Method, electronic device, and computer program product for image marking standardization

Abstract

A method, an electronic device, and a computer program product for image marking standardization. The method for image marking standardization includes: receiving a defect pattern; marking the image according to the defect pattern to generate a first judgment result; marking the image according to the defect pattern to generate a second judgment result; comparing the first judgment result and the second judgment result to obtain a comparison result; and updating the defect pattern according to the comparison result to standardize the defect pattern. The image marking method, electronic device, and computer program product of the present invention can improve the marking stability of the training data of the trained image recognition algorithm, thereby improving the correct rate of image recognition of the trained image recognition algorithm.

Description

Technical Field

[0001] The present invention relates to a method for image marking, and more particularly to an image marking method using a convolutional neural network for image recognition. Background Art

[0002] In the field of artificial intelligence machine learning in existing industrial production, due to specialized division of labor and a large amount of data, usually a professional marker performs data marking, a researcher trains an image recognition model and confirms the effect, and then a technician on the production line compares the marking results marked by the professional marker and the image recognition model. The purpose of the researcher to confirm the effect is to confirm whether the image recognition model has successfully achieved knowledge transfer to achieve the effect of replacing human thinking and decision-making. Therefore, although existing artificial intelligence machine learning usually suffices to meet its intended purpose, they are not entirely satisfactory in every aspect.

[0003] Therefore, a method for image marking standardization, a computer program product, and an electronic device applying the method are needed to solve the above problems. Summary of the Invention

[0004] The method for image marking standardization according to an embodiment of the present invention includes: receiving a defect appearance; marking an image according to the defect appearance to generate a first judgment result; marking the image according to the defect appearance to generate a second judgment result; comparing the first judgment result and the second judgment result to obtain a comparison result; and updating the defect appearance according to the comparison result to standardize the defect appearance.

[0005] The method as described above, wherein the step of marking the image includes: receiving a control signal from a user interface (UI), and according to the control signal, marking the image as a defective image or a non-defective image, or deleting the image.

[0006] The method as described above, wherein the step of marking the image further includes: receiving a control signal from the user interface and storing the image in a corresponding storage space according to the control signal.

[0007] The method as described above, wherein the step of generating the first judgment result includes: receiving at different time points the first judgment result and a re-judgment result obtained by judging the image according to the defect appearance; and when the re-judgment result is the same as the first judgment result, marking the image according to the first judgment result; wherein the first judgment result records whether the image is a defective image.

[0008] The method as described above, the method for standardizing image marking further includes: obtaining a specification version of a defect pattern; wherein the specification version records the update date of the defect pattern.

[0009] The method as described above, wherein the step of generating the second judgment result includes: using the image as training data to generate a trained image recognition algorithm; and using the image recognition algorithm to determine whether the image is a defect image to obtain the second judgment result.

[0010] The method as described above, the method for standardizing image marking further includes: auditing the marking result; wherein the step of auditing the marking result includes: receiving a first auditing result and a second auditing result at different time points; the first auditing result and the second auditing result record the number of defect images and non-defect images included in the same set of multiple images at different time points; and when the ratio (moving rate) of the first auditing result being different from the second auditing result is lower than the first threshold, it is determined that the marking result is qualified.

[0011] The method as described above, wherein the step of updating the specification version of the defect pattern includes: increasing or decreasing the number of images corresponding to at least one defect pattern; or increasing or decreasing the number of types of defect patterns defined in the defect setting file.

[0012] An electronic device according to an embodiment of the present invention, for standardizing image marking, the electronic device includes: a user interface and a processor. The user interface receives a defect pattern and receives a first judgment result generated by marking an image according to the defect pattern. The processor receives the defect pattern and the first judgment result from the user interface, and marks the image according to the defect pattern to generate a second judgment result. The processor compares the first judgment result and the second judgment result to obtain a comparison result. The processor updates the defect pattern according to the comparison result to standardize the defect pattern.

[0013] A computer program product according to an embodiment of the present invention, for standardizing image marking, applicable to a computer having a processor, the computer program product includes: a first receiving instruction, a second receiving instruction, a judgment instruction, a comparison instruction, and an update instruction. The first receiving instruction causes the processor to receive a defect pattern. The second receiving instruction causes the processor to receive a first judgment result generated by marking an image according to the defect pattern. The judgment instruction causes the processor to mark the image according to the defect pattern to generate a second judgment result. The comparison instruction causes the processor to compare the first judgment result and the second judgment result to obtain a comparison result. The update instruction causes the processor to update the defect pattern according to the comparison result to standardize the defect pattern.

[0014] The image marking method of the present invention, the electronic device and the computer program product for image marking standardization can improve the marking stability of the training data of the trained image recognition algorithm, thereby improving the correct rate of image recognition of the trained image recognition algorithm. Description of the Drawings

[0015] Figure 1 It is a flowchart of the method for image marking standardization according to an embodiment of the present invention.

[0016] Figure 2 It is a schematic diagram of the user interface of the method for image marking standardization according to an embodiment of the present invention.

[0017] Figure 3 It is a schematic diagram of the structure of the deep residual network of the method for image marking standardization according to an embodiment of the present invention.

[0018] Figure 4 It is a schematic diagram of the structure of the convolution block and the identify block of the deep residual network of the method for image marking standardization according to an embodiment of the present invention.

[0019] Figure 5A 、 Figure 5B 、 Figure 5C It is a verification flowchart of the convolutional neural network of the method for image marking standardization according to an embodiment of the present invention.

[0020] Figure 6 It is a setting diagram when the method for image marking standardization according to an embodiment of the present invention is applied to the production line.

[0021] Figure 7 It is a curve graph of the overkill rate and the moving rate of the product after the method for image marking standardization according to an embodiment of the present invention is applied to the production line.

[0022] Figure 8A 、 Figure 8B It is a schematic diagram of the defect mode of the method for image marking standardization according to an embodiment of the present invention.

