TWI703514B - Artificial intelligence recheck system and method thereof - Google Patents

Artificial intelligence recheck system and method thereof Download PDF

Info

Publication number
TWI703514B
TWI703514B TW107117692A TW107117692A TWI703514B TW I703514 B TWI703514 B TW I703514B TW 107117692 A TW107117692 A TW 107117692A TW 107117692 A TW107117692 A TW 107117692A TW I703514 B TWI703514 B TW I703514B
Authority
TW
Taiwan
Prior art keywords
artificial intelligence
result
unit
results
manual
Prior art date
Application number
TW107117692A
Other languages
Chinese (zh)
Other versions
TW202004574A (en
Inventor
汪光夏
呂彥德
何羽立
Original Assignee
牧德科技股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 牧德科技股份有限公司 filed Critical 牧德科技股份有限公司
Priority to TW107117692A priority Critical patent/TWI703514B/en
Priority to CN201810635724.1A priority patent/CN109060817B/en
Publication of TW202004574A publication Critical patent/TW202004574A/en
Application granted granted Critical
Publication of TWI703514B publication Critical patent/TWI703514B/en

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G01N2021/8883—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges involving the calculation of gauges, generating models
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G01N2021/8887—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges based on image processing techniques

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
  • Image Analysis (AREA)

Abstract

本發明在提出一種人工智慧複檢系統，可在光學檢測裝置判斷NG影像或數量資料後，先經由人工智慧(AI)單元模擬預測人工複檢的結果，再傳送至人工複檢站進行實際的人工複檢，藉此減少人工複檢時所需再次檢測的數量，並提升操作人員工作效率。 The present invention proposes an artificial intelligence re-inspection system. After the optical detection device judges NG images or quantitative data, the artificial intelligence (AI) unit simulates and predicts the artificial re-inspection result, and then transmits it to the manual re-inspection station for actual Manual re-inspection, thereby reducing the number of re-inspection required during manual re-inspection, and improving the work efficiency of operators.

Description

人工智慧複檢系統及其方法 Artificial intelligence review system and method

本發明係關於一種用於光學檢測裝置的人工智慧複檢系統及其方法，更特別的是關於一種能模擬預測人工複檢結果，使得操作人員在實際複檢時能夠減少複檢除錯的數量，並藉此提升人工複檢工作效率。 The present invention relates to an artificial intelligence re-inspection system and method for optical inspection devices, and more particularly to an artificial re-inspection system that can simulate and predict the results of manual re-inspection, so that operators can reduce the number of re-inspection and debugging during actual re-inspection , And to improve the efficiency of manual re-inspection.

光學辨識系統如自動光學檢測機(Automated Optical Inspection,AOI)及外觀終檢機(Automatic Final Inspection,AFI)等檢測機台，如今已經被普遍應用在電子業之電路板組裝生產線上的檢測流程中，用以取代以往的人工目測檢視作業，它利用影像技術比對待測物與標準影像是否有差異來判斷待測物有否符合標準，自動光學檢測設備除了提供高穩定度的檢測結果外，還大幅節省了檢測時間及人力成本。 Optical identification systems such as Automatic Optical Inspection (AOI) and Automatic Final Inspection (AFI) and other inspection machines are now widely used in the inspection process of circuit board assembly lines in the electronics industry , To replace the previous manual visual inspection operation, it uses imaging technology to compare whether the object to be tested is different from the standard image to determine whether the object to be tested meets the standard. In addition to providing high-stability detection results, the automatic optical inspection equipment also Significantly save testing time and labor costs.

然而，習知之光學檢測流程如第1圖所示，待測物(包括但不限於印刷電路板、軟板等)經由光學檢測裝置100檢測後，需再將光學檢測裝置之檢測結果及可能存在缺陷的待複檢物(原待測物的全部或一部分)透過運送單元200送至人工視覺複檢站300，經人工複檢辨識後確認是否為真的缺陷或僅是光學檢測機台的誤判，進而操作人員再將待複檢物進行標識或是修復的動作，故光學檢測裝置100判斷缺陷的能力優劣高度相關於複檢站300所需投入的複檢人員數量。 However, the conventional optical inspection process is shown in Figure 1. After the object to be tested (including but not limited to printed circuit boards, flexible boards, etc.) is detected by the optical inspection device 100, the inspection results of the optical inspection device and the possible existence The defective object to be re-inspected (all or part of the original object to be tested) is sent to the manual visual re-inspection station 300 through the transport unit 200, and after manual re-inspection and identification, it is confirmed whether it is a real defect or just a misjudgment by the optical inspection machine , And the operator then marks or repairs the object to be re-inspected, so the ability of the optical inspection device 100 to determine defects is highly related to the number of re-inspectors required by the re-inspection station 300.

假若光學檢測裝置判斷報點的靈敏度過高，則後端需投入的複檢人員勢必 增加，此舉不僅降低了企業生產效率亦同時提高了企業營運成本；而若是將光學檢測裝置判斷報點的靈敏度降低，則又容易導致瑕疵漏檢影嚮了後續產品的生產。故如何在保持光學檢測裝置判斷報點相對靈敏度的前提下，還能夠減少複檢人員需複檢的瑕疵數量，此實為業界亟待解決的問題，本發明案即為此提出解決方案。 If the optical detection device judges that the sensitivity of the reporting point is too high, the re-inspectors who need to invest in the back-end are bound to Increasing, this not only reduces the production efficiency of the enterprise, but also increases the operating cost of the enterprise; and if the sensitivity of the optical inspection device to determine the reporting point is reduced, it will easily lead to the failure to detect defects and affect the production of subsequent products. Therefore, how to reduce the number of defects that the re-inspector needs to re-examine while maintaining the relative sensitivity of the optical inspection device to determine the reported point is indeed a problem to be solved in the industry. The present invention proposes a solution for this.

本發明之一目的在於光學檢測流程中，減少複檢站人員所需再次複檢除錯的數量，故可減少複檢站所需之複檢人員。 One purpose of the present invention is to reduce the number of re-inspection and debugging errors required by the re-inspection station personnel in the optical inspection process, so the re-inspection personnel required in the re-inspection station can be reduced.

本發明之另一目的在於提供一種人工智慧(AI)複檢系統，透過人工智慧建模訓練，在相對維持光學檢測裝置的判斷靈敏度前提下，於光學檢測裝置判斷報點後先行模擬預測人工複檢的結果，使得實際需要透過人工複檢的報點數量減少，以提高複檢效率。 Another object of the present invention is to provide an artificial intelligence (AI) review system, through artificial intelligence modeling training, under the premise of relatively maintaining the judgment sensitivity of the optical detection device, after the optical detection device judges the report point, it simulates and predicts the artificial review. The results of the inspection have reduced the number of points that actually need to be manually re-inspected to improve the efficiency of the re-inspection.

