the embodiment of the present application.

According to one aspect of the embodiment of the present application, it is an article detection method.
To position the detection target in the article image by the deep learning positioning network model and acquire the positioning data of the detection target.
Calculate the feature value of the detection target based on the positioning data, and
Provided is an article detection method including comparing a calculated feature value with a preset detection threshold value and determining whether or not the detection target has a defect from the comparison result.

According to another aspect of the embodiment of the present application, it is an article detection device.
A positioning unit that positions the detection target in the article image and acquires the positioning data of the detection target by the deep learning positioning network model.
A calculation unit that calculates the feature value of the detection target based on the positioning data,
Provided is an article detection device including a comparison unit for comparing a calculated feature value with a preset detection threshold value and determining whether or not the detection target has a defect based on the comparison result.

According to another aspect of the embodiment of the present application, the electronic device includes a memory and a processor, and a computer program is stored in the memory.
An operation of positioning a detection target in an article image by a deep learning positioning network model and acquiring positioning data of the detection target.
An operation of calculating the feature value of the detection target based on the positioning data, and
Provided is an electronic device arranged so as to compare a calculated feature value with a preset detection threshold value and perform an operation of determining whether or not the detection target has a defect based on the comparison result. ..

One of the beneficial effects of the examples of the present application is as follows. Using the deep learning positioning network model, the detection target in the article image is positioned to acquire the positioning data of the detection target, the feature value of the detection target is calculated based on the positioning data, and the calculation is performed. By comparing the feature value with a preset detection threshold and determining whether or not the detection target has a defect from the comparison result, in a practical application, the requirement of production speed is satisfied. It is possible to quickly determine whether or not the detection target has a defect.
Specific embodiments of the present application are disclosed in detail with reference to the description and drawings described below, and it is also shown that the principles of the present application are used. It should be understood that the embodiments of the present application are not so limited in scope. Within the spirit of the appended claims and within the scope of the claims, the embodiments in the present application include many modifications, modifications and equivalents.

It is a schematic diagram of the article detection method in the Example of this application. It is an example figure of the positioning data in the Example of this application. It is an example figure of the data statistics in the Example of this application. It is a schematic diagram of the distribution type detection in the Example of this application. It is a schematic diagram of the article detection device in the Example of this application. It is a schematic diagram of the electronic device in the Example of this application.

The included drawings are provided for further understanding of the embodiments of the present application, and constitute a part of the specification, exemplify the embodiments of the present invention, and together with the text description of the present invention. It explains the principle. Obviously, the accompanying drawings described below are merely some examples of the present application, and other attached drawings will be based on those attached drawings, assuming that those skilled in the art will not make any creative effort. Can be obtained.

With reference to the accompanying drawings, the above and other features of the present application will become clearer according to the following description. Although the specification and drawings specifically disclose specific embodiments of the present application and show some embodiments in which the principles of the present application can be adopted, it should also be noted that the present application is not limited to the described embodiments. Should. Conversely, the present application shall include all modifications, modifications and equivalents that fall within the appended claims.

In the embodiments of the present application, the terms "first", "second", etc. are used to distinguish different elements by name, such as the spatial arrangement or temporal order of those elements. And their elements are not limited by their terminology. The term "and / or" includes any or all combinations of one or more of the related terms. The terms "including", "providing", etc. refer to the presence or absence of a feature, element, element or component of the description, excluding the presence or addition of one or more other features, elements, elements or components. do not have.

In the examples of the present application, "one", "corresponding" and the like indicating a singular aspect may include a plurality of aspects, and should be broadly understood as "type 1" or "class 1", but "1". It is not limited to the meaning of "tsu". In addition, the term "above" should be understood to include both singular and plural aspects, unless otherwise described above and below. In addition, the term "based" should be understood as "at least partly based on ..." unless otherwise explained in the upper and lower sentences.

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

(Example according to the first aspect)
The embodiments of the present application provide a method for detecting an article. FIG. 1 is a schematic diagram of an article detection method according to an embodiment of the present application. As shown in FIG. 1, the method includes the following steps.
101: Positioning the detection target in the article image by the deep learning positioning network model and acquiring the positioning data of the detection target.
102: Calculate the feature value of the detection target based on the positioning data, and
103: The calculated feature value is compared with a preset detection threshold value, and it is determined from the comparison result whether or not the detection target has a defect.