[0023] Main Component Symbol Description:

[0024] Steps S100, S102, S104, S106, S108

[0025] 200 User Interface

[0026] 202 Image Display

[0027] 204, 206 Image Marking Objects

[0028] 208 Image Deletion Object

[0029] 210 Image Marking Display

[0030] 212 Image Opening Object

[0031] 214 Marking Completion Object

[0032] 216 Marking Progress Display

[0033] 218 Previous Page Object

[0034] 220 Next Page Object

[0035] 300 Deep Residual Network

[0036] 302 Convolutional Layer

[0037] 304 Residual Layer

[0038] 306 Global Pooling Layer

[0039] 308 Fully Connected Layer

[0040] 310 Image

[0041] 312 Numerical Value

[0042] 400 Convolutional Block

[0043] 402, 404, 406, 408 Convolutional Layers

[0044] 410 Identification Block

[0045] 412, 414, 416 Convolutional Layers

[0046] 420, 422, 424, 430, 432 Feature Images

[0047] S500, S502, S504, S506, S508 Steps

[0048] S510, S512, S514, S516, S518, S520 Steps

[0049] S530, S532, S534, S536, S538 Steps

[0050] 600 Production Line

[0051] 602 Company Intranet

[0052] 604, 610 Computers

[0053] 606 Router

[0054] 608 AOI Device

[0055] 612 Edge Computing Device

[0056] 700, 702 Curves

[0057] FP Flatness, Link Bar Gap, SC Pin Missing, FPHook deformation, Hook Broken, Snap Broken, Link Bar Width Fail, Hook not engaged, Snap not engaged, Link Bar Missing, SC Pin Broken, FP HOLE BROKEN Machine Error Code, Specification Sheet Error Code

[0058] PIN-MISSING1-10, FP_DEFORMATION, Defect Sub-state

[0059] FP_DEFORMATION2-6, FP_WARP, FP_WARP1-5,

[0060] HK-BROKEN1-7, SNAP-BROKEN, SNAP-BROKEN1-7,

[0061] LINK_BAR_WIDTH_RB, LINK_BAR_WIDTH_LT

[0062] LINK_BAR_WIDTH_LB, LINK_BAR_WIDTH_RT

[0063] LINK_BAR_WIDTH_1-2, HOOK NOT ENGAGED,

[0064] HOOK NOT ENGAGED1-4, SNAP NOT ENGAGED,

[0065] SNAP NOT ENGAGED1-3, LINK_BAR_MISSING,

[0066] LINK_BAR_MISSING1-2, PIN_BROKEN, PIN_BROKEN2-3,

[0067] SC-PIN-BROKEN, SC-PIN-BROKEN1, HOLE BROKEN Detailed Implementation Manner

[0068] The present invention will be described with reference to the accompanying drawings, in which like reference numerals throughout the drawings designate similar or identical elements. The above-mentioned drawings are not drawn to actual scale and merely provide an illustration of the present invention. Some forms of the invention are described below as references for illustrative applications. This means that many specific details, relationships, and methods are set forth to provide a complete understanding of this invention. However, those of ordinary skill in the relevant art will recognize that the present invention can still be implemented without one or more of the specific details or in other ways. By way of other examples, well-known structures or operations are not listed in detail to avoid obscuring the present invention. The present invention is not limited by the order of acts or events described, as some acts may occur in a different order or simultaneously with other acts or events. In addition, not all of the acts or events described need to be performed in the same manner as in the existing invention.

[0069] The method for standardizing image marking according to an embodiment of the present invention is applicable to multiple production lines for producing multiple products. In some embodiments, the products are an electronic device, such as a notebook computer, a smart mobile device, or a tablet computer. In the production line, multiple automated optical inspection (AOI) devices use optical instruments to obtain the surface state (e.g., represented in the form of an image) of the finished or semi-finished products of the products, and then use computer image processing technology to detect defects such as foreign objects or pattern abnormalities. Since automated optical inspection is a non-contact inspection, the semi-finished products of the products can be inspected in the intermediate process. In other words, the method for standardizing image marking of the present invention uses the images captured by the automated optical inspection devices for subsequent processing to achieve the purpose of marking the images as defective images or non-defective images.

[0070] Figure 1 is a flowchart of the method for standardizing image marking according to an embodiment of the present invention. As Figure 1As shown, the method for standardizing image marking according to an embodiment of the present invention first receives a defect setting file, and the defect setting file defines a defect pattern of a defect image (step S100). In some embodiments, the method for standardizing image marking according to an embodiment of the present invention also obtains a specification version of the defect pattern, and the specification version records the update date of the defect pattern. Next, the method for standardizing image marking according to an embodiment of the present invention marks an image based on the defect pattern received in step S100, and generates a first judgment result (step S102). More specifically, in step S102, the method for standardizing image marking according to an embodiment of the present invention receives a first judgment result and a rejudgment result obtained by judging the image according to the defect pattern at different time points. When the rejudgment result is the same as the first judgment result, the image is marked based on the first judgment result and the rejudgment result. The first judgment result and the rejudgment result record whether the image is a defect image. Next, the method for standardizing image marking according to an embodiment of the present invention marks the same image again based on the defect pattern received in step S100, and generates a second judgment result (step S104). More specifically, in step S104, the method for standardizing image marking according to an embodiment of the present invention generates a trained image recognition algorithm using the image as training data, and uses the image recognition algorithm to judge whether the image is a defect image, and obtains the second judgment result. After that, the method for standardizing image marking according to an embodiment of the present invention compares the first judgment result obtained in step S102 with the second judgment result obtained in step S104, and obtains a comparison result (step S106). Finally, according to the comparison result of the second judgment result and the first judgment result (for example, the comparison result is that the second judgment result is different from the first judgment result), the defect pattern is updated (for example, the specification version of the defect pattern is updated), and the defect pattern of the defect image is redefined to standardize the defect pattern (step S108).

[0071] In some embodiments, Figure 1In step S100, the defect patterns (or defect items) in the defect setting file of the image may, for example, include component skew, component breakage, component missing, component open solder, component displacement, foreign objects around the component, through-hole foreign object blockage, residual glue, lack of glue, overflow of glue, assembly misalignment, etc., but the present invention is not limited thereto. In some embodiments, the method for standardizing image marking of the present invention may define different defect patterns according to different production line processes. For example, the above-mentioned component skew, component breakage, component missing, and component open solder are common defect patterns in the surface mount technology (SMT) process. The above-mentioned residual glue, lack of glue, overflow of glue, and assembly misalignment are common defect patterns in the assembly process. In some embodiments, in order to satisfy that the image recognition algorithm in step S108 can replace the original manual judgment after being deployed on the production line, in step S100, the method for standardizing image marking according to an embodiment of the present invention constructs a complete data matrix according to the product defect definition and machine language logic. In some embodiments, the defect setting file includes a plurality of defect patterns and images corresponding to these defect patterns. For example, the defect setting file records the defect patterns of component skew, component breakage, and foreign objects around the component. The defect setting file also records a plurality of images corresponding to component skew, a plurality of images corresponding to component breakage, and a plurality of images corresponding to foreign objects around the component. In some embodiments, the defect setting file includes a plurality of folders, each folder corresponding to a defect pattern respectively, and each folder stores a plurality of images corresponding to the defect pattern.