為達上述之目的，本發明提出一種人工智慧複檢系統，包含：光學檢測裝置，用以拍攝至少一待測物並將判斷出之複數NG結果或影像傳送至顯示裝置；顯示裝置，可顯示光學檢測檢測裝置判斷之複數NG結果或影像，以供操作人員進行人工複檢，並再挑選出人工複檢後之NG結果或是OK結果；人工智慧(AI)單元，其根據人工複檢後之NG結果及OK結果進行建模訓練；及運算單元，其根據人工智慧(AI)單元之建模訓練，運算得出一預測之人工複檢結果。 In order to achieve the above objective, the present invention provides an artificial intelligence review system, which includes: an optical detection device for photographing at least one object to be tested and transmitting the determined plural NG results or images to a display device; the display device can display The multiple NG results or images judged by the optical inspection device are used for manual re-inspection by the operator, and then select the NG result or the OK result after the manual re-inspection; the artificial intelligence (AI) unit is based on the manual re-inspection Perform modeling training on the NG results and OK results; and an arithmetic unit, which calculates a predicted manual recheck result based on the modeling training of the artificial intelligence (AI) unit.

為達上述之目的，本發明提出一種人工智慧複檢方法，包含：步驟S100：將光學檢測裝置判斷之複數NG結果或影像傳送至顯示裝置；步驟S200：操作人員根據步驟S100的複數NG結果或影像進行複檢，再挑選出人工複檢後之NG結果或是OK結果；步驟S300：將步驟S200經人工複檢挑選出之NG結果及OK結果傳送至人工智慧(AI)單元進行建模訓練；步驟S400：判斷人工智慧(AI)單元是否已完成建模訓練，若尚未完成建模訓練則重覆步驟S100~S300；若判斷已完成建模訓練則輸出經人工智慧(AI)單元預測之人工複檢結果。 In order to achieve the above objective, the present invention proposes an artificial intelligence re-inspection method, which includes: Step S100: Transmit a plurality of NG results or images determined by the optical inspection device to a display device; Step S200: The operator according to the complex NG results or images of step S100 The image is rechecked, and then the NG result or the OK result after the manual recheck is selected; Step S300: The NG result and the OK result selected by the manual recheck in step S200 are sent to the artificial intelligence (AI) unit for modeling training Step S400: Determine whether the artificial intelligence (AI) unit has completed the modeling training, if the modeling training has not been completed, repeat steps S100~S300; if it is determined that the modeling training has been completed, output the artificial intelligence (AI) unit predicted Review the results.

於本發明之一實施例中，以客戶端的第一批待測電路板做為人工智慧(AI)單元110的訓練資料，尤其是以該些電路板經光學檢測裝置100檢測後判定的瑕疵點及經人工複檢後之實際瑕疵點做為訓練資料。 In one embodiment of the present invention, the first batch of circuit boards to be tested of the client are used as the training data of the artificial intelligence (AI) unit 110, especially the defects determined after the circuit boards are detected by the optical inspection device 100 And the actual defects after manual re-inspection are used as training data.

於本發明之另一實施例中，將客戶端的第一批待測電路板做為人工智慧(AI)單元110的訓練資料1~m，並將該些訓練完成後之資料儲存為建模資料1；將客戶端的第二批待測電路板做為訓練資料1~m，並將該些訓練完成後之資料儲存為建模資料2，以此類推並將該些資料儲存為建模資料m。客戶端使用者可根據該些已訓練好的建模資料，直接選取較符合或近似於目前待測物的建模資料，以節省重新訓練的時間。 In another embodiment of the present invention, the first batch of circuit boards to be tested of the client are used as the training data 1~m of the artificial intelligence (AI) unit 110, and the training data is stored as modeling data 1. Use the second batch of circuit boards to be tested in the client as training data 1~m, and save the training data as modeling data 2, and so on, and save these data as modeling data m . The client user can directly select the modeling data that is more consistent or similar to the current object to be tested based on the trained modeling data, so as to save time for retraining.

100:光學檢測裝置 100: Optical detection device

200:運送單元 200: transport unit

300:人工視覺複檢站 300: Artificial vision review station

110:人工智慧(AI)單元 110: Artificial Intelligence (AI) Unit

1101:複數訓練資料 1101: Plural training data

1102:類神經網路學習模組 1102: Neural Network Learning Module

1103:建模資料儲存模組 1103: Modeling data storage module

1104:複數建模資料 1104: Complex Modeling Data

120:啟動單元 120: start unit

130:校驗訓練結果 130: Verify training results

140:運算單元 140: arithmetic unit

150:儲存單元 150: storage unit

1401:運算模組 1401: Computing Module

1402:預測結果顯示模組 1402: Forecast result display module

1403:使用者操作模組 1403: User operation module

S100~S400:步驟 S100~S400: steps

T100~T300:步驟 T100~T300: steps

U100~U300:步驟 U100~U300: steps

第1圖係習知之光學檢測裝置及人工視覺複檢站之相關示意圖。 Figure 1 is a schematic diagram of the conventional optical inspection device and the artificial visual reinspection station.

第2圖係為本發明經人工智慧(AI)單元訓練後之複檢系統架構圖。 Figure 2 is the architecture diagram of the recheck system of the present invention after the artificial intelligence (AI) unit training.

第3圖係為本發明實施例中類神經網路資料輸入、輸出示意圖。 Fig. 3 is a schematic diagram of data input and output of the similar neural network in the embodiment of the present invention.

第4圖係為本發明中人工智慧(AI)單元及與週邊資訊串流示意圖。 Figure 4 is a schematic diagram of the artificial intelligence (AI) unit and the surrounding information streaming in the present invention.

第5圖係為本發明中人工智慧(AI)單元完成建模訓練後之架構示意圖。 Figure 5 is a schematic diagram of the structure of the artificial intelligence (AI) unit of the present invention after completing the modeling training.

第6圖係為本發明中人工智慧(AI)單元之複檢流程圖。 Figure 6 is a flow chart of the recheck of the artificial intelligence (AI) unit in the present invention.

第7圖係為本發明中人工智慧(AI)單元建模訓練之另一實施例示意圖。 Figure 7 is a schematic diagram of another embodiment of the artificial intelligence (AI) unit modeling training of the present invention.

第8圖係為本發明中人工智慧(AI)單元完成建模訓練後之另一實施例示意圖。 Figure 8 is a schematic diagram of another embodiment of the artificial intelligence (AI) unit of the present invention after completing the modeling training.

第9圖係為本發明中人工智慧(AI)單元完成建模訓練後之另一實施例示意圖。 Figure 9 is a schematic diagram of another embodiment of the artificial intelligence (AI) unit of the present invention after completing the modeling training.

