A still camera system for deep learning of a single chip microcomputer

Abstract

The present application relates to the field of optical measurement, and more particularly to a kind of static image instrument system for the depth learning of single-chip microcomputer, comprising: image acquisition unit to collect the surface image of production line print;With the contour deformation characteristic value of the surface image, determine whether the operation of the stroboscopic static image instrument meets the preset standard operation feedback unit;With the image pre-data storage unit for storing reference surface image;With the data processing unit for constructing data set;With the data storage unit for storing the data set of target screening type;With the data entry unit for obtaining pre-training data set;With the scale characteristic value of the pre-training data set, when the stability of pre-training data set does not meet the preset standard, reduce the time interval, or, adjust the weight of the data set corresponding to each target screening type control unit, so as to improve image detection precision.

Description

TECHNICAL FIELD

[0001] The application relates to the field of optical measurement, in particular to a still picture instrument system for deep learning of a single-chip microcomputer. BACKGROUND

[0002] A stroboscopic still picture instrument is an instrument for detecting the appearance of a printed matter, which is an optical measurement device for changing a vibrating, high-speed rotating or periodically moving component into a component that is not moving.

[0003] Traditional visual inspection systems have long been challenged by motion blur and environmental light interference. The conventional scheme uses a continuous light source with an ordinary industrial camera, which can realize dynamic acquisition, but is limited by the target displacement and environmental stray light within the exposure time, and the image quality is difficult to guarantee, resulting in insufficient defect detection accuracy.

[0004] The introduction of stroboscopic still picture technology effectively breaks through this limitation. By freezing the moving target with high-frequency flash and precise timing control, the motion blur effect is significantly reduced, providing a hardware foundation for high-quality image acquisition. By applying the still picture instrument to high-speed production equipment and synchronizing the working frequency of the still picture instrument with the running speed of the equipment, one can directly see the appearance of the high-speed running product. Although the measured object is moving at high speed, it appears to be moving slowly or relatively stationary.

[0005] However, relying solely on the stroboscopic still picture instrument for defect detection still has limitations. Traditional algorithms based on manually designed features for defect classification are difficult to cope with the diverse defect patterns on complex printed matter surfaces. With the development of deep learning technology, defect recognition methods based on convolutional neural networks have shown great potential, but the model generalization ability is still restricted by the distribution of training data.

[0006] Chinese Patent Publication No. CN116183616A discloses a high-precision low-cost LED still picture instrument device based on a single-chip microcomputer, which includes a still picture instrument housing, an LED lamp assembly, a self-designed circuit board, a domestic STC32G128K type single-chip microcomputer, a lithium battery power module, a control assembly, and a driving circuit assembly. The application belongs to the field of optical measurement and mechanical electronics, and specifically refers to a high-precision low-cost LED still picture instrument device based on a single-chip microcomputer, which effectively solves the problems of printed packaging quality detection, paper product surface detection, and monitoring of cavitation phenomena in turbine propulsion systems, and breaks through the limitations of high cost, dim light, and inability to use in bright environments of the stroboscopic still picture instrument on the market. Moreover, all materials used in the application are domestic.

[0007] It can be seen that the above technical scheme only realizes the basic image acquisition function, does not consider the influence of the deformation of the collected image on the image quality detection of the printed packaging surface, does not consider the influence of the overexposure artifact on the image detection, and does not consider the multi-dimensional analysis and feedback optimization of the image data, thereby causing the problem of low image detection precision. SUMMARY

[0008] Therefore, the present application provides a still image camera system for deep learning of a single-chip microcomputer to overcome the problem in the prior art that the influence of the deformation of the collected image on the image quality detection of the printed packaging surface is not considered, the influence of the overexposure artifact on the image detection is not considered, and the multi-dimensional analysis and feedback optimization of the image data are not considered, thereby causing the problem of low image detection precision.