[0072] For example, taking AOI as an example, if only one detection frame code is used as the basis for classifying the image, for example, using detection frame code A to represent component skew, detection frame code B to represent component missing, and detection frame code C to represent foreign objects around the component, it is often found in practice that the detection frame code does not necessarily match the actual component defect phenomenon, thus resulting in a situation where the actual image (image) does not match the detection frame code, and causing the image recognition algorithm in subsequent step S108 to fail to achieve the expected image recognition level. In some embodiments, the method for standardizing image marking of the present invention combines the defect items of the product with the product specification information to form a visualization file (that is, a defect setting file including various defect patterns and images corresponding to various defect patterns), and uses the above visualization file as the basis for performing subsequent steps.

[0073] In some embodiments, the defect patterns in step S100 are stored in a defect setting file. The defect setting file records the defect items of the product and the product specification information. The specification information of the product can be, for example, the acceptable range of the component appearance length. In some embodiments, the defect setting file also records the updated versions of the defect items. For example, the defect setting file records that the defect item A is added or updated in the defect setting file at a time a, and the defect item A is updated from the original version 1.2 to version 1.3 at time a, and the defect item B is added or updated in the defect setting file at a time b, and the defect item B is updated from the original version 2.0 to 3.0 at time point b. In some embodiments, the defect setting file also records the storage locations of each defect item in the storage space (such as a memory). In some embodiments, the defect pattern in step S100 is defined from the product image captured by the AOI device, but the present invention is not limited thereto. For example, the present invention can also define the defect pattern of the product from the image captured by any image capture device.

[0074] In some embodiments, the method for standardizing image marking in the embodiments of the present invention obtains a specification version of the defect pattern after step S100. In other words, the defect setting file further includes the specification version of the defect pattern, and the specification version records the update or import date of the defect pattern. For example, assume that the defect pattern of component skew is imported or updated in the defect setting file on a first date. Therefore, in the defect setting file, the defect pattern of component skew will be marked as "component skew (first date)". The specification version of the defect pattern in the defect setting file can effectively control the import time of each defect pattern. When the second judgment result determined by the image recognition algorithm in step S106 is different from the first judgment result in step S102, in order to find the image corresponding to the defect pattern that causes the difference between the first judgment result and the second judgment result (i.e., the training data of the image recognition algorithm), to establish a new defect standard for the image corresponding to the defect pattern, that is, to establish a new defect standard for the defect pattern, and finally to standardize the defect pattern.

[0075] In some embodiments, in step S102, the method for standardizing image marking according to the embodiments of the present invention further receives a first judgment result and a re-judgment result determined based on the defect pattern at different time points. In other words, step S102 also examines the consistency of the received first judgment result and re-judgment result. In other words, step S102 also examines the marking ability of an object (such as a computer or a person) that generates the first judgment result and the re-judgment result (based on the same defect standard). When the re-judgment result is different from the first judgment re-judgment result, the method for image marking according to the embodiments of the present invention marks the image. The first judgment result re-judgment records whether the image is a defective image or a non-defective image. In some embodiments, the method for standardizing image marking according to the embodiments of the present invention is that an image marking technician (Annotation Technician: AT) operates a user interface (user interface: UI), so that the computer in the production line can receive the first judgment result and the re-judgment result determined by judging the image according to the defect pattern at different time points. In some embodiments, the image marking technician is only responsible for summarizing the image specifications and executing the image marking, but does not participate in the formulation of the image specifications. Generally speaking, the formulation of the image specifications is carried out by each unit at the factory end according to the product specification and production practice. For example, each unit at the factory end may include product design engineers, process engineers, artificial intelligence R & D engineers, product assembly and testing personnel, supplier management engineers, and quality engineers, but the present invention is not limited thereto. In some embodiments, the method for standardizing image marking according to the present invention is that one or more computers read the defect setting file in step S100, and determine whether the (input) image is a defective image according to the defect pattern defined in the defect setting file.

[0076] When an image marking technician encounters product process specification changes or disputes in determining whether an image is a defective image, the image marking method of the present invention will initiate a mechanism for a marking meeting. The above-mentioned product design engineer, process engineer, artificial intelligence R & D engineer, product assembly and testing personnel, supplier management engineer, and quality engineer will discuss the target feature image in dispute, and establish a specification version number for the meeting result of this marking meeting, so that the marking judgment standard (or marking standard) can be accurately adjusted. In some embodiments, when there is a dispute about the target feature image in the marking meeting, the image marking method of the present invention will perform a correlation analysis of the target feature image (the specification determination is performed by the corresponding factory unit according to the defect to which the target feature image is associated), and the product design engineer or quality engineer participating in the marking meeting will make a final conclusion according to the product specifications. For example, if there is a dispute about an image associated with an SMT component placement defect, the SMT process engineer will perform the specification determination. In some embodiments, the result of the marking meeting will be updated to the defect setting file, so that the defect pattern of the defective image can be synchronously updated.

[0077] In some embodiments, the image marking method of the embodiments of the present invention can update the defect pattern and the specification version of the defect pattern of the defective image by increasing or decreasing the number of images corresponding to at least one defect pattern, or increasing or decreasing the number of types of defect patterns defined in the defect setting file. For example, when an image is determined to be a component skew, the image will be added to the defect pattern of component skew (stored in the folder of the defect pattern of component skew), and the image will be marked with the import date (for example, stored in the date corresponding to the folder of component skew).

[0078] In some embodiments, in step S102, when a first image is determined by an image marking technician to be a defective image at a time point C, the image marking technician inputs the first judgment result into the production line computer through the user interface, so the first judgment result will record that the first image is a defective image. When the same first image is determined to be a non-defective image at a time point D, the rejudgment result will record that the first image is a non-defective image.

[0079] In step S102, when the re-verification result at time point D is the same as the first judgment result at time point C, that is, the first judgment result records that the first image is a defective image, and the re-verification result also records that the first image is a defective image; or the first judgment result records that the first image is a non-defective image, and the re-verification result also records that the first image is a non-defective image, then the image marking standardization method of the embodiment of the present invention performs a marking action on the first image. In some embodiments, the image marking technician marks the first image through the user interface. In other words, the image marking method of the present invention finally determines whether the image (such as the first image) is a defective image or a non-defective image in step S102. In some embodiments, during the process of receiving the re-verification result in step S102 of the image marking method of the present invention, the marking ability of the above-mentioned image marking technician or one or more computers is evaluated.