為充分瞭解本發明之目的、特徵及功效，茲藉由下述具體之實施例，並配合所附之圖式，對本發明做一詳細說明，說明如後：請參閱第2圖，係本發明增加人工智慧(AI)單元110模擬預測人工複檢的示意圖，首先在人工智慧(AI)單元110尚未完成建模訓練前，複數個待測物(例如印刷電路板、軟性電路板等)一一經由光學檢測裝置100取像拍照後(在此假設為電路板1、電路板2……電路板1000)，將其定義分別存在於電路板1、電路板2……電路板1000的斷路、短路、粘黏異物等瑕疵點傳送至人工視覺複檢站300進行人工複檢，而該些經人工複檢後確認為真實缺陷的結果除了可供後續的瑕 疵剔除復原外，亦可做為人工智慧(AI)單元110的建模訓練資料，例如將光學檢測裝置100判斷的電路板1瑕疵點及經人工複檢後的電路板1瑕疵點作為訓練資料1；將光學檢測裝置100判斷的電路板2瑕疵點及經人工複檢後的電路板2瑕疵點作為訓練資料2……以此類推。藉由人工智慧(AI)單元110來模擬預測人工複檢的電路板瑕疵結果，而具體的人工智慧(AI)訓練方式，例如類神經網路則以第3圖來做進一步的說明。 In order to fully understand the purpose, features, and effects of the present invention, the following specific embodiments are used in conjunction with the accompanying drawings to give a detailed description of the present invention. The description is as follows: please refer to Figure 2, which is the present invention Add a schematic diagram of artificial intelligence (AI) unit 110 to simulate and predict artificial recheck. First, before the artificial intelligence (AI) unit 110 has completed the modeling training, multiple objects to be tested (such as printed circuit boards, flexible circuit boards, etc.) one by one After taking an image and taking a photo via the optical inspection device 100 (assuming it is circuit board 1, circuit board 2... circuit board 1000), define the open circuit and short circuit on circuit board 1, circuit board 2... circuit board 1000 respectively Defects such as, sticky foreign bodies, etc. are sent to the artificial visual re-inspection station 300 for manual re-inspection, and the results confirmed as real defects after manual re-inspection are available for subsequent defects In addition to the defect removal and restoration, it can also be used as the modeling training data of the artificial intelligence (AI) unit 110. For example, the defect of the circuit board 1 judged by the optical inspection device 100 and the defect of the circuit board 1 after manual re-inspection are used as training data 1. Take the defects of the circuit board 2 judged by the optical inspection device 100 and the defects of the circuit board 2 after manual re-inspection as the training data 2... and so on. The artificial intelligence (AI) unit 110 is used to simulate and predict the result of the artificial re-examination of the circuit board defect, and the specific artificial intelligence (AI) training method, such as a neural network, is further illustrated in Figure 3.

繼續參閱第2圖，當訓練資料累積到一定數量時，可經由一啟動單元120來決定人工智慧(AI)單元110是否已建模訓練完成，假若尚未完成建模訓練(啟動單元120關閉)，則經光學檢測裝置100檢測後之複數待測物仍經由運送單元(圖未繪)送至人工視覺複檢站300進行人工複檢；假若已完成建模訓練則開啟啟動單元120，先將光學檢測裝置100所判定的複數NG影像或數量資料透過人工智慧(AI)單元110模擬預測人工複檢的結果後，再將此模擬預測之結果傳送至人工視覺複檢站300進行真正的人工複檢動作，使得實際需要人工複檢的瑕疵點數量減少。 Continuing to refer to Figure 2, when the training data has accumulated to a certain amount, an activation unit 120 can be used to determine whether the artificial intelligence (AI) unit 110 has completed the modeling training. If the modeling training has not been completed (the activation unit 120 is closed), Then the plurality of objects to be tested after being detected by the optical inspection device 100 are still sent to the artificial visual re-inspection station 300 via the transport unit (not shown) for manual re-inspection; if the modeling training has been completed, the activation unit 120 is turned on, and the optical After the multiple NG images or quantity data determined by the detection device 100 are simulated and predicted by the artificial intelligence (AI) unit 110, the result of the artificial recheck is then transmitted to the artificial visual recheck station 300 for real manual recheck Action, which reduces the number of defects that actually need to be re-examined manually.

於判斷人工智慧(AI)單元110是否已完成建模訓練時，可參照第2圖中的校驗訓練結果130，操作人員可將人工智慧(AI)單元110模擬預測之人工複檢結果和實際經過人工複檢的結果比對，若兩者之比對結果(例如瑕疵點的數量)小於一預設值，則可由操作人員判斷該人工智慧(AI)單元110已完成建模訓練而選擇開啟啟動單元120，接著讓所有經由光學檢測裝置100判斷出的瑕疵點均先經由人工智慧(AI)單元模擬預測後(此時模擬預測之人工複檢結果應少於原光學檢測裝置100判斷出的數量)，再將該模擬預測之結果傳送至人工複檢站300進行實際的人工複檢，藉此達到減少人工複檢數量及提升操作人員工作效率之目的。 When judging whether the artificial intelligence (AI) unit 110 has completed the modeling training, refer to the verification training result 130 in Figure 2. The operator can simulate the artificial intelligence (AI) unit 110 to predict the artificial recheck result and the actual After comparing the results of manual rechecking, if the comparison result of the two (such as the number of defects) is less than a preset value, the operator can judge that the artificial intelligence (AI) unit 110 has completed the modeling training and choose to turn it on Start the unit 120, and then let all the defects determined by the optical inspection device 100 be simulated and predicted by an artificial intelligence (AI) unit (at this time, the artificial re-inspection result of the simulated prediction should be less than that determined by the original optical inspection device 100 Quantity), and then transmit the result of the simulation prediction to the manual re-inspection station 300 for actual manual re-inspection, thereby achieving the purpose of reducing the number of manual re-inspection and improving the work efficiency of operators.

再請參閱第3圖~第4圖，第3圖為人工智慧(AI)單元110之建模訓練實施例說明，本實施例中以類神經網路訓練為例，然不依此為限。在第3圖中該人工智慧(AI)單元110具有第一層的訓練資料輸入節點D1及D2，接著有第二層的神經元D3、D4及D5，在第三層具有建模訓練後的資料輸出節點D6。 Please refer to Figs. 3 to 4 again. Fig. 3 is an example of modeling training of artificial intelligence (AI) unit 110. In this embodiment, neural network training is taken as an example, but it is not limited thereto. In Figure 3, the artificial intelligence (AI) unit 110 has the first layer of training data input nodes D1 and D2, followed by the second layer of neurons D3, D4, and D5, and the third layer has modeled training Data output node D6.

類神經網路的各個節點與各個神經元均有相對應的權重，主要就是利用輸入資料與輸出資料之間的關係特性，建模訓練出各輸入、輸出節點及神經元的權重及偏權值，而後再配合從光學檢測裝置100搜集到的待測物瑕疵點資料作為各輸入節點的輸入資料，用以進行人工複檢結果的模擬預測。 Each node and each neuron of the class neural network has a corresponding weight, mainly using the relationship between input data and output data to model and train the weight and partial weight of each input, output node and neuron , And then cooperate with the defect data of the object to be tested collected from the optical inspection device 100 as the input data of each input node for the simulation prediction of the manual re-inspection result.

例如在輸入節點D1與神經元D3具有一權重W13，輸入節點D2與神經元D5具有一權重W25；而神經元和輸出節點亦有一權重，例如神經元D4與輸出節點D6的權重為W46，神經元D5與輸出節點D6的權重為W56。除此之外，各別的輸入、輸出節點及神經元均具有各自的偏權值，例如D1的偏權值為θ1，神經元D4偏權值為θ4，輸出節點D6的偏權值為θ6。而每個節點傳輸至下個節點的傳輸數值計算方式如下，假設共有n個節點將各自的傳輸數值傳輸至該節點X，則節點Y傳輸至下一節點的傳輸數值y之公式為：

Figure 107117692-A0305-02-0007-1
For example, the input node D1 and the neuron D3 have a weight W13, the input node D2 and the neuron D5 have a weight W25; and the neuron and the output node also have a weight, for example, the weight of the neuron D4 and the output node D6 is W46, The weight of element D5 and output node D6 is W56. In addition, the respective input and output nodes and neurons have their own partial weights. For example, the partial weight of D1 is θ1, the partial weight of neuron D4 is θ4, and the partial weight of output node D6 is θ6. . The calculation method of the transmission value transmitted from each node to the next node is as follows. Assuming that there are n nodes in total transmitting their respective transmission values to the node X, the formula of the transmission value y from node Y to the next node is:
Figure 107117692-A0305-02-0007-1

其中，Wi為將傳輸數值傳輸至節點Y的n個節點中之第i個節點與節點Y所對應的權重，Xi為該第i個節點傳輸至節點Y的傳輸數值，θ則為節點Y的偏權值。 Among them, Wi is the weight corresponding to the i-th node of the n nodes that transmit the transmission value to node Y and node Y, Xi is the transmission value of the i-th node transmitted to node Y, and θ is the value of node Y Partial weight.