[0009] To achieve the above-mentioned purpose, the present application provides a still image camera system for deep learning of a single-chip microcomputer, comprising:

[0010] an image pre-data storage unit for storing a reference surface image of a printed matter in a defect-free state;

[0011] an image acquisition unit for collecting surface images of printed matters on a production line at preset time intervals using a stroboscopic still image camera;

[0012] a running feedback unit connected to the image acquisition unit and the image pre-data storage unit, respectively, for determining whether the running of the stroboscopic still image camera meets the preset standard according to the contour deformation characteristic value of the surface image;

[0013] a data processing unit connected to the image pre-data storage unit and the running feedback unit, for screening a plurality of surface images collected within a preset time period according to target screening information to construct a data set, wherein the target screening information includes a target screening type and target screening data related to the target screening type;

[0014] a data storage unit connected to the data processing unit for storing the data set of the target screening type;

[0015] a data entry unit connected to the data storage unit for inputting data in the data set to a ResNet network model to obtain a pre-training data set;

[0016] a control unit connected to the data entry unit for reducing the time interval when the stability of the pre-training data set does not meet the preset standard according to the scale characteristic value of the pre-training data set, or adjusting the weight of the data set corresponding to each target screening type.

[0017] Further, the operation feedback unit determines whether the operation of the stroboscopic still image device meets the preset standard according to the contour deformation characteristic value of the surface image, wherein,

[0018] If the contour deformation characteristic value is less than a first preset contour deformation characteristic value, it is determined that the operation of the stroboscopic still image device meets the preset standard.

[0019] If the contour deformation characteristic value is greater than or equal to the first preset contour deformation characteristic value and less than a second preset contour deformation characteristic value, it is determined that the operation of the stroboscopic still image device does not meet the preset standard, and the operation of the stroboscopic still image device is further determined according to the artifact area ratio of the surface image.

[0020] If the contour deformation characteristic value is greater than or equal to the second preset contour deformation characteristic value, it is determined that the operation of the stroboscopic still image device does not meet the preset standard, and the exposure time of the stroboscopic still image device is reduced according to the difference between the contour deformation characteristic value and the second preset contour deformation characteristic value.

[0021] The contour deformation characteristic value is a ratio of a contour area of the surface image to a contour area of the reference surface image.

[0022] Further, the operation feedback unit determines whether the operation of the stroboscopic still image device meets the preset standard according to the artifact area ratio of the surface image.

[0023] If the artifact area ratio is less than a preset artifact area ratio, it is determined that the operation of the stroboscopic still image device meets the preset standard.

[0024] If the artifact area ratio is greater than or equal to the preset artifact area ratio, it is determined that the operation of the stroboscopic still image device does not meet the preset standard, and the illumination brightness of the stroboscopic still image device is increased according to the difference between the artifact area ratio and the preset artifact area ratio.

[0025] The artifact area ratio is a ratio of an artifact pixel number to a total image pixel number of the surface image, wherein the artifact pixel number is a number of pixel points whose gray value is less than a gray value of a corresponding pixel point of the reference surface image.

[0026] Further, the operation feedback unit sets several time length adjustment modes for the reduction of the exposure time of the stroboscopic still image device, and each time length adjustment mode has a different reduction amplitude of the exposure time of the stroboscopic still image device.

[0027] Further, the target screening type includes the following multiple or all:

[0028] The average hue value of the surface image exceeds a preset average hue value; or

[0029] a variance of gray values of the surface image exceeds a preset variance of gray values; or

[0030] a texture repeatability error of the surface image exceeds a preset texture repeatability error; or

[0031] an edge integrity coefficient of the surface image is lower than a preset edge integrity coefficient; or

[0032] an area proportion of an abnormal local contrast region of the surface image exceeds a preset area proportion.

[0033] Further, the target screening data include the following multiple or all:

[0034] an average hue value of the surface image; or

[0035] a variance of gray values of the surface image; or

[0036] a texture repeatability error of the surface image; or

[0037] an edge integrity coefficient of the surface image; or

[0038] an area proportion of an abnormal local contrast region of the surface image.

[0039] Further, the control unit determines whether the stability of the pre-training data set meets a preset standard according to a scale characteristic representation value of the pre-training data set, wherein

[0040] if the scale characteristic representation value is less than a first preset scale characteristic representation value, it is determined that the stability of the pre-training data set meets the preset standard;

[0041] if the scale characteristic representation value is greater than or equal to the first preset scale characteristic representation value and less than a second preset scale characteristic representation value, it is determined that the stability of the pre-training data set does not meet the preset standard, and the time interval is reduced according to a difference between the scale characteristic representation value and the first preset scale characteristic representation value;

[0042] if the scale characteristic representation value is greater than or equal to the second preset scale characteristic representation value, it is determined that the stability of the pre-training data set does not meet the preset standard, and the weight of the data set corresponding to each target screening type is adjusted.