[0080] For example, an image marking technician A determines that a second image is a defective image at time point C according to the defective pattern defined in step S100. The image marking technician inputs the judgment result of "the second image is a defective image" into the production line computer through the user interface. Then, in step S102, the image marking technician A re-verifies the same second image, and the image marking technician inputs the re-verification result of the second image into the production line computer through the user interface again. In the re-verification result, when the second image is also a defective image, the image marking standardization method of the embodiment of the present invention determines that the marking ability of the image marking technician A is qualified, and the image marking technician A will be allowed to continue to stay on the production line. On the contrary, when the image marking technician A determines that the second image is a non-defective image in step S102, that is, the first judgment result in step S102 is different from the re-verification result, the image marking standardization method of the embodiment of the present invention determines that the image marking technician A is unqualified and asks the image marking technician A to temporarily leave the production line.

[0081] Figure 2 It is a schematic diagram of a user interface (UI) of the image marking standardization method according to an embodiment of the present invention. The image marking technician A in the above embodiment operates a user interface as Figure 2 to mark the image as the defective image or the non-defective image, or to delete the image. As Figure 2As shown, the user interface 200 can be displayed on the display screens connected to one or more computers in the production line. In some embodiments, the user interface 200 can also be displayed on a mobile device carried by the image marking technician to mark different images at any time. The user interface 200 includes an image display 202, image marking objects 204 and 206, an image deletion object 208, an image marking display 210, an image opening object 212, a marking completion object 214, a marking progress display 216, a previous page object 218, and a next page object 220. In some embodiments, the image display 202 is used to display images (such as the above-mentioned first image, second image, and third image). The image marking technician observes the defect pattern of the image through the image display 202, so that the image marking technician can mark the image according to the defect pattern of the image.

[0082] In some embodiments, the image marking object 204 is used to mark the image being displayed in real time in the image display 202 as a defective image, while the image marking object 206 is used to mark the image being displayed in real time in the image display 202 as a non-defective image. When the image marking technician presses the image marking objects 204 and 206 in the user interface 200, the image being displayed in real time in the image display 202 will be automatically stored in a storage space. For example, when a fourth image is displayed in the image display 202 and the image marking technician presses the image marking object 204, the fourth image will be stored in a defective folder. When the image marking technician presses the image marking object 206, the fourth image will be stored in a non-defective folder.

[0083] The image deletion object 208 is used to delete the images being displayed in real time in the image display 202. For example, if the image marking technician determines that the image being displayed in real time in the image display 202 does not belong to the product produced on this production line, the image marking technician can delete the image being displayed in real time in the image display 202 through the image deletion object 208. The image marking display 210 can display the statistical number of real-time defective images or non-defective images. In some embodiments, the image opening object 212 can read the images in the storage space and display the read images in the image display 202. When the image marking technician has completed the marking of all images, the marking process can be ended through the marking completion object 214. In some embodiments, the marking progress display 216 can display the current marking progress. For example, the image marking technician reads multiple images from a storage space (such as a folder) through the image opening object 212. The marking progress display 216 can display the number of images in the read folder and the number of images for which the marking is not yet completed. In some embodiments, the image marking technician can use the previous page object 218 and the next page object 220 to switch the images being displayed in real time in the image display 202. The layout settings and functions of the user interface 200 of the image marking method according to the embodiments of the present invention are only illustrative and do not limit the present invention.

[0084] In some embodiments, the image marking method of the present invention also audits the marking results of the images. When the marking operation of the images is performed by an image marking technician, the marking results of the same image by the image marking technician may fluctuate due to human factors, or the images themselves have highly controversial characteristics. If highly controversial images are used as the training data for the image recognition algorithm in step S104, the judgment accuracy of the image recognition algorithm will be affected. Therefore, the image marking method of the present invention audits the marking results of each image marking technician.

[0085] For example, the image marking method according to an embodiment of the present invention audits the marking results of image marking technician A. First, image marking technician A determines whether a fifth image among a plurality of images is a defective image (first determination) (e.g., at time point E), and outputs the determination result to the production line computer through the user interface. The production line computer receives the control signal of the user interface and obtains a first audit result. Then, image marking technician A (e.g., at time point F) determines again whether the fifth image among the plurality of images is a defective image (second determination), and outputs the determination result to the production line computer through the user interface. The production line computer receives the control signal of the user interface and obtains a second audit result. In other words, the image marking method of the present invention allows the same image marking technician to determine the same image at different time points and compares whether the two audit results are the same. In some embodiments, the first audit result and the second audit result record the number of defective images and non-defective images included in the same set of a plurality of images at different time points (e.g., time points E and F). When the ratio (moving rate) of the first audit result being different from the second audit result is lower than a first threshold, the image marking method of the present invention determines that the marking result of image marking technician A is qualified. In some embodiments, the first threshold is 20%. In some embodiments, the first threshold can be dynamically adjusted according to the comparison result of the first determination result in different steps S102 and the second determination result in step S104 in previous tests (i.e., comparing whether the first determination result is the same as the second determination result in step S106) to effectively control the marking ability of the same image marking technician for the same image.

[0086] For example, the image marking standardization method according to an embodiment of the present invention audits 5,643 images marked by image marking technician A. According to the above audit steps, the image marking standardization method of the present invention finds that 508 images are non-defective images in the first determination and defective images in the second determination. In addition, 1,063 images are defective images in the first determination and non-defective images in the second determination. Therefore, the ratio of the first audit result being different from the second audit result is (508 + 1,063) / 5,643 = 27.84%. Since 27.84% is greater than the first threshold (e.g., 20%), the image marking standardization method according to the embodiment of the present invention determines that the marking result of image marking technician A is unqualified. In some embodiments, when the marking result of image marking technician A is unqualified, image marking technician A must first leave the production line until the moving rate of the first audit result and the second audit result he tested is lower than 20%.

[0087] In some embodiments, the image marking standardization method of the present invention enables different image marking technicians to judge the same multiple images, so as to audit the marking results of different image marking technicians. For example, technician A of image marking judges whether a sixth image among multiple images is a defective image, and outputs his judgment result to the production line computer through the user interface. The production line computer receives the control signal of the user interface and obtains a third audit result. Technician B of image marking judges whether the sixth image among multiple images is a defective image, and outputs his judgment result to the production line computer through the user interface. The production line computer receives the control signal of the user interface and obtains a fourth audit result. The image marking standardization method of the embodiments of the present invention compares the third audit result and the fourth audit result. When the proportion that the third audit result is different from the fourth audit result is lower than a second threshold, the image marking method of the present invention determines that the marking results of technicians A and B of image marking are qualified. In some embodiments, the second threshold is 5%. In some embodiments, the second threshold can be dynamically adjusted according to the comparison result of the first judgment result in different step S102 and the second judgment result in step S104 in the previous test (that is, comparing whether the first judgment result is the same as the second judgment result in step S106), so as to effectively control the marking ability of different image marking technicians for the same image.