It should be noted here that FIG. 1 above is merely for exemplifying the embodiments of the present application, and the present application is not limited thereto. For example, the execution order between each step can be adjusted, some other steps can be added, or some steps can be reduced. Those skilled in the art can appropriately modify the contents according to the above contents, but the description is not limited to the above description in FIG.

In some practical examples, an article on the production line may be photographed to form an article image. Each article image may include a plurality of detection targets. For example, when the detection target is a part being manufactured and manufactured, it is possible to take a picture of each of the 48 articles to form one article image.

In some embodiments, a deep learning positioning network model may be installed in the AI device, and an article image may be input to the deep learning positioning network model. The AI device having the deep learning positioning network model may be a real hardware device such as a server or a PC, or a virtual software device such as a thread or process.

In some embodiments, the deep learning positioning network model may be implemented by a convolutional neural network (CNN), such as one or more convolutional layers, pool layers, and all connected layers. It may be included. Further, for example, the YOLO algorithm or the like may be used. Related techniques may be referred to for deep learning positioning network models. For example, Google Net model and the like may be referred to.

In some embodiments, the deep learning positioning network model may be used to position the detection target in the article image and acquire the positioning data of the detection target. For example, the positioning data of the detection target includes the type of the detection target, the center point coordinates of the rectangular frame in which the detection target is located, the width of the rectangular frame in which the detection target is located, and the rectangular frame in which the detection target is located. The height may be included. Further, the positioning data to be detected may further include a reliability indicating a reliability probability with respect to the output positioning data.

FIG. 2 is an example diagram of positioning data in the embodiment of the present application. As shown in FIG. 2, for one data sample (article image) of the collected N data samples, one or more positioning data [class, rectangle, x, by the deep learning positioning network model y, w, h] can be obtained, where class represents the type of detection target, confidence represents the reliability, and x and y represent the center point coordinates of the rectangular frame in which the detection target is located. w represents the width of the rectangular frame in which the detection target is located, and h represents the height of the rectangular frame in which the detection target is located.

For example, data [parts 1,0.9, 20, 30, 15, 16] corresponding to the detection target 202 can be acquired for the article image 201, where the type is "part 1". The reliability is 0.9, the center point coordinates of the rectangular frame are (20, 30), the width of the rectangular frame is 15, and the height of the rectangular frame is 16. Data [parts 1,0.95,100,40,15,19] corresponding to the detection target 203 can be further acquired for the article image 201, where the type is "part 1". The reliability is 0.95, the center point coordinates of the rectangular frame are (100, 40), the width of the rectangular frame is 15, and the height of the rectangular frame is 19.

However, the present application is not limited to these, and specific data may be specified according to actual needs. For example, the positioning data does not have to include type and reliability. Further, for example, the positioning data is three-dimensional data including the length l of the rectangular frame where the detection target is located, the width w of the rectangular frame where the detection target is located, the height h of the rectangular frame where the detection target is located, and the like. You may.

In some embodiments, the feature value of the detection target may be calculated based on the positioning data. For example, the feature value of the detection target is the ratio of the width to the height of the rectangular frame in which the detection target is located, the center point deviation of the rectangular frame in which the detection target is located, and the rectangle in which the detection target is located. The area deviation of the frame may be included. However, the present application is not limited to these, and specific data may be specified according to actual needs.

The present application will be further described below by taking as an example the ratio of the width to the height of the rectangular frame in which the detection target is located.

For example, if the corresponding positioning data for the detection target 202 in the article image 201 is [Part 1, 0.9, 20, 30, 15, 16], the feature value of the detection target 202 is 15/16 = 0. It is possible to detect that it is .9375. Further, if the corresponding positioning data with respect to the detection target 203 in the article image 201 is [Part 1,0.95,200,300,15,19], the feature value of the detection target 203 is 15/19 = 0. It is possible to detect that it is .7895.

In some examples, the calculated feature value may be compared with a preset detection threshold value, and it may be determined whether or not the detection target has a defect from the comparison result. Here, the detection threshold value may be predetermined according to the empirical value or the product requirement.

For example, assuming that the detection threshold value is set to 0.9, the feature value of the detection target 02 in the article image 201 is 0.9375> 0.9, so that the detection target 202 has no defects. I can judge. Further, since the feature value 0.7895 <0.9 with respect to the detection target 203 in the article image 201, it can be determined that the detection target 203 has a defect.