[0043] Further, the scale characteristic representation value is determined by a sample number variance of the pre-training data set and a sample number average value of the pre-training data set.

[0044] Further, the control unit adjusts the weight of the data set corresponding to each target screening type in the following manner: arranging the data set corresponding to each target screening type in descending order according to the sample size of the data set, and reallocating the weight by using a weighted distribution loss function according to the sample size.

[0045] Further, the reduction range of the time interval is positively correlated with the scale characteristic difference value, wherein the scale characteristic difference value is the difference between the scale characteristic value and the first preset scale characteristic value.

[0046] Compared with the prior art, the present application has the beneficial effects that the present application realizes real-time parameter adjustment of the static image instrument by integrating image acquisition and operation feedback; and realizes multi-dimensional dynamic detection of the surface image of the printed matter and adaptive optimization of the deep learning model by image data processing, contour deformation characteristic value, multi-dimensional data screening and weight distribution optimization of the training data set, thereby improving the packaging quality detection precision.

[0047] Further, the present application can dynamically evaluate the operation state of the stroboscope and the image quality by setting the double judgment mechanism of the contour deformation characteristic value and the artifact area ratio, thereby improving the image acquisition efficiency.

[0048] Further, when the contour deformation is not up to standard, the present application triggers secondary artifact area analysis, and realizes combination of gray value difference, thereby effectively solving the problem of image deformation caused by the shooting instrument.

[0049] Further, the exposure time is dynamically adjusted according to the contour deformation characteristic value and the preset difference value, so as to avoid the interference of artifacts caused by overexposure or underexposure, thereby realizing accurate regulation and control of the exposure time.

[0050] Further, the present application covers the defect types by integrating multi-dimensional image feature extraction such as geometry, texture and color, detects deformation defects by contour area ratio, identifies problems such as uneven ink and overprint deviation by texture feature analysis, and realizes color difference defect detection by color feature analysis, thereby improving the image detection precision.

[0051] Further, the present application quantifies the stability of the data set by the scale characteristic value, and assigns a higher weight to the small sample data set under the condition that the stability of the pre-training data set does not meet the preset standard, solves the class imbalance problem, improves the detection sensitivity of the minority class defects, and reallocates the weight after arranging the sample size in descending order, thereby reducing the training shock risk and accelerating the model convergence, thereby improving the detection reliability. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a unit connection schematic diagram of the static image instrument system for deep learning of the single-chip microcomputer according to the embodiments of the present application.

[0053] Figure 2 A flow chart for determining whether the operation of the stroboscopic still camera meets the preset standard according to the contour deformation characteristic value of the surface image of the embodiment of the present application;

[0054] Figure 3 A flow chart for twice determining whether the operation of the stroboscopic still camera meets the preset standard according to the artifact area proportion of the surface image of the embodiment of the present application;

[0055] Figure 4 A flow chart for determining whether the stability of the pre-training data set meets the preset standard of the embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to make the objects and advantages of the present application clearer, the present application will be further described below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0057] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application and not to limit the protection scope of the present application.

[0058] Please refer to Figure 1 , Figure 2 , Figure 3 and Figure 4 , which are respectively a unit connection schematic diagram of a stroboscopic still camera system for deep learning of a single-chip microcomputer according to an embodiment of the present application, a flow chart for determining whether the operation of the stroboscopic still camera meets the preset standard according to the contour deformation characteristic value of the surface image of the embodiment of the present application, a flow chart for twice determining whether the operation of the stroboscopic still camera meets the preset standard according to the artifact area proportion of the surface image of the embodiment of the present application, and a flow chart for determining whether the stability of the pre-training data set meets the preset standard of the embodiment of the present application.

[0059] An embodiment of the present application is a stroboscopic still camera system for deep learning of a single-chip microcomputer, comprising:

[0060] An image pre-data storage unit is used to store the reference surface image of the printed matter in the defect-free state;

[0061] An image acquisition unit is used to collect the surface image of the printed matter on the production line by using the stroboscopic still camera according to the preset time interval 5ms / time, the stroboscopic exposure time 500μs;

[0062] An operation feedback unit is connected with the image acquisition unit and the image pre-data storage unit respectively, and is used to determine whether the operation of the stroboscopic still camera meets the preset standard according to the contour deformation characteristic value of the surface image.