[0088] Next, the image marking method of the present invention uses images (such as the above-mentioned first image, second image, third image, fourth image, fifth image, and sixth image after being marked) as training data to generate a trained image recognition algorithm. In some embodiments, the image recognition algorithm is a Convolution Neural Network (CNN). In some embodiments, the Convolution Neural Network is also a Deep Residual Network (RestNet). Figure 3 It is a schematic structural diagram of the Deep Residual Network of the image marking standardization method of the embodiments of the present invention. As Figure 3 shown, the Deep Residual Network 300 includes a convolutional layer 302, a residual layer 304, a global pooling layer 306, and a dense layer 308. In some embodiments, the convolutional layer 302 performs feature extraction on the input image 310 with a feature filter of 7*7 size. Since the size of the feature filter of the convolutional layer 302 is relatively large (such as 7*7), the convolutional layer 302 can extract the local features of the image 310 and input the local features of the image 310 into the residual layer 304. In some embodiments, the residual layer 304 is formed by overlapping and connecting in series multiple convolutional blocks and multiple identity blocks. In some embodiments, the image 310 can be a square image with a size of 224*224 or a rectangular image with a size of 336*112.

[0089] Figure 4 Schematic diagram of the convolution block 400 and the identify block 410 of the deep residual network of the image marking standardization method according to an embodiment of the present invention. As Figure 4 shown, the convolution block 400 includes a convolutional layer 402, a convolutional layer 404, a convolutional layer 406, and a convolutional layer 408. The convolutional layers 402, 406, 408 have feature filters of size 1*1, and the convolutional layer 404 has a feature filter of size 3*3 for extracting detailed features of the feature image 420. In some embodiments, the feature image 420 is the feature extracted from Figure 3 the image 310. The feature image 420 undergoes feature extraction by the convolutional layers 402, 404, 406 to generate the feature image 422. On the other hand, the feature image 420 directly undergoes feature extraction by the convolutional layer 408 to generate the feature image 424. Then, the convolution block 400 superimposes the feature image 422 and the feature image 424 to generate the feature image 426, and outputs the feature image to the next layer.

[0090] The identify block 410 includes a convolutional layer 412, a convolutional layer 414, and a convolutional layer 416. The convolutional layers 412, 416 have feature filters of size 1*1, and the convolutional layer 414 has a feature filter of size 3*3. In some embodiments, the feature image 430 is the feature extracted from Figure 3 the image 310. In some embodiments, the feature image 430 undergoes feature extraction by the convolutional layers 412, 414, 416 to generate the feature image 432. The identify block 410 outputs the feature image 432 to the next layer. In some embodiments, the feature image 432 serves as the final output of the residual layer 304 and is output to Figure 3 the global pooling layer 306. When the image 310 is a square image of 224*224, the global pooling layer 306 describes the image 310 with 1344 feature image values. When the image 310 is a rectangular image of 336*112, the global pooling layer 306 describes the image 310 with 2048 feature image values. In some embodiments, when the image 310 is a square image of 224*224, the fully connected layer 308 assembles 1344 feature image values into the values corresponding to the defective image and the non-defective image to obtain the value 312.

[0091] Finally, the image marking standardization method of the present invention converts the value 312 into a probability using the softmax function. For example, the value 312 can be represented as (-2.6, 3.1), where -2.6 is the value when the deep residual network 300 determines that the image 310 is a non-defective image, and 3.1 is the value when the deep residual network 300 determines that the image 310 is a defective image. The image marking standardization method of the present invention converts the value 312 into a probability (0.003, 0.997) through the softmax function. In other words, the probability that the deep residual network 300 determines that the image 310 is a non-defective image is 0.3%, and the probability that it determines that the image 310 is a defective image is 99.7%. Therefore, the deep residual network 300 determines that the image 310 is a defective image. The softmax function is a generalization of logistic regression and is thus applicable to multi-class classification problems, such as classifying an image as a defective image or a non-defective image.

[0092] Back to Figure 1 , when the image marking standardization method of the present invention uses an image (such as the first, second, third, fourth, fifth, sixth images marked above, or Figure 3 the image 310) as training data to generate a trained convolutional neural network (such as the deep residual network 300), in step S104, the image marking standardization method of the embodiment of the present invention then uses the trained image recognition algorithm (such as a convolutional neural network) to determine whether the same image is a defective image, and obtains a second judgment result. The image marking standardization method of the embodiment of the present invention then compares the first judgment result in step S102 and the second judgment result in step S104 in step S106, and in step S108, based on the first judgment result in step S102 and the second judgment result in step S104, updates the defect setting file defined in step S100 to update the specification version of the defect pattern and redefine the defect pattern of the defective image, so that the defect pattern can be standardized and have a consistent standard. It should be noted that the defect pattern of the defect setting file received in step S100, the first judgment result received in step S102, and the comparison result of the first judgment result and the second judgment result in step S106 are all generated by the same object according to the same defect standard.

[0093] For example, during a marking meeting, image labeling technician B confirms the defect patterns and defect criteria for an image and enters the defect labels into the production line computer in the form of a defect configuration file via the user interface. After passing the marking competency assessment (i.e., the first judgment result in step S102 is the same as the re-judgment result), image labeling technician B then judges an image based on the defect criteria from the marking meeting and enters the judgment result into the production line computer via the user interface, generating the first judgment result in step S102. In step S108, image labeling technician B then confirms the comparison of the first and second judgment results from step S106 using the same defect criteria. Therefore, because the defect configuration file in step S100, the first judgment result in step S102, and the comparison of the first and second judgment results in step S106 are all generated by image labeling technician B based on the defect criteria established during the marking meeting, the labeling stability of the image recognition algorithm's training data is improved, thereby enhancing the image recognition accuracy of the trained image recognition algorithm.

[0094] For example, a seventh image is judged as a defective image in step S102, but is judged as a non-defective image in step S104. Therefore, the image marking standardization method of an embodiment of the present invention must correct the specification settings for defect patterns in the defect setting file. For example, the image marking standardization method of an embodiment of the present invention may increase or decrease the number of images corresponding to at least one defect pattern, or increase or decrease the number of types of defect patterns defined in the defect setting file. In some embodiments, the seventh image is judged as a defective image in step S102, and is also judged as a defective image in step S104. Therefore, the image marking standardization method of the present invention determines that the image recognition accuracy of the image recognition algorithm (such as a convolutional neural network, a deep residual network 300) generated in step S104 is applicable to the actual production line. For example, the trained convolutional neural network can be transmitted to multiple computers in the production line through the network.