以神經元D5為例，輸入節點D1與神經元D5所對應的權重為W15，輸入節點D2與神經元D5所對應的權重為W25，而神經元D5的偏權值為θ5。假設輸入節點D1傳輸至神經元D5的傳輸數值為X1，而輸入節點D2傳輸至神經元D5的傳輸數值為X2，則神經元D5傳輸至輸出節點D6的傳輸數值為：(W15．X1+W25．X2)-θ5 Taking neuron D5 as an example, the weight corresponding to input node D1 and neuron D5 is W15, the weight corresponding to input node D2 and neuron D5 is W25, and the partial weight of neuron D5 is θ5. Assuming that the transmission value from the input node D1 to the neuron D5 is X1, and the transmission value from the input node D2 to the neuron D5 is X2, the transmission value from the neuron D5 to the output node D6 is: (W15. X1 + W25 ．X2)-θ5

在本實施中以類神經網路作為建模訓練方式，藉由大量已知的輸入資料及輸出資料關係塑模出各個權重及各個偏權值，於塑模出各個權重及偏權值後就可以即時的輸入資料進行輸出資料的預測。 In this implementation, a neural network is used as the modeling training method, and each weight and each partial weight are modeled by a large number of known input data and output data relationships. After each weight and partial weight are modeled, It is possible to input data in real time to predict output data.

第4圖為本發明基於類神經網路之人工複檢預測的人工智慧(AI)單元110及運算單元140周邊資訊流通示意圖，在本發明案中複數訓練資料1~訓練資料m(1101)即為經光學檢測裝置100判定之瑕疵點、人工複檢後所判定之真實瑕疵點、及人工複檢後所判定為正確OK的資料，其做為訓練資料由類神經網路學習模組1102訓練後，可再經由建模資料儲存模組1103分類儲存於複數建模資料1~建模資料m之中(1104)。 Figure 4 is a schematic diagram of information circulation around the artificial intelligence (AI) unit 110 and the computing unit 140 based on the artificial recheck prediction of the artificial neural network of the present invention. In the present invention, the plural training data 1~training data m(1101) are The flaws determined by the optical inspection device 100, the real flaws determined after manual re-inspection, and the data determined to be correct after manual re-inspection, which are used as training data to be trained by the neural network learning module 1102 Then, it can be classified and stored in the complex modeling data 1~the modeling data m through the modeling data storage module 1103 (1104).

舉例來說，例如客戶端有第一批待測電路板，經人工智慧(AI)單元110建模訓練後可儲存為建模資料1；同樣的，客戶端第二批待測電路板，可經人工智慧(AI)單元110建模訓練後儲存為建模資料2，以此類推......。而建模資料1~m(1104)客戶於日後可依待測電路板的批號或是近似的待測物，直接從建模資料1~m(1104)當中撰擇較符合的建模資料而毋需再重新訓練，以節省訓練所需的時間。 For example, if the client has the first batch of circuit boards to be tested, it can be stored as modeling data 1 after modeling and training by the artificial intelligence (AI) unit 110; similarly, the second batch of circuit boards to be tested can be After modeling training by artificial intelligence (AI) unit 110, it is stored as modeling data 2, and so on. The modeling data 1~m (1104) customers can directly select the more consistent modeling data from the modeling data 1~m (1104) according to the lot number of the circuit board to be tested or the similar DUT in the future. No need to retrain to save the time required for training.

然而需說明的是，訓練人工智慧(AI)單元110所需之資料並非以光學檢測裝置100判定之瑕疵點(NG)及人工複檢後判定之真實瑕疵點(NG)為限，在建模訓練人工智慧(AI)單元110時，亦可增加光學檢測裝置100判定之OK資料或是人工複檢後判定之OK資料做為訓練資料，其目的為使得經訓練後的人工智慧(AI)單元110更趨近於人工複檢之結果。 However, it should be noted that the data required for training the artificial intelligence (AI) unit 110 is not limited to the flaws (NG) determined by the optical inspection device 100 and the real flaws (NG) determined after manual re-inspection. When training the artificial intelligence (AI) unit 110, the OK data determined by the optical inspection device 100 or the OK data determined after manual re-inspection can also be added as training data. The purpose is to make the trained artificial intelligence (AI) unit 110 is closer to the result of manual re-examination.

再次看到第4圖，經由類神經網路學習模組1102訓練或是透過建模資料儲存模組1103選取的的資料，再經由運算模組1401計算後，由預測結果顯示模組1402顯示於可供操作人員觀看的地方，例如可呈現於一顯示裝置上，以方便操作人員透過使用者操作模組1403來決定是否開啟啟動單元120，此即為校驗訓練結果130，而操作人員於判斷是否開啟啟動單元120時，可將人工智慧(AI)單元110模擬預測之人工複檢結果和實際經過人工複檢的結果比對，若兩者之比對結果(例如瑕疵點的數量)小於一預設值，則判斷開啟啟動單元120。 See Figure 4 again. The data trained by the neural network learning module 1102 or selected by the modeling data storage module 1103 is calculated by the calculation module 1401 and displayed by the prediction result display module 1402. The place that can be viewed by the operator, for example, can be displayed on a display device, so that the operator can decide whether to turn on the activation unit 120 through the user operation module 1403. This is the verification training result 130, and the operator can determine When the activation unit 120 is turned on, the artificial intelligence (AI) unit 110 can compare the artificial re-inspection result simulated and predicted with the actual manual re-inspection result, if the comparison result (such as the number of defects) is less than one With a preset value, it is determined to turn on the activation unit 120.