[0063] a data processing unit connected with the image pre-data storage unit and the operation feedback unit, configured to screen a plurality of surface images collected within a preset time length 3s according to target screening information to construct a data set, wherein the target screening information comprises a target screening type and target screening data related to the target screening type;

[0064] a data storage unit connected with the data processing unit, configured to store the data set of the target screening type;

[0065] a data entry unit connected with the data storage unit, configured to input data in the data set to a ResNet network model to obtain a pre-training data set;

[0066] a control unit connected with the data entry unit, configured to reduce the time interval or adjust the weight of the data set corresponding to each target screening type when the stability of the pre-training data set does not meet a preset standard according to a scale characteristic representation value of the pre-training data set.

[0067] In this embodiment, the specific structure of the operation feedback unit and the control unit is not limited, and each unit thereof can be composed of a logic component, including a field programmable component, a computer or a microprocessor in the computer.

[0068] Specifically, the operation feedback unit determines whether the operation of the stroboscopic still camera meets a preset standard according to a contour deformation representation value of the surface image, wherein,

[0069] if the contour deformation representation value is less than a first preset contour deformation representation value 0.94, it is determined that the operation of the stroboscopic still camera meets the preset standard;

[0070] if the contour deformation representation value is greater than or equal to the first preset contour deformation representation value and less than a second preset contour deformation representation value 1.08, it is determined that the operation of the stroboscopic still camera does not meet the preset standard, and whether the operation of the stroboscopic still camera meets the preset standard is determined again according to a artifact area ratio of the surface image;

[0071] if the contour deformation representation value is greater than or equal to the second preset contour deformation representation value, it is determined that the operation of the stroboscopic still camera does not meet the preset standard, and the exposure time length of the stroboscopic still camera is reduced according to the difference between the contour deformation representation value and the second preset contour deformation representation value;

[0072] the contour deformation representation value is a ratio of a contour area of the surface image to a contour area of the reference surface image, wherein the contour area is obtained by edge detection.

[0073] In this embodiment, the first preset profile deformation representation value is selected as 0.94, and the second preset profile deformation representation value is selected as 1.08.

[0074] Specifically, the running feedback unit determines whether the running of the stroboscopic still image device meets the preset standard according to the artifact area proportion twice.

[0075] If the artifact area proportion is less than the preset artifact area proportion 0.11, it is determined that the running of the stroboscopic still image device meets the preset standard.

[0076] If the artifact area proportion is greater than or equal to the preset artifact area proportion, it is determined that the running of the stroboscopic still image device does not meet the preset standard, and the light brightness of the stroboscopic still image device is increased according to the difference between the artifact area proportion and the preset artifact area proportion.

[0077] The artifact area proportion is the ratio of the number of artifact pixels to the total number of image pixels of the surface image, wherein the number of artifact pixels is the number of pixel points whose gray value is less than the gray value of the corresponding pixel point of the reference surface image.

[0078] In this embodiment, the preset artifact area proportion is selected as 0.11, but the above value is not limited thereto, and those skilled in the art can adjust the value according to actual needs.

[0079] Specifically, the running feedback unit sets several time length adjustment modes for the reduction of the exposure time length of the stroboscopic still image device, wherein,

[0080] If the profile deformation representation difference value is less than the first preset profile deformation representation difference value 0.08, the exposure time length of the stroboscopic still image device is reduced to a corresponding value by using a first time length adjustment coefficient 0.98.

[0081] If the profile deformation representation difference value is greater than or equal to the first preset profile deformation representation difference value and less than the second preset profile deformation representation difference value 0.16, the exposure time length of the stroboscopic still image device is reduced to a corresponding value by using a second time length adjustment coefficient 0.96.

[0082] If the profile deformation representation difference value is greater than or equal to the second preset profile deformation representation difference value, the exposure time length of the stroboscopic still image device is reduced to a corresponding value by using a third time length adjustment coefficient 0.94.

[0083] The profile deformation representation difference value is the difference between the profile deformation representation value and the second preset profile deformation representation value.

[0084] Specifically, the target screening type includes the following multiple or all:

[0085] The average hue value of the surface image exceeds a preset average hue value, wherein the preset average hue value is 1.05 times the average hue value of a reference surface image, and the multiple in this embodiment is 1.05, but the above value is not limited thereto, and a person skilled in the art can adjust the value according to actual needs.

[0086] The gray value variance of the surface image exceeds a preset gray value variance, wherein the preset gray value variance is 1.13 times the gray value variance of a reference surface image, but the above multiple value is not limited thereto, and a person skilled in the art can adjust the value according to actual needs.