As a result, the detection target in the article image is positioned by the deep learning positioning network model, the positioning data of the detection target is acquired, the feature value of the detection target is calculated based on the positioning data, and the calculated feature value is calculated in advance. It is compared with the set detection threshold, and it is determined whether or not the detection target has a defect from the comparison result. Therefore, it is not necessary to calculate the edge shape of the target, and a large amount of calculation is not required, so that the real-time property is good. Further, in a practical application, the demand for production speed can be satisfied, and it is possible to quickly determine whether or not the detection target has a defect.

In some examples, it may be included to statistic (take statistics) the feature values of a plurality of detection targets, and to specify or update the detection threshold value from the result of the statistic. For example, the data distribution map may be calculated, and the normal section and the abnormal section may be divided according to the data distribution in the distribution map to specify the detection threshold value.

In some embodiments, statisticizing the feature values of a plurality of detection targets calculates the mean value of the plurality of feature values, and calculates the standard difference of the plurality of positional feature values based on the mean value. Including doing. The detection threshold can be identified or updated based on the standard deviation, which allows for more accurate detection.

For example, if Xi = w / h (an integer in which i is larger than 0), the average value may be calculated for n feature values.

Moreover, the standard difference of n feature values may be calculated based on the average value.

FIG. 3 is an example diagram of data statistics in the examples of the present application. As shown in FIG. 3, for example, a numerical value of 68.3% is in the [-1σ, 1σ] interval, a numerical value of 95.5% is in the [-2σ, 2σ] interval, and 99.73%. The value of is within the [-3σ, 3σ] interval. For example, if it is required that the [-3σ, 3σ] section is a normal section and the other sections are abnormal sections during production, the detection threshold value can be determined with reference to 3σ. Further, for example, if it is required that the [-6σ, 6σ] section is a normal section and the other sections are abnormal sections during production, the detection threshold value can be determined with reference to 6σ.

Taking 6σ as an example, the solution for determining whether or not the detection target has a defect may be as shown in Table 1.

The above has outlined how to perform article detection, but the following will explain how to further accelerate article detection.

In some embodiments, the article image may be cut out into a plurality of sub-images, and the plurality of sub-images may be input to a plurality of deep learning positioning network models to detect the detection target in the sub-images.

FIG. 4 is a schematic diagram of distributed detection in the examples of the present application. As shown in FIG. 4, for example, an article image is cut out into four sub-images (sub-images 1, 2, 3, 4) and then input to each of the four AI devices, and each AI device performs deep learning positioning. Along with having a network model, it is also possible to carry out the article detection method shown in FIG.

For example, an article image containing 48 detection targets is cut out into 4 sub-images, and each sub-image contains 12 detection targets, and each sub-image is input to one AI device. Article detection can be performed in parallel.

In some embodiments, if the detections for the plurality of sub-images are all completed within a predetermined time, the detection results for the plurality of sub-images may be synchronized. Further, the predetermined time may be started after the detection of the first sub-image among the plurality of sub-images is completed. However, the present application is not limited to these.

For example, a free list [Nan, Nan, Nan, Nan] corresponding to each of # 1 to # 4 of the AI device is created, and each time the detection by one AI device is completed, the completed time stamp is performed at the corresponding position. May be recorded. If the list is not empty (ie, the detection by the last one AI device is complete), the four data may be synchronized.

For example, as shown in FIG. 4, AI device # 2 first completes detection on sub-image 2, then AI device # 3 completes detection on sub-image 3, followed by AI device #. 1 completes the detection on the sub-image 1, and finally AI device # 4 completes the detection on the sub-image 4, in which case the data can be synchronized with the results of the four detections. That is, the results of these four detections may be combined to obtain the detection result of the article image.

In some embodiments, if the detection on at least one sub-image is not completed within a predetermined time, it may be determined that the detection on the article image is abnormal. Further, the predetermined time may be started after the detection of the first sub-image among the plurality of sub-images is completed. However, the present application is not limited to these.

For example, when the detection by the first AI device is completed, the timeout timer is activated (as shown by 401 in FIG. 4), and even if the time length T of the timer is the predetermined time. good. When the timer times out, if the data synchronization is not completed, an error notification is performed and it is determined that there is an abnormality in the detection of the article image.

As a result, it becomes possible to detect articles locally in parallel by a plurality of AI devices, further accelerating the detection and improving the real-time performance.