再請看到第5圖，其為人工智慧(AI)單元110完成建模訓練後之架構示意圖，此時光學檢測置100判斷出之瑕疵點資料先輸出至人工智慧(AI)單元110預測模擬一人工複檢之結果，再將此預測結果送至人工視覺複檢站300進行真正的人工複檢並剔除瑕疵及復原，透過人工智慧(AI)單元110的預測模擬，可藉此減少需要操作人員複檢之瑕疵點數量，進而減少所需的操作人員。而由於人工智慧(AI)單元110已開啟，此時亦可將自動光學檢測裝置100檢測後判定之瑕疵點(NG)資料及經人工視覺複檢站300判定後之實際瑕疵點(NG)資料傳送至儲存單元150，該儲存單元150可位於運算單元140中；或可以雲端方式進行上傳或下載動作，例如該儲存單元150可雲端放置於設備商場地處；光學檢測裝置100、人工視覺複檢站300可放置於客戶端工廠處，並可進行遠端的數據資料分析。同樣的， 該儲存單元150不僅可儲存NG資料，亦可儲存自動光學檢測裝置100檢測後之OK資料或是經人工複檢後之OK資料，用以進行遠端數據資料分析。 Please see Figure 5 again, which is a schematic diagram of the architecture after the artificial intelligence (AI) unit 110 has completed the modeling training. At this time, the defect data determined by the optical inspection set 100 is first output to the artificial intelligence (AI) unit 110 for prediction and simulation The result of a manual re-inspection is sent to the artificial visual re-inspection station 300 for real manual re-inspection and defect removal and restoration. The artificial intelligence (AI) unit 110 predictive simulation can reduce the need for operations The number of defect points for personnel re-inspection, thereby reducing the number of operators required. Since the artificial intelligence (AI) unit 110 has been turned on, the defect (NG) data determined by the automatic optical inspection device 100 and the actual defect (NG) data determined by the artificial vision re-inspection station 300 can also be used at this time Send to the storage unit 150, the storage unit 150 can be located in the computing unit 140; or can upload or download in the cloud mode, for example, the storage unit 150 can be placed in the cloud in the equipment store; optical inspection device 100, manual visual review The station 300 can be placed at the client factory and can perform remote data analysis. same, The storage unit 150 can not only store NG data, but also store the OK data after the automatic optical inspection device 100 detects or the OK data after manual recheck, for remote data analysis.

再請看到第6圖，為本發明實施方式之流程圖，首先先將至少一個待測物傳送至自動光學檢測裝置100進行瑕疵檢測，於步驟S100中係將光學檢測裝置100判斷之複數NG結果或影像傳送至顯示裝置；於步驟S200中，操作人員根據前一步驟S100的複數NG結果或影像進行複檢，再挑選出人工複檢後之NG結果或是OK結果；於步驟S300中，係將前述步驟S200經人工複檢挑選出之NG結果及OK結果傳送至人工智慧(AI)單元進行建模訓練；接者步驟S400係判斷人工智慧(AI)單元是否已完成建模訓練，若尚未完成建模訓練則重覆步驟S100~S300，反之則輸出經人工智慧(AI)單元預測模擬之人工複檢結果。 Please see Fig. 6 again, which is a flowchart of the embodiment of the present invention. First, at least one object to be tested is transferred to the automatic optical inspection device 100 for defect detection, and in step S100, the optical inspection device 100 determines the plural NG The result or image is transmitted to the display device; in step S200, the operator performs a re-examination according to the multiple NG results or images in the previous step S100, and then selects the NG result or the OK result after manual re-examination; in step S300, The NG result and the OK result selected by the manual recheck in the aforementioned step S200 are sent to the artificial intelligence (AI) unit for modeling training; the next step S400 is to determine whether the artificial intelligence (AI) unit has completed the modeling training, if If the modeling training has not been completed, steps S100~S300 shall be repeated, otherwise, the artificial recheck result of artificial intelligence (AI) unit prediction and simulation will be output.

同樣的，在第6圖的步驟S100~步驟S300中，可進一步再增加相應的OK資料做為人工智慧(AI)單元110的訓練資料，例如可再將光學檢測裝置100檢測後之OK資料及人工複檢後之OK資料做為人工智慧(AI)單元110的訓練資料(圖未繪)，藉此使得人工智慧(AI)單元110的模擬預測更趨近人工複檢判讀之結果。 Similarly, in step S100 to step S300 in Figure 6, the corresponding OK data can be further added as training data for the artificial intelligence (AI) unit 110. For example, the OK data after the optical inspection device 100 can be detected and The OK data after the manual recheck is used as the training data of the artificial intelligence (AI) unit 110 (not shown in the figure), thereby making the artificial intelligence (AI) unit 110 simulation prediction closer to the result of the manual recheck and interpretation.

第7圖為本發明案人工智慧(AI)單元110建模之另一實施例之說明，首先光學檢測裝置100先將至少一待測物上之複數影像資料判讀為OK或是NG，接著將該些NG資料傳送至人工視覺複檢站300，由人工複檢再次判斷該些NG資料實際上是否其實為OK的資料(即光學檢測裝置100的誤檢)，亦或是由光學檢測裝置100判斷之NG資料為真實NG資料(即光學檢測裝置100的NG判斷和人工複檢之NG判斷相同)，接著將該些人工複檢判斷之OK及NG資料做為人工智慧(AI)單元110之建模訓練資料，再接著由校驗訓練結果130來判斷人工智慧(AI)單元是否已完成訓練，該校驗訓練結果130的方式例如可為：將人工視覺複檢站 300經人工比對出的NG資料和人工智慧(AI)單元110建模訓練後判定之NG資料進行比對，若兩者的誤差值在一預定範圍內，則判斷人工智慧(AI)單元110已完成訓練。 Figure 7 is an illustration of another embodiment of the artificial intelligence (AI) unit 110 modeling of the present invention. First, the optical inspection device 100 first judges the complex image data on at least one object to be measured as OK or NG, and then The NG data are sent to the artificial visual recheck station 300, and the manual recheck again determines whether the NG data is actually OK data (that is, the misdetection of the optical inspection device 100), or the optical inspection device 100 The judged NG data is real NG data (that is, the NG judgment of the optical inspection device 100 is the same as the NG judgment of the manual recheck), and then the OK and NG data judged by the manual recheck are used as the artificial intelligence (AI) unit 110 Model training data, and then use the verification training result 130 to determine whether the artificial intelligence (AI) unit has completed the training. The method of verifying the training result 130 may be: 300 The NG data obtained by manual comparison is compared with the NG data determined after the modeling training of the artificial intelligence (AI) unit 110. If the error value of the two is within a predetermined range, the artificial intelligence (AI) unit 110 is determined Training has been completed.

然而在第7圖中人工智慧(AI)單元110的建模訓練資料並非以人工視覺複檢站300判斷之OK或NG資料為限，必要時可再加入光學檢測裝置100判斷之OK資料(如第7圖中虛線所示)，加入光學檢測裝置100判斷之OK資料做為人工智慧(AI)單元110的訓練資料，其目的在讓訓練後的人工智慧(AI)單元110預測模擬的結果更接近人工視覺複檢站300的判斷結果，差別在於公式(1)中各節點權重或偏權值可能會有所調整變動。 However, the modeling training data of the artificial intelligence (AI) unit 110 in Figure 7 is not limited to the OK or NG data judged by the artificial vision recheck station 300. If necessary, the OK data judged by the optical inspection device 100 can be added (e.g. 7), the OK data judged by the optical inspection device 100 is added as the training data of the artificial intelligence (AI) unit 110, the purpose of which is to make the artificial intelligence (AI) unit 110 predict the simulation results after training. Close to the judgment result of the artificial visual recheck station 300, the difference is that the weight or partial weight of each node in formula (1) may be adjusted and changed.