[0087] The texture repeatability error of the surface image exceeds a preset texture repeatability error, and the preset texture repeatability error is set to 0.08 mm; or

[0088] The edge integrity coefficient of the surface image is lower than a preset edge integrity coefficient, and the preset edge integrity coefficient is set to 0.90; or

[0089] The local contrast abnormal area area ratio of the surface image exceeds a preset area ratio, and the preset area ratio is 0.25.

[0090] Specifically, the target screening data includes the following multiple or all:

[0091] The average hue value of the surface image, wherein the average hue value is obtained by Halcon; or

[0092] The gray value variance of the surface image, which is obtained by MATLAB; or

[0093] The texture repeatability error of the surface image, wherein the texture repeatability error is obtained by MATLAB; or

[0094] The edge integrity coefficient of the surface image, wherein the edge integrity coefficient is obtained by an image processing software; or

[0095] The local contrast abnormal area area ratio of the surface image, wherein the surface image is divided into a plurality of unit cells, the standard deviation of the gray value of each unit cell is calculated, and the area ratio of the standard deviation of the gray value exceeding the standard deviation of the gray value of the corresponding unit cell of the reference image is counted.

[0096] Specifically, the control unit determines whether the stability of the pre-training data set meets a preset standard according to the scale characteristic representation value of the pre-training data set, wherein

[0097] If the scale characteristic value is less than a first preset scale characteristic value 0.65, it is determined that the stability of the pre-training data set meets the preset standard.

[0098] If the scale characteristic value is greater than or equal to the first preset scale characteristic value and less than a second preset scale characteristic value 1.08, it is determined that the stability of the pre-training data set does not meet the preset standard, and the time interval is reduced according to the difference between the scale characteristic value and the first preset scale characteristic value.

[0099] If the scale characteristic value is greater than or equal to the second preset scale characteristic value, it is determined that the stability of the pre-training data set does not meet the preset standard, and the weight of the data set corresponding to each target screening type is adjusted.

[0100] Specifically, the scale characteristic value is calculated by the following formula:

[0101]

[0102] In the formula, P represents the scale characteristic value; λ1 represents the first evaluation coefficient, λ1 = 0.47; λ2 represents the second evaluation coefficient, λ2 = 0.63; σ 2 1 represents the sample number variance of each target type in the pre-training data set; σ 2 0 represents the preset sample number variance; μ0 represents the preset sample number average; μ1 represents the sample number average of each target type in the pre-training data set.

[0103] Specifically, the control unit adjusts the weight of the data set corresponding to each target screening type in the following manner: arranging the data sets corresponding to each target screening type in descending order according to the sample amount of the data sets, and redistributing the weights using a weighted distribution loss function according to the sample amount.

[0104] Specifically, the reduction range of the time interval is positively correlated with the scale characteristic difference value, wherein the positive correlation is, for example, linear positive correlation or nonlinear positive correlation. The linear slope of the linear positive correlation is not specifically limited. It can be understood that the larger the scale characteristic difference value, the larger the reduction range of the time interval.

[0105] The scale characteristic difference value is the difference between the scale characteristic value and the first preset scale characteristic value.

[0106] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0107] The above only describes the preferred embodiments of the present application and is not intended to limit the present application; the present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A still imager system for deep learning using a microcontroller, characterized in that, include: Image pre-data storage unit, used to store reference surface images of printed materials in a defect-free state; The image acquisition unit is used to acquire surface images of printed materials on the production line at preset time intervals using a stroboscope. A feedback unit, connected to both the image acquisition unit and the image pre-data storage unit, is used to determine whether the operation of the stroboscopic still imager meets preset standards based on the contour deformation characterization values ​​of the surface image. If the contour deformation characterization value is less than the first preset contour deformation characterization value, then the operation of the stroboscopic still imager is determined to meet the preset standard. If the contour deformation characterization value is greater than or equal to the first preset contour deformation characterization value and less than the second preset contour deformation characterization value, it is determined that the operation of the stroboscopic still imager does not meet the preset standard, and the operation of the stroboscopic still imager is determined a second time based on the proportion of artifact area of ​​the surface image. If the contour deformation characterization value is greater than or equal to the second preset contour deformation characterization value, it is determined that the operation of the stroboscopic still imager does not meet the preset standard, and the exposure time of the stroboscopic still imager is reduced according to the difference between the contour deformation characterization value and the second preset contour deformation characterization value. The contour deformation characterization value is the ratio of the contour area of ​​the surface image to the contour area of ​​the reference surface image. A data processing unit, which is connected to the image pre-data storage unit and the operation feedback unit, is used to screen several surface images collected within a preset time period according to the target screening information to construct a dataset. The target screening information includes the target screening type and the target screening data related to the target screening type. A data storage unit, connected to the data processing unit, is used to store the dataset of the target filtering type; A data input unit, which is connected to the data storage unit, is used to input the data in the dataset into the ResNet network model to obtain a pre-trained dataset; The control unit, which is connected to the data input unit, is used to reduce the time interval or adjust the weight of the dataset corresponding to each target screening type when the stability of the pre-training dataset does not meet the preset standard based on the scale characteristic characterization value of the pre-training dataset.