In some embodiments, the article image or sub-image may be gradation to reduce the size of the image before the article image or sub-image is input to the deep learning positioning network model. For example, an 800 * 600 * 3 RGB image may be converted into a 400 * 300 * 1 gradation image.

As a result, the size of the image can be reduced by performing preprocessing before entering the model, the amount of calculation can be further reduced, the real-time property can be improved, and the speed of detection can be accelerated. ..

In some embodiments, after acquiring the positioning data of the detection target by the deep learning positioning network model by inputting the article image or the sub image, another detailed position detection process can be performed. For example, based on the center point coordinates of the rectangular frame, the width of the rectangular frame, and the height of the rectangular frame output by the deep learning positioning network model, a region having an image size smaller than that of the article image or the sub image is determined, and the determined region is determined. Position is detected for. As a result, the coordinates of the center point, the width of the rectangular frame, and the height of the rectangular frame can be obtained in detail while reducing the image size and the amount of calculation. The above-mentioned region may be a rectangular frame in which predetermined margins are added to the width of the rectangular frame and the height of the rectangular frame output by the deep learning positioning network model. As the detailed position detection process, a method such as edge detection or segmentation can be used. In addition, another deep learning positioning network model can be used.

In some embodiments, the detection range may be reduced by cutting out the article image or sub-image before inputting the article image or sub-image into the deep learning positioning network model. For example, by cutting out an article image having a wide field of view, it is possible to acquire a cut-out image mainly of a detection target.

As a result, the detection range can be reduced by performing preprocessing before entering the model, the amount of calculation can be further reduced, the real-time property can be improved, and the detection speed can be accelerated. ..

Each of the above examples is merely an example of an embodiment of the present application, and the present application is not limited thereto. It may be appropriately modified based on each of the above embodiments. For example, each of the above embodiments may be used individually, or one or a plurality of the above embodiments may be combined.

As can be seen from the above embodiment, the detection target in the article image is positioned by the deep learning positioning network model, the positioning data of the detection target is acquired, and the feature value of the detection target is calculated based on the positioning data. , The calculated feature value is compared with a preset detection threshold, and it is determined from the comparison result whether or not the detection target has a defect. Thereby, in a practical application, the demand for production speed can be satisfied, and it can be quickly determined whether or not the detection target has a defect.

(Example according to the second aspect)
An embodiment of the present application provides an article detection device, which may be, for example, an electronic device or one or more members or assemblies disposed in the electronic device. Please omit the description of the same contents as those of the embodiment according to the first aspect.

FIG. 5 is a schematic view of an article detection device according to an embodiment of the present application. As shown in FIG. 5, the article detection device 500 is
A positioning unit 501 that positions a detection target in an article image and acquires positioning data of the detection target by a deep learning positioning network model.
A calculation unit 502 that calculates the feature value of the detection target based on the positioning data, and
It is provided with a comparison unit 503 that compares the calculated feature value with a preset detection threshold value and determines whether or not the detection target has a defect based on the comparison result.

In some embodiments, the detection target positioning data includes the type of the detection target, the center point coordinates of the rectangular frame in which the detection target is located, the width of the rectangular frame in which the detection target is located, and the detection target. May include the height of the rectangular frame in which is located.

In some embodiments, the feature values of the detection target are the ratio of the width to the height of the rectangular frame in which the detection target is located, or the deviation of the center point of the rectangular frame in which the detection target is located, and the detection. It may include the area deviation of the rectangular frame in which the object is located.

In some embodiments, the article detector 500 further comprises, as shown in FIG.
A statistical unit 504 that statistics the feature values of a plurality of detection targets and specifies or updates the detection threshold value from the statistical results may be further provided.

In some embodiments, the statistics unit 504 statistics the feature values of the plurality of detection targets by calculating the average value of the plurality of feature values, and based on the average value, the plurality of positional features. It may include calculating the standard difference of values.
In some embodiments, the article detector 500 further comprises, as shown in FIG.
The article image may be cut out into a plurality of sub-images, and each of the plurality of sub-images may be input to a plurality of deep learning positioning network models to include a preprocessing unit 505 for detecting a detection target in the sub-images.

In some embodiments, the pre-processing unit 505 synchronizes the detection results for the plurality of sub-images when all the detections for the plurality of sub-images are completed within a predetermined time, and at least one sub-image. If the detection on the article is not completed within a predetermined time, it may be determined that the detection on the article image is abnormal.