[0095] Figure 5A 、 Figure 5B and Figure 5C This is a flow chart for verifying the convolutional neural network of the image labeling standardization method according to an embodiment of the present invention. In some embodiments, when the convolutional neural network generated in step S104 is introduced into the actual production line, the products produced by the production line have already entered the mass production stage. Therefore, when verifying the convolutional neural network (such as the deep residual network 300) with images, the original operation mode of the production line must not be affected, and no interference should be applied until the image labeling standardization method of the present invention completes the verification of the convolutional neural network. For example, Figure 5A This is a verification method that uses fully manual assisted judgment.

[0096] likeFigure 5A As shown, in step S500, the image marking method of the present invention captures an image of the products produced in the production line with an automatic optical inspection (AOI) device. Then, in step S502, the AOI device determines whether the captured image is a defective image. When the AOI device determines that the image is a non-defective image, the image marking standardization method of the present invention directly transfers the product corresponding to the non-defective image to the next test station (e.g., through the production line conveyor belt) (step S508). When the AOI device determines that the image is a defective image, the image marking technician performs a secondary review on the image in step S504.

[0097] If the image marking technician confirms that the image is a defective image, the image marking technician inputs his / her judgment result into the production line computer through the user interface, so that the production line computer receives the control signal from the user interface and confirms that the image is a defective image (step S506). Then, the image marking standardization method of the embodiment of the present invention can pick out the product corresponding to the defective image and repair its defective state. If the image marking technician confirms that the image is a non-defective image, the image marking technician inputs his / her judgment result into the production line computer through the user interface, so that the production line computer receives the control signal from the user interface and confirms that the image is a non-defective image, and then enters step S508. The image marking standardization method of the present invention directly transfers the product corresponding to the non-defective image to the next test station. Figure 5A The verification method shown is the method before the introduction of the convolutional neural network for auxiliary verification.

[0098] Figure 5B It is the verification method when the convolutional neural network is introduced for auxiliary verification. As Figure 5B As shown, in step S510, the image marking method of the present invention captures an image of the products produced in the production line with the AOI device. Then, in step S512, the AOI device determines whether the captured image is a defective image. When the AOI device determines that the image is a non-defective image, the image marking standardization method of the present invention directly transfers the product corresponding to the non-defective image to the next test station (step S520). When the AOI device determines that the image is a defective image, the convolutional neural network (CNN) (which is a type of artificial intelligence (AI)) performs a secondary review on the image in step S514.

[0099] Since the correctness of the image recognition of the convolutional neural network has not been confirmed in step S514, therefore Figure 5BAmong them, regardless of the re-verification result of the convolutional neural network, it is necessary to proceed to step S516, where the image marking technician re-verifies the image again. If the image marking technician still confirms that the image is a defective image, the image marking technician inputs his judgment result into the production line computer through the user interface, so that the production line computer receives the control signal of the user interface and confirms that the image is a defective image (step S518). Then, the image marking standardization method of the embodiment of the present invention can pick out the product corresponding to the defective image and repair its defective state. If the image marking technician confirms that the image is a non-defective image, the image marking technician inputs his judgment result into the production line computer through the user interface, so that the production line computer receives the control signal of the user interface and confirms that the image is a non-defective image, then it enters step S520. The image marking standardization method of the embodiment of the present invention directly transfers the product corresponding to the non-defective image to the next test station.

[0100] Figure 5C It is a verification method when formally verifying the import of the convolutional neural network. In Figure 5C Among them, the image marking method of the present invention completely replaces manual verification with a trained convolutional neural network. As Figure 5C shown, in step S530, the image marking standardization method of the present invention uses an AOI device to capture images of the products produced in the production line. Then, in step S532, the AOI device determines whether the captured image is a defective image. When the AOI device determines that the image is a non-defective image, the image marking method of the present invention directly transfers the product corresponding to the non-defective image to the next test station (step S538). When the AOI device determines that the image is a defective image, formal verification is performed with the convolutional neural network, and the convolutional neural network performs a secondary re-verification on the image in step S534. Figure 5C The difference between step S534 of Figure 5B and step S514 of Figure 5B is that in step S514, since the correctness of the image recognition of the convolutional neural network has not been confirmed, therefore in Figure 5B among them, regardless of the re-verification result of the convolutional neural network, it is necessary to proceed to step S516, where the image marking technician re-verifies the image again. And in step S534, the correctness of the image recognition of the convolutional neural network has been verified, so the convolutional neural network can officially perform the work of image recognition.

[0101] If the convolutional neural network confirms that the image is a defective image (step S536), then the image marking standardization method of the present invention can pick out the product corresponding to the defective image and repair its defective state. If the convolutional neural network confirms that the image is a non-defective image, the image marking standardization method of the present invention directly transfers the product corresponding to the non-defective image to the next test station for subsequent testing.

[0102] Figure 6 This is a setting diagram when the image marking standardization method of the embodiment of the present invention is applied to a production line. As Figure 6 shown, a computer 604, a router 606, an AOI device 608, a computer 610, and an edge computing device 612 are set in the production line 600. The computer 604 is coupled between a company intranet 602 and the router 606, and the router 606 is coupled between the computer 604 and the AOI device 608, between the computer 604 and the computer 610, and between the computer 604 and the edge computing device 612. The convolutional neural network (such as the deep residual network 300) of the image marking standardization method of the present invention is executed on the computer 610 and the edge computing device 612. The computer 610 and the edge computing device 612 obtain the images of the product captured by the AOI device through the router 606.

[0103] For example, when the computer 610 and the edge computing device 612 receive an eighth image from the AOI device 608 through the router 606, the computer 610 determines whether the eighth image is a defective image, and the edge computing device 612 performs a re-judgment on the eighth image. In some embodiments, the edge computing device 612 transmits a re-judgment result of the eighth image to the computer 610 through the router 606, and the computer 610 marks the eighth image as a defective image or a non-defective image according to its own judgment result of the eighth image and the judgment result of the edge computing device 612. In addition, the computer 604 can, through a method of edge computing, summarize the marking results of the computer 610 and the edge computing device 612 on the eighth image to a server (not shown) through the company intranet 602. In other words, the defective states of the products produced by each production line can be stored in the server (for example, stored in a defect setting file). The Figure 6 production line setting of the present invention is only for illustration and is not a limitation of the present invention. The present invention can also use other production line setting methods to achieve Figure 1 the effects of steps S100 to S108.