第8圖為人工智慧(AI)單元110已完成建模訓練後之示意圖，首先光學檢測裝置100先將至少一待測物上的複數影像資料判讀為OK或是NG資料，接著將光學檢測裝置100判斷之NG資料傳送至人工智慧(AI)單元110進行人工複檢的模擬預測，第8圖中所示的係將人工智慧(AI)單元110模擬預測的NG結果再次送進人工視覺複檢站300進行人工判讀，必要時亦可將虛線所示之人工智慧(AI)單元110判定的OK結果送進人工視覺複檢測300進行人工再次判讀，以確認是否有人工智慧(AI)單元110誤判成OK的資料。 Figure 8 is a schematic diagram after the artificial intelligence (AI) unit 110 has completed the modeling training. First, the optical inspection device 100 first interprets the complex image data on at least one object to be tested as OK or NG data, and then the optical inspection device The NG data judged by 100 are sent to the artificial intelligence (AI) unit 110 for artificial reexamination simulation prediction. The NG result of the artificial intelligence (AI) unit 110 simulation prediction is sent to the artificial vision reexamination again. The station 300 performs manual interpretation. If necessary, the OK result determined by the artificial intelligence (AI) unit 110 shown by the dotted line can be sent to the artificial vision re-inspection 300 for manual re-interpretation to confirm whether the artificial intelligence (AI) unit 110 misjudged The information is OK.

在第8圖中，經由人工智慧(AI)單元110的先行預測模擬，已可大幅減少人工視覺複檢站300所需投入複檢的人力需求，較佳的甚至可以完全取代人力複檢。 In Fig. 8, the artificial intelligence (AI) unit 110 has advanced predictive simulation, which can greatly reduce the manpower requirement of the artificial visual re-inspection station 300 for re-inspection, and preferably can even completely replace the human re-inspection.

第9圖為本發明案人工智慧(AI)單元110建模完成後之另一實施例，首先光學檢測裝置100先將至少一待測物上的複數影像資料判讀為OK或是NG資料，接著將光學檢測裝100判斷之NG資料傳送至人工智慧(AI)單元110進行 人工複檢的模擬預測並再區分為OK或是NG資料，要注意到的是，人工智慧(AI)單元110在判讀資料時，通常是以一個數值大小或是百分比的型式來判定是OK或NG資料，若OK資料的機率大於50%則判定為OK資料，反之則判定為NG資料。例如下表一中第1列欄位中OK數值為0.65(即65%機率為OK)時，則人工智慧(AI)單元110即判定此為OK資料，當然這代表者也有35%的機率會判錯。若將一實際為NG的資料誤判為OK資料，此為光學檢測設備商最不願意見到的underkill情況(即漏檢了瑕疵)。為避免此情況發生，於軟體設計中可再預設參數(參數範圍為自設並非限制)，例如可將OK機率介於50%~99%之間的影像資料[50~99]判定為「unsure」，只有OK機率介於99%~100%之間的影像資料[99~100]才真正判定為OK資料。 Figure 9 is another embodiment of the artificial intelligence (AI) unit 110 of the present invention after the modeling is completed. First, the optical inspection device 100 first interprets the complex image data on at least one object to be measured as OK or NG data, and then Send the NG data judged by the optical inspection device 100 to the artificial intelligence (AI) unit 110 for processing The simulated predictions of manual re-examination are then classified as OK or NG data. It should be noted that when the artificial intelligence (AI) unit 110 interprets the data, it usually judges whether it is OK or NG in the form of a numerical value or a percentage. NG data, if the probability of OK data is greater than 50%, it is judged as OK data, otherwise, it is judged as NG data. For example, when the value of OK in the first column of the following table 1 is 0.65 (that is, 65% probability is OK), the artificial intelligence (AI) unit 110 will determine that this is OK data. Of course, this representative also has a 35% probability Wrong. If a data that is actually NG is misjudged as an OK data, this is an underkill situation that optical inspection equipment manufacturers are least willing to see (that is, a defect is missed). In order to avoid this situation, the parameters can be preset in the software design (the parameter range is self-set and not limited). For example, the image data [50~99] whose OK probability is between 50%~99% can be judged as " Unsure", only the image data [99~100] whose OK probability is between 99%~100% can be judged as OK data.

請再同時第9圖及下表一，同樣的若人工智慧(AI)單元110將一實際為OK的資料誤判為NG資料，此亦為我們不願意看到會造成待測物製造商成本增加的overkill情況，為避免此情況發生，於軟體設計中亦可預設參數(參數範圍為自設並非限制)，例如將NG機率介於50%~95%之間的影像資料[50~95]判定為「unsure」，只有NG機率介於95%~100%之間的影像資料[95~100]才真正判定為NG資料。 Please refer to Figure 9 and Table 1 at the same time. Similarly, if the artificial intelligence (AI) unit 110 misjudges a data that is actually OK as NG data, this is also because we don’t want to see that it will increase the cost of the DUT manufacturer In order to avoid this situation, the parameters can also be preset in the software design (the parameter range is self-set and not limited), for example, the image data whose NG probability is between 50%~95%[50~95] It is judged as "unsure", only the image data [95~100] whose NG probability is between 95%~100% can be judged as NG data.

Figure 107117692-A0305-02-0012-2
Figure 107117692-A0305-02-0012-2

在第9圖中虛線部分為可視情況再加入人工智慧複檢流程的部分，例如經人工智慧(AI)單元110判定幾乎確定為OK的資料[99~100]，可選擇性的再傳送至人工視覺複檢站300進行複檢；另外，經人工智慧(AI)單元110判定為「unsure」的資料亦可選擇性的歸類為「unsure」資料或也可直接判定為NG資料。 The dotted line in Figure 9 is the part that can be added to the artificial intelligence review process as appropriate. For example, the data that is determined to be OK by the artificial intelligence (AI) unit 110 [99~100] can be selectively sent to manual The visual reinspection station 300 performs re-inspection; in addition, the data determined as "unsure" by the artificial intelligence (AI) unit 110 can also be selectively classified as "unsure" data or directly determined as NG data.

人工視覺複檢站300主要目的即是以人工方式檢測前述該些「unsure」或是NG的資料，經由訓練好後之人工智慧(AI)單元110先行模擬、分類，已可大幅減少人工視覺複檢站300所需投入的人力，不僅可提高人工複檢的效率，亦可降低待測物製造商所需投入的人事成本。 The main purpose of the artificial vision review station 300 is to manually detect the aforementioned "unsure" or NG data. The artificial intelligence (AI) unit 110 after training can simulate and classify in advance, which can greatly reduce the artificial visual review. The manpower required by the inspection station 300 can not only improve the efficiency of manual re-inspection, but also reduce the personnel costs required by the manufacturer of the test object.

綜上所述，本發明在上文中已以較佳實施例揭露，然熟習本項技術者應理解的是，該實施例僅用於描繪本發明，而不應解讀為限制本發明之範圍。應注意的是，舉凡與該實施例等效之變化與置換，均應設為涵蓋於本發明之範疇內。 In summary, the present invention has been disclosed in the above preferred embodiments, but those familiar with the art should understand that the embodiments are only used to describe the present invention and should not be interpreted as limiting the scope of the present invention. It should be noted that all changes and substitutions equivalent to this embodiment should be included in the scope of the present invention.