2. The still image system for deep learning using a microcontroller according to claim 1, characterized in that, The operation feedback unit makes a secondary determination based on the proportion of artifact area in the surface image to determine whether the operation of the stroboscopic still imager meets the preset standard, wherein... If the proportion of artifact area is less than the preset proportion of artifact area, it is determined that the operation of the stroboscopic still imager meets the preset standard. If the proportion of artifact area is greater than or equal to the preset proportion of artifact area, it is determined that the operation of the stroboscope does not meet the preset standard, and the illumination of the stroboscope is increased according to the difference between the proportion of artifact area and the preset proportion of artifact area. The artifact area ratio refers to the ratio of the number of artifact pixels to the total number of pixels in the surface image, wherein the number of artifact pixels is the number of pixels in the surface image whose grayscale value is less than the grayscale value of the corresponding pixel in the reference surface image.

3. The still image system for deep learning using a microcontroller according to claim 2, characterized in that, The operation feedback unit has several duration adjustment methods for reducing the exposure time of the stroboscopic still imager, and each duration adjustment method reduces the exposure time of the stroboscopic still imager by a different amount.

4. The still image system for deep learning using a microcontroller according to claim 3, characterized in that, The target filtering types include several or all of the following: The average hue value of the surface image exceeds a preset average hue value; or The variance of the grayscale values ​​of the surface image exceeds the preset variance of the grayscale values; or The texture repeatability error of the surface image exceeds a preset texture repeatability error; or The edge integrity coefficient of the surface image is lower than a preset edge integrity coefficient; or The area of ​​the local contrast abnormal region in the surface image exceeds the preset area ratio.

5. The still image system for deep learning using a microcontroller according to claim 4, characterized in that, The target screening data includes some or all of the following: The average hue value of the surface image; or The variance of the grayscale values ​​of the surface image; or The texture repeatability error of the surface image; or The edge integrity coefficient of the surface image; or The percentage of the area of ​​local contrast anomalies in the surface image.

6. The still image system for deep learning using a microcontroller according to claim 5, characterized in that, The control unit determines whether the stability of the pre-training dataset meets a preset standard based on the size characteristic representation value of the pre-training dataset, wherein, If the scale characteristic representation value is less than the first preset scale characteristic representation value, then the stability of the pre-trained dataset is determined to meet the preset standard. If the scale characteristic representation value is greater than or equal to the first preset scale characteristic representation value and less than the second preset scale characteristic representation value, then it is determined that the stability of the pre-training dataset does not meet the preset standard, and the time interval is reduced according to the difference between the scale characteristic representation value and the first preset scale characteristic representation value. If the scale characteristic representation value is greater than or equal to the second preset scale characteristic representation value, it is determined that the stability of the pre-trained dataset does not meet the preset standard, and the weights of the datasets corresponding to each target screening type are adjusted.

7. The still image system for deep learning using a microcontroller according to claim 6, characterized in that, The scale characteristic value is determined by the variance of the number of samples in the pre-training dataset and the average number of samples in the pre-training dataset.

8. The still image system for deep learning using a microcontroller according to claim 7, characterized in that, The control unit adjusts the weights of the datasets corresponding to each target screening type by sorting the datasets in descending order according to the sample size of each target screening type, and then redistributing the weights using a weighted loss function based on the sample size.

9. The still image system for deep learning using a microcontroller according to claim 8, characterized in that, The reduction in the time interval is positively correlated with the difference in scale characteristic representation, wherein the difference in scale characteristic representation is the difference between the scale characteristic representation value and the first preset scale characteristic representation value.

Images