[0104] The present invention also discloses an electronic device for image marking standardization. The electronic device includes a user interface and a processor. In some embodiments, the user interface is used to receive Figure 1 a defective state in step S100. In addition, the user interface can receive Figure 1 a first judgment result generated by marking an image according to the defective state in step S102. In some embodiments, the processor can receive the defective state from the user interface and also receive the first judgment result from the user interface. The processor can mark the image according to the defective state to generate Figure 1A second determination result in step S104. The processor then compares the first determination result and the second determination result to obtain Figure 1 A comparison result in step S106. Finally, the processor updates the defect pattern according to the comparison result to standardize the defect pattern (which can correspond to Figure 1 Step S108).

[0105] The present invention also discloses a computer program product for image marking standardization. The above computer program product is applicable to a computer having a processor. The above computer program product includes a first receiving instruction, a second receiving instruction, a determination instruction, a comparison instruction, and an update instruction. When the processor of the computer executes the first receiving instruction, the processor receives Figure 1 A defect pattern in step S100. When the processor executes the second receiving instruction, the processor receives Figure 1 A first determination result generated by marking an image according to the defect pattern in step S102. When the processor executes the determination instruction, the processor marks the image according to the defect pattern to generate Figure 1 A second determination result in step S104. When the processor executes the comparison instruction, the processor compares the first determination result and the second determination result to obtain Figure 1 A comparison result in step S106. When the processor executes the update instruction, the processor updates the defect pattern according to the comparison result to standardize the defect pattern (which can correspond to Figure 1 Step S108).

[0106] When the image marking standardization method of the present invention is applied to an actual production line, it can effectively reduce the overkill rate and leak rate during image recognition. In some embodiments, the overkill rate represents the probability of judging an image that should be a non-defective image as a defective image. The leak rate represents the probability of judging an image that should be a defective image as a non-defective image. Figure 7 It is a curve graph of the overkill rate and re-judgment moving rate of the product after the image marking standardization method of the embodiment of the present invention is applied to the production line. As Figure 7As shown, curve 700 is the rejudgment moving rate after the image marking method of the present invention is applied to the production line. Curve 702 is the product overkill rate after the image marking method of the present invention is applied to the production line. The time point when the image marking standardization method of the present invention is applied to the actual production line is t1. According to curves 700 and 702, it can be clearly observed that after time point t1, the rejudgment moving rate drops to 17.10%, and the product overkill rate drops to 2.44%, and the rejudgment moving rate and the product overkill rate continue to decline after time point t1. In other words, the image marking standardization method of the present invention can effectively improve the overkill rate, omission rate, and rejudgment moving rate when the production line produces products, making the production efficiency of the production line more stable. The image marking standardization method of the present invention can improve the marking stability of the training data of the trained convolutional neural network, thereby improving the image recognition accuracy of the trained convolutional neural network.

[0107] Figure 8A and Figure 8B is a schematic diagram of the defective state of the image marking standardization method of the embodiment of the present invention. Referring also to Figure 6 , Figure 8A and Figure 8B , Figure 6 The computer 610 receives a total of 12 defective states belonging to the AOI test station, including state one, surface flatness; state two, poor clearance of the flat connecting rod; state three, missing pins; state four, deformed hooks; state five, damaged hooks; state six, damaged buckles; state seven, poor width of the flat connecting rod; state eight, the hook not being properly engaged; state nine, the buckle not being properly engaged; state ten, missing flat connecting rods; state eleven, damaged pins; and state twelve, damaged corners or holes blocked. The convolutional neural network (such as the deep residual network 300) of the image marking standardization method of the present invention is executed on the computer 610 and the edge computing device 612.

[0108] For example, when the computer 610 determines that the image from the AOI device 608 conforms to the defects of state one and state two, the computer 610 will generate machine error codes "FP Flatness" and "Link Bar Gap", and the above machine error codes are the same as the pre-set specification table codes. Since state one and state two are set to not require rejudgment, there is no need for rejudgment by the edge computing device 612, and the image marking standardization method of the present invention directly adopts the determination result of the computer 610. In some embodiments, when the computer 610 determines that the image from the AOI device 608 conforms to the defect of state three, the computer 610 will generate the machine error code "SC Pin Missing", and on the display device of the computer 610 (such as Figure 2In the user interface 200, defect sub-mode types are displayed, such as a total of 10 defect sub-mode types like "PIN-MISSING 1-10". Since mode three is set to require re-verification, it still needs to go through the re-verification of the edge computing device 612, and when the determination results of the computer 610 and the edge computing device 612 are the same, the image marking standardization method of the present invention adopts the determination result of the computer 610.

[0109] Similarly, when the computer 610 determines that the image from the AOI device 608 conforms to the defects of mode four, the computer 610 will generate a machine error code of "FP Hook deformation", and will display the defect sub-mode types in its display device, such as a total of 12 defect sub-mode types like "FP_DEFORMATION", "FP_DEFORMATION2-6", "FP_WARP", "FP_WARP1-5", etc. When the computer 610 determines that the image from the AOI device 608 conforms to the defects of mode five, the computer 610 will generate a machine error code of "Hook Broken", and will display the defect sub-mode types in its display device, such as a total of 7 defect sub-mode types like "HK_BROKEN1-7", etc.

[0110] In some embodiments, when the computer 610 determines that the image from the AOI device 608 conforms to the defects of mode six, the computer 610 will generate a machine error code of "Snap Broken", and will display the defect sub-mode types in its display device, such as a total of 8 defect sub-mode types like "SNAP-BROKEN", "SNAP-BROKEN1-7", etc. When the computer 610 determines that the image from the AOI device 608 conforms to the defects of mode seven, the computer 610 will generate a machine error code of "Link Bar Width Fail", and will display the defect sub-mode types in its display device, such as a total of 6 defect sub-mode types like "LINK_BAR_WIDTH_RB", "LINK_BAR_WIDTH_LT", "LINK_BAR_WIDTH_LB", "LINK_BAR_WIDTH_RT", "LINK_BAR_WIDTH_1-2", etc.

[0111] In some embodiments, when computer 610 determines that the image from AOI device 608 conforms to the defects of pattern eight, computer 610 will generate a machine error code of "Hook not engaged", and will display the defect sub-pattern types in its display device, such as a total of 5 defect sub-patterns including "HOOK NOT ENGAGE", "HOOK NOT ENGAGE1-4", etc. When computer 610 determines that the image from AOI device 608 conforms to the defects of pattern nine, computer 610 will generate a machine error code of "Snap not engaged", and will display the defect sub-pattern types in its display device, such as a total of 4 defect sub-patterns including "SNAP NOT ENGAGE", "SNAP NOT ENGAGE1-3", etc. When computer 610 determines that the image from AOI device 608 conforms to the defects of pattern ten, computer 610 will generate a machine error code of "Link Bar Missing", and will display the defect sub-pattern types in its display device, such as a total of 3 defect sub-patterns including "LINK_BAR_MISSING", "LINK_BAR_MISSING1-2", etc.