S100~S400:步驟 S100~S400: steps

Claims (14)

一種人工智慧複檢系統，包含：光學檢測裝置，用以拍攝至少一待測物並將判斷出之複數NG結果或影像傳送至顯示裝置；顯示裝置，可顯示光學檢測裝置判斷之複數NG結果或影像，以供操作人員進行人工複檢直接以人工目測檢查相對應待測物，並再挑選出人工複檢後之NG結果或是OK結果；人工智慧(AI)單元，其根據人工複檢後之NG結果及OK結果進行建模訓練，其中，於人工複檢後之NG結果與人工智慧(AI)單元建模訓練後預測之NG結果差值小於一預設值時，自動進行模擬預測；及運算單元，其根據人工智慧(AI)單元之建模訓練，運算得出一預測之人工複檢結果。 An artificial intelligence re-examination system, comprising: an optical detection device for photographing at least one object to be tested and transmitting the determined plural NG results or images to a display device; the display device can display the plural NG results or images determined by the optical detection device Image, for the operator to perform manual re-inspection directly to manually inspect the corresponding object to be tested, and then select the NG result or OK result after manual re-inspection; artificial intelligence (AI) unit, which is based on manual re-inspection Perform modeling training for the NG results and OK results, where the difference between the NG results after manual recheck and the NG results predicted after the artificial intelligence (AI) unit modeling training is less than a preset value is automatically simulated and predicted; And an arithmetic unit, which calculates a predicted manual recheck result based on the modeling training of the artificial intelligence (AI) unit. 如請求項第1項所述之人工智慧複檢系統，該人工智慧(AI)單元之建模訓練資料另包括有光學檢測裝置判斷之複數OK結果或影像、光學檢測裝置判斷之複數NG結果或影像兩者其一或是組合。 For example, in the artificial intelligence review system described in claim 1, the modeling training data of the artificial intelligence (AI) unit further includes a plurality of OK results or images judged by an optical inspection device, and plural NG results judged by an optical inspection device or One or a combination of the two images. 如請求項第1項、第2項所述之人工智慧複檢系統，另包括至少一個儲存單元，用以儲存光學檢測裝置判斷出之複數NG或OK影像資料及操作人員進行人工複檢後之NG或OK結果。 For example, the artificial intelligence review system described in item 1 and item 2 of the request, further includes at least one storage unit for storing the plural NG or OK image data determined by the optical inspection device and the manual review by the operator NG or OK result. 如請求項第3項所述之人工智慧複檢系統，該至少一個儲存單元可位於運算單元中，或是以雲端方式提供該些影像或結果儲存。 For the artificial intelligence review system described in claim 3, the at least one storage unit may be located in the computing unit, or provide the images or result storage in the cloud. 如請求項第1項、第2項所述之人工智慧複檢系統，該人工智慧(AI)單元另包括一啟動單元，用以讓操作人員決定該人工智慧(AI)單元是否已完成建模訓練，若經人工複檢後之NG結果與人工智慧(AI)單元建模訓練後預測之NG結果差值小於一預設值，則判定完成建模訓練。 Such as the artificial intelligence review system described in item 1 and item 2, the artificial intelligence (AI) unit further includes an activation unit for the operator to determine whether the artificial intelligence (AI) unit has completed modeling Training, if the difference between the NG result after manual recheck and the NG result predicted after artificial intelligence (AI) unit modeling training is less than a preset value, it is determined that the modeling training is completed. 如請求項第5項所述之人工智慧複檢系統，其另包括一資料庫單元，用以儲存或輸出經建模訓練完成後之人工智慧(AI)單元訓練模組。 For example, the artificial intelligence review system described in claim 5 further includes a database unit for storing or outputting artificial intelligence (AI) unit training modules after modeling training is completed. 一種人工智慧複檢方法，包含：步驟S100：將光學檢測裝置判斷之複數NG結果或影像傳送至顯示裝置；步驟S200：操作人員根據步驟S100的複數NG結果或影像進行複檢直接以人工目測檢查相對應待測物，再挑選出人工複檢後之NG結果或是OK結果；步驟S300：將步驟S200經人工複檢挑選出之NG結果及OK結果傳送至人工智慧(AI)單元進行建模訓練；步驟S400：判斷人工智慧(AI)單元是否已完成建模訓練，若尚未完成建模訓練則重覆步驟S100~S300；若判斷已完成建模訓練則輸出經人工智慧(AI)單元預測之人工複檢結果，其中，當人工複檢後之NG結果與人工智慧(AI)單元建模訓練後之預測NG結果差值小於一預設值，則於步驟S400中判斷已完成建模訓練。 An artificial intelligence re-inspection method, including: Step S100: Transmit the plural NG results or images judged by the optical inspection device to a display device; Step S200: The operator performs re-inspection according to the plural NG results or images in Step S100 and directly performs manual visual inspection Corresponding to the object to be tested, select the NG result or the OK result after manual recheck; Step S300: Send the NG result and OK result selected by the manual recheck in step S200 to the artificial intelligence (AI) unit for modeling Training; Step S400: Determine whether the artificial intelligence (AI) unit has completed the modeling training, if the modeling training has not been completed, repeat steps S100~S300; if it is determined that the modeling training has been completed, output the prediction by the artificial intelligence (AI) unit Manual re-examination result, where, when the difference between the NG result after manual re-examination and the predicted NG result after artificial intelligence (AI) unit modeling training is less than a preset value, it is determined in step S400 that the modeling training has been completed. 如請求項第7項所述之人工智慧複檢方法，在步驟S300時，另可將光學檢測裝置判斷之複數OK結果或影像、光學檢測裝置判斷之複數NG結果或影像擇一或組合傳送至人工智慧(AI)單元進行建模訓練。 For example, in the artificial intelligence review method described in claim 7, in step S300, the multiple OK results or images judged by the optical detection device, and the plural NG results or images judged by the optical detection device may be alternatively or combined to be sent to Artificial intelligence (AI) unit for modeling training. 如請求項第7項所述之人工智慧複檢方法，若於步驟S400中判斷尚未完成建模訓練而重覆步驟S100~S300時，則先將光學檢測裝置判斷之複數NG結果或影像傳送至顯示裝置供操作人員進行人工複檢。 For example, in the artificial intelligence review method described in claim 7, if it is determined in step S400 that the modeling training has not been completed and steps S100~S300 are repeated, the plural NG results or images determined by the optical inspection device are first transmitted to The display device is for the operator to manually recheck. 一種人工智慧複檢方法，包括：步驟T100：將光學檢測裝置拍攝之複數影像區分為OK結果或是NG結果；步驟T200：將步驟T100中的NG結果傳送至已建模訓練好的人工智慧(AI)單元，並藉由人工智慧(AI)單元將步驟T100中判定為NG的結果進一步再區分為OK結果或是NG結果；步驟T300：將步驟T200中經人工智慧(AI)單元判定的NG結果傳送至人工視覺複檢站，由操作人員直接以人工目測檢查相對應待測物將人工智慧(AI)單元判定的NG結果進一步再區分為OK結果或是NG結果。 An artificial intelligence re-inspection method, including: Step T100: distinguish the plural images taken by the optical inspection device as OK results or NG results; Step T200: transmit the NG results in step T100 to the artificial intelligence that has been modeled and trained ( AI) unit, and the result determined as NG in step T100 is further divided into OK result or NG result by artificial intelligence (AI) unit; step T300: NG determined by artificial intelligence (AI) unit in step T200 The result is transmitted to the artificial visual reinspection station, and the NG result determined by the artificial intelligence (AI) unit is further classified into an OK result or an NG result by the operator directly using the manual visual inspection of the corresponding object to be tested. 如請求項第10項所述的人工智慧複檢方法，在步驟T300中另將人工智慧(AI)單元區分為OK的結果傳送至人工視覺複檢站，由操作人員再該些結果再進一步區分為OK結果或是NG結果。 For the artificial intelligence review method described in item 10 of the request, in step T300, the result of the artificial intelligence (AI) unit being classified as OK is sent to the artificial visual review station, and the operator will further distinguish the results The result is OK or NG. 一種人工智慧複檢方法，包括： 步驟U100：將光學檢測裝置拍攝之複數影像區分為OK結果或是NG結果；步驟U200：將步驟U100中的NG結果傳送至已建模訓練好的人工智慧(AI)單元，並藉由人工智慧(AI)單元將步驟U100中判定為NG的結果進一步再區分為OK結果或是NG結果或是Unsure結果；步驟U300：將步驟U200中經人工智慧(AI)單元判定的NG結果及Unsure結果傳送至人工視覺複檢站，由操作人員將人工智慧(AI)單元判定的NG結果及Unsure結果直接以人工目測檢查相對應待測物進一步再區分為OK結果或是NG結果。 An artificial intelligence review method, including: Step U100: Distinguish the multiple images taken by the optical inspection device as OK results or NG results; Step U200: Send the NG results in step U100 to the artificial intelligence (AI) unit that has been modeled and trained, and use artificial intelligence The (AI) unit further differentiates the result determined as NG in step U100 into OK result, NG result or Unsure result; Step U300: Transmit the NG result and Unsure result determined by the artificial intelligence (AI) unit in step U200 At the artificial visual re-inspection station, the operator will directly use the artificial intelligence (AI) unit to determine the NG result and the Unsure result to manually inspect the corresponding object to be tested and further differentiate it into an OK result or an NG result. 如請求項第12項所述的人工智慧複檢方法，在步驟U300中另將人工智慧(AI)單元區分為OK的結果傳送至人工視覺複檢站，由操作人員再該些結果再進一步區分為OK結果或是NG結果。 As the artificial intelligence review method described in item 12 of the request item, in step U300, the result of the artificial intelligence (AI) unit being classified as OK is sent to the artificial visual review station, and the operator will further distinguish the results The result is OK or NG. 如請求項第12項所述的人工智慧複檢方法，其中該些Unsure結果的範圍可因需求而做不同設定。 For the artificial intelligence review method described in claim 12, the range of the Unsure results can be set differently according to requirements.
TW107117692A 2018-05-24 2018-05-24 Artificial intelligence recheck system and method thereof TWI703514B (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
TW107117692A TWI703514B (en) 2018-05-24 2018-05-24 Artificial intelligence recheck system and method thereof
CN201810635724.1A CN109060817B (en) 2018-05-24 2018-06-20 Artificial intelligence reinspection system and method thereof