In some embodiments, the pre-processing unit 505 further gradations the article image or sub-image to reduce the size of the image before inputting the article image into the deep learning positioning network model. And / or, the detection range may be reduced by cutting out the article image or the sub-image.

It should be noted that although each member or module related to the present invention has been described above, the present invention is not limited to them. The article detector 500 may further include other members or modules, but see the relevant art for the specific content of those members or modules.

Further, for convenience of explanation, FIG. 5 illustrates only the connection relationship between each member or module or the transmission destination of the signal, but as can be understood by those skilled in the art, a connection by a bus or the like is shown. Related techniques may be adopted. Each of the above members or modules may be realized by hardware devices such as a processor, a memory, a transmitter, and a receiver. The examples of the present application do not limit them.

(Example according to the third aspect)
The embodiments of the present application provide an electronic device provided with the article detection device according to the embodiment according to the second aspect. Since the contents of the article detection device are included in this embodiment, the description is omitted. The electronic device may be, for example, a computer, server, workstation, laptop computer, smart phone, etc., but the embodiments of the present application are not limited thereto.

FIG. 6 is a schematic diagram of an electronic device according to an embodiment of the present application. As shown in FIG. 6, the electronic device 600 may include a processor (eg, a central processing unit (CPU)) 610 and a memory 620, the memory 620 being coupled to the processor 610. Here, various data may be stored in the memory 620. Further, a program 621 for information processing is further stored, and the program 620 is executed under the control of the processor 610.

In some embodiments, the function of the article detection device 500 may be integrated and realized in the processor 610. Here, the processor 610 is arranged so as to realize the article detection method according to the first embodiment.

In some embodiments, the article detector 500 may be located separately from the processor 610. For example, the article detection device 500 may be arranged as a chip connected to the processor 610, and the function of the article detection device 500 may be realized by control by the processor 610.

In some embodiments, the processor 610 is controlled by a deep learning positioning network model to position a detection target in an article image to acquire positioning data for the detection target, features of the detection target based on the positioning data. It is arranged so as to perform control for calculating a value and control for comparing the calculated feature value with a preset detection threshold value and determining whether or not the detection target has a defect based on the comparison result. You may.

In some embodiments, the processor 610 may be arranged to perform control to statistic the feature values of a plurality of detection targets and control to specify or update the detection threshold value from the result of the statistic.

In some embodiments, the processor 610 is arranged to perform control to calculate the average value of the plurality of feature values and control to calculate the standard difference of the plurality of position feature values based on the average value. May be done.

In some embodiments, the processor 610 controls to cut out the article image into a plurality of sub-images, inputs the plurality of sub-images into a plurality of deep learning positioning network models, respectively, and detects a detection target in the sub-images. It may be arranged to perform detection control.

In some embodiments, the processor 610 controls to synchronize the detection results for the plurality of sub-images when the detection for the plurality of sub-images is completed within a predetermined time, and

If the detection on at least one sub-image is not completed within a predetermined time, it may be arranged so as to control to determine that the detection on the article image is abnormal.

In some embodiments, the processor 610 gradations the article image or sub-image to reduce the size of the image before inputting the article image or sub-image into the deep learning positioning network model. And / or, it may be arranged so as to control to reduce the detection range by cutting out the article image or the sub image.

The above-mentioned device and method in the present invention may be realized by hardware, or may be realized by linking hardware and software. The present application relates to the following computer-readable programs. When the program is executed by the logic unit, the logic unit can realize the above-mentioned device or component, or the logic unit can realize the various methods or steps described above. The present application relates to a storage medium for storing the above program, and relates to, for example, a hard disk, a magnetic disk, an optical disk, a DVD, a flash memory, and the like.

It is possible to connect the methods / devices described in the examples of the present application and directly appear in the hardware, the software module executed by the processor, or a combination of both. For example, one or more in the functional block diagram shown in the drawing and / or one or more combinations in the functional block diagram may correspond to each software module in the computer program process and correspond to each hardware module. You may. Each of these software modules can correspond to each step shown in the drawing. These hardware modules can be realized by solidifying their software modules using, for example, a field programmable gate array (FPGA).