[0112] In some embodiments, when computer 610 determines that the image from AOI device 608 conforms to the defects of pattern eleven, computer 610 will generate a machine error code of "SC Pin Broken", and will display the defect sub-pattern types in its display device, such as a total of 5 defect sub-patterns including "PIN-BROKEN", "PIN-BROKEN2-3", "SC-PIN-BROKEN", "SC-PIN-BROKEN1", etc. When computer 610 determines that the image from AOI device 608 conforms to the defects of pattern twelve, computer 610 will generate a machine error code of "FP HOLE BROKEN", and will display the defect sub-pattern type in its display device, such as the defect sub-pattern of "HOLE BROKEN". The image marking standardization method of the present invention can be obtained through Figure 8A and Figure 8B the information displayed in the test window to know the actual defect pattern of the image from AOI device 608.

[0113] Although the embodiments of the present invention are described as above, we should understand that the above-presented are only examples, not limitations. Many changes based on the above exemplary embodiments of the present invention can be implemented without violating the spirit and scope of the present invention. Therefore, the breadth and scope of the present invention should not be limited by the embodiments described above. More precisely, the scope of the present invention should be defined by the scope of the appended claims and their equivalents.

[0114] Although the foregoing invention has been illustrated and described with reference to one or more related implementations, equivalent changes and modifications will occur to those skilled in the art in light of the foregoing specification and drawings. In addition, although a particular feature of the invention has been illustrated in one of the related implementations, the foregoing feature may be combined with one or more other features, such that it may be desirable and useful in any known or particular application.

[0115] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. We should be more aware that the foregoing terms, as defined in a commonly used dictionary, should be interpreted as having the same meaning in the context of the relevant art. Unless clearly defined herein, the foregoing terms are not to be construed as idealized or overly formal.

Claims

1. A method for image marking standardization, the method for image marking standardization comprising: Receiving a defect pattern; Marking an image according to the defect pattern to generate a first judgment result; Marking the image according to the defect pattern to generate a second judgment result; Comparing the first judgment result and the second judgment result to obtain a comparison result; And Updating the defect pattern according to the comparison result to standardize the defect pattern; The method for image marking standardization further comprises: Receiving a first audit result and a second audit result at different time points; wherein, the first audit result and the second audit result record the number of defective images and non-defective images included in the same group of multiple images at different time points; and When the ratio of the first audit result being different from the second audit result is lower than a first threshold, determining that the marking result is qualified; Wherein, the step of generating the second judgment result comprises: Generating a trained image recognition algorithm using the image as training data; and Using the image recognition algorithm to determine whether the image is a defective image to obtain the second judgment result; Wherein, the step of updating the defect pattern according to the comparison result to standardize the defect pattern comprises: Updating the defect setting file according to the comparison result, updating the specification version of the defect pattern, and redefining the defect pattern of the defective image to standardize the defect pattern; Wherein, the defect setting file includes the defect pattern.

2. The method for standardizing image markings as described in claim 1, wherein, The step of marking the image comprises: receiving a control signal from a user interface, and according to the control signal, marking the image as a defective image or a non-defective image, or deleting the image.

3. The method for standardizing image markers as described in claim 2, wherein, The step of marking the image further comprises: receiving the control signal from the user interface, and storing the image in a corresponding storage space according to the control signal.

4. The method for standardizing image markers as described in claim 1, wherein, The step of generating the first judgment result comprises: Receiving the first judgment result and a recheck result judged according to the defect pattern for the image at different time points; and When the recheck result is the same as the first judgment result, marking the image according to the first judgment result; Wherein, the first judgment result records whether the image is a defective image.

5. The method for image marking standardization according to claim 1, the method for image marking standardization further comprises: Obtaining the specification version of the defect pattern; wherein, the specification version records the update date of the defect pattern.

6. The method for standardizing image markers as described in claim 5, wherein, The step of updating the specification version of the defect pattern comprises: Increasing or decreasing the number of images corresponding to at least one of the defect patterns; or Increasing or decreasing the number of types of the defect patterns defined in the defect setting file.

7. The method for image marking standardization according to claim 1, the method for image marking standardization further comprises: When the second judgment result is the same as the first judgment result, transmitting the trained convolutional neural network to multiple computers in the production line through a network.

8. The method for standardizing image markers as claimed in claim 7, wherein, The convolutional neural network is a deep residual network.

9. An electronic device for image marking standardization, the electronic device comprising: A user interface that receives a defect pattern and a first judgment result generated by marking an image according to the defect pattern; And A processor that receives the defect pattern and the first judgment result from the user interface, and marks the image according to the defect pattern to generate a second judgment result; Wherein, the processor compares the first judgment result and the second judgment result to obtain a comparison result; the processor updates the defect pattern according to the comparison result to standardize the defect pattern; Wherein, the processor receives a first verification result and a second verification result at different time points. The first verification result and the second verification result record the number of defective images and non-defective images included in the same group of multiple images at different time points; and When the proportion that the first verification result is different from the second verification result is lower than a first threshold, the processor determines that the marking result is qualified; Wherein, the processor uses the image as training data to generate a trained image recognition algorithm, and uses the image recognition algorithm to determine whether the image is a defective image to obtain the second judgment result; Wherein, the processor updates the defect setting file according to the comparison result, updates the specification version of the defect pattern, and redefines the defect pattern of the defective image to standardize the defect pattern; Wherein, the defect setting file includes the defect pattern.

10. A computer program product for image marking standardization and applicable to a computer with a processor. The computer program product includes: A first receiving instruction that enables the processor to receive a defect pattern; A second receiving instruction that enables the processor to receive a first judgment result generated by marking an image according to the defect pattern; A judgment instruction that enables the processor to mark the image according to the defect pattern to generate a second judgment result; A comparison instruction that enables the processor to compare the first judgment result and the second judgment result to obtain a comparison result; And An update instruction that enables the processor to update the defect pattern according to the comparison result to standardize the defect pattern; Wherein, the processor receives a first verification result and a second verification result at different time points. The first verification result and the second verification result record the number of defective images and non-defective images included in the same group of multiple images at different time points; and When the proportion that the first verification result is different from the second verification result is lower than a first threshold, the processor determines that the marking result is qualified; Wherein, the processor uses the image as training data to generate a trained image recognition algorithm, and uses the image recognition algorithm to determine whether the image is a defective image to obtain the second judgment result; Wherein, the processor updates the defect setting file according to the comparison result, updates the specification version of the defect pattern, and redefines the defect pattern of the defective image to standardize the defect pattern; Among them, the defect setting file includes the defect pattern.

Images