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
TW107117692A TWI703514B (en) 2018-05-24 2018-05-24 Artificial intelligence recheck system and method thereof

Publications (2)

Publication Number Publication Date
TW202004574A TW202004574A (en) 2020-01-16
TWI703514B true TWI703514B (en) 2020-09-01

Family

ID=64821036

Family Applications (1)

Application Number Title Priority Date Filing Date
TW107117692A TWI703514B (en) 2018-05-24 2018-05-24 Artificial intelligence recheck system and method thereof

Country Status (2)

Country Link
CN (1) CN109060817B (en)
TW (1) TWI703514B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI780580B (en) * 2021-01-19 2022-10-11 大陸商富泰華工業（深圳）有限公司 Image reinspection method, computer device, and storage medium

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109919925A (en) * 2019-03-04 2019-06-21 联觉(深圳)科技有限公司 Printed circuit board intelligent detecting method, system, electronic device and storage medium
CN110579479A (en) * 2019-08-09 2019-12-17 康代影像科技(苏州)有限公司 PCB maintenance system and maintenance method based on false point defect detection
CN112083002A (en) * 2020-08-26 2020-12-15 苏州中科全象智能科技有限公司 Capacitance appearance detection device and method based on artificial intelligence technology
CN112139050B (en) * 2020-09-08 2021-08-03 佛山读图科技有限公司 Cooperative light inspection method and system for high-speed product transmission
CN113386131A (en) * 2021-06-01 2021-09-14 江苏科瑞恩自动化科技有限公司 Full-automatic test handling system and control method thereof
CN113567446B (en) * 2021-07-06 2022-07-19 北京东方国信科技股份有限公司 Method and system for grading component defect detection quality
CN115100095B (en) * 2021-12-29 2023-08-22 苏州真目人工智能科技有限公司 PCB detection method based on non-supervision algorithm

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7580124B2 (en) * 2002-07-09 2009-08-25 Kla-Tencor Technologies Corp. Dual stage defect region identification and defect detection method and apparatus
US7761182B2 (en) * 2005-01-21 2010-07-20 Photon Dynamics, Inc. Automatic defect repair system
CN101915769A (en) * 2010-06-29 2010-12-15 华南理工大学 Automatic optical inspection method for printed circuit board comprising resistance element
CN103676868A (en) * 2013-12-09 2014-03-26 华南理工大学 Automatic monitoring and intelligent analyzing system used in FPC manufacturing critical process
TW201430768A (en) * 2013-01-30 2014-08-01 Hitachi High Tech Corp Defect observation method and defect observation device
CN106290378A (en) * 2016-08-23 2017-01-04 东方晶源微电子科技（北京）有限公司 Defect classification method and defect inspecting system
TW201732979A (en) * 2016-01-04 2017-09-16 克萊譚克公司 Optical die to database inspection
CN107677679A (en) * 2017-09-22 2018-02-09 武汉精测电子技术股份有限公司 Sorting technique and device the defects of L0 pictures in a kind of AOI detection
CN107992900A (en) * 2017-12-18 2018-05-04 深圳市盛波光电科技有限公司 Sample acquiring method, training method, device, medium and the equipment of defects detection
TW201928308A (en) * 2017-12-25 2019-07-16 群光電子股份有限公司 Light source detection system and method thereof

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH0712530A (en) * 1993-06-24 1995-01-17 Nec Corp Soldering inspection apparatus
US5751910A (en) * 1995-05-22 1998-05-12 Eastman Kodak Company Neural network solder paste inspection system
US20100194562A1 (en) * 2009-01-30 2010-08-05 Jong-Moon Lee Failure recognition system
CN104167052B (en) * 2014-04-30 2017-01-04 昆山古鳌电子机械有限公司 A kind of bill handling device and recognition methods
CN105425087B (en) * 2015-11-17 2018-02-16 杭州西力电能表制造有限公司 It is easy to device and the detection control method detected to the electric elements with stitch
CN107991795B (en) * 2017-11-22 2020-07-10 航天科工智能机器人有限责任公司 Method for detecting liquid crystal module by using automatic optical detection system of liquid crystal module

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7580124B2 (en) * 2002-07-09 2009-08-25 Kla-Tencor Technologies Corp. Dual stage defect region identification and defect detection method and apparatus
US7761182B2 (en) * 2005-01-21 2010-07-20 Photon Dynamics, Inc. Automatic defect repair system
CN101915769A (en) * 2010-06-29 2010-12-15 华南理工大学 Automatic optical inspection method for printed circuit board comprising resistance element
TW201430768A (en) * 2013-01-30 2014-08-01 Hitachi High Tech Corp Defect observation