Software modules can be located in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disks, removable disks, CD-ROMs or any other type of memory medium known in the art. be. By combining one memory medium with the processor, the processor can read information from the memory medium and can also write information to the memory medium. Alternatively, the memory medium may be a component of the processor. The processor and memory medium may be located in the ASIC. The software module may be stored in the storage of the device or may be stored in a memory card that can be inserted into the device. For example, when a MEGA-SIM card having a relatively large capacity or a large-capacity flash memory device is adopted, the software module may be stored in the MEGA-SIM card or a large-capacity flash memory.

A general purpose processor, digital signal processor (DSP), specific for realizing the functions described in the present application for one or more of the functional blocks and / or one or more combinations of functional blocks described in the accompanying drawings. It may be implemented as an application-specific integrated circuit (ASIC), field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any suitable combination. One or more of the functional blocks and / or one or more combinations of the functional blocks described in the accompanying drawings may be realized as a combination of calculation devices. For example, it may be implemented as a combination of DSP and microprocessor, multiple microprocessors, one or more microprocessors in combination with DSP communication, or any other such arrangement.

The present application has been described above by linking specific embodiments. However, it should be understood by those skilled in the art that all of these descriptions are merely exemplary and are not a limitation on the scope of protection of the present application. Those skilled in the art can make various modifications and modifications to the present application based on the spirit and principles of the present application, but these modifications and modifications are also within the scope of the present invention.

The preferred embodiment of the present application has been described above with reference to the accompanying drawings. Since many features and advantages of those embodiments are apparent in the details of the specification above, the appended claims are the true spirit and scope of those embodiments. It aims to cover all of those features and benefits contained within. Moreover, since a person skilled in the art should be able to easily conceive many modifications and modifications, the embodiments in the present application are not limited to the precise structures and operations illustrated and described, but are within the scope thereof. It can cover all the appropriate modifications, deformations, and equivalents contained in it.

Claims

It is an article detection method
To position the detection target in the article image by the deep learning positioning network model and acquire the positioning data of the detection target.
Calculate the feature value of the detection target based on the positioning data, and
An article detection method comprising comparing a calculated feature value with a preset detection threshold value and determining whether or not the detection target has a defect from the comparison result.
The positioning data of the detection target includes the type of the detection target, the center point coordinates of the rectangular frame in which the detection target is located, the width of the rectangular frame in which the detection target is located, and the height of the rectangular frame in which the detection target is located. The method according to claim 1, comprising the above.
The feature value of the detection target is
The ratio of the width to the height of the rectangular frame in which the detection target is located, or the deviation of the center point of the rectangular frame in which the detection target is located, and
The method according to claim 1, wherein the area deviation of the rectangular frame in which the detection target is located is included.
Statistics of feature values of multiple detection targets, and
The method of claim 1, further comprising identifying or updating the detection threshold from the statistical results.
Statistics of feature values of multiple detection targets is
Calculating the average value of multiple feature values, and
The method of claim 4, wherein the standard difference of the plurality of positional feature values is calculated based on the average value.
Cutting out the article image into a plurality of sub-images, and
The method according to claim 1, further comprising inputting each of the plurality of sub-images into a plurality of deep learning positioning network models to detect a detection target in the sub-images.
When all the detections on the plurality of sub-images are completed within a predetermined time, the results of the detections on the plurality of sub-images are synchronized, and
The method according to claim 6, wherein if the detection on at least one sub-image is not completed within a predetermined time, it is determined that the detection on the article image is abnormal.
Before inputting the article image or the sub-image into the deep learning positioning network model, the article image or the sub-image is graded to reduce the size of the image and / or the article image or the sub-image. The method according to claim 6 or 7, further comprising reducing the range of detection by cutting out.
It is an article detection device,
A positioning unit that positions the detection target in the article image and acquires the positioning data of the detection target by the deep learning positioning network model.
A calculation unit that calculates the feature value of the detection target based on the positioning data,
An article detection device including a comparison unit that compares a calculated feature value with a preset detection threshold value and determines whether or not the detection target has a defect based on the comparison result.
Equipped with memory and processor
A computer program is stored in the memory.
The processor is driven by the computer program.
An operation of positioning a detection target in an article image by a deep learning positioning network model and acquiring positioning data of the detection target.
Based on the positioning data, the operation of calculating the feature value of the detection target, and comparing the calculated feature value with a preset detection threshold value, and based on the comparison result, whether or not the detection target has a defect. An electronic device that performs an operation to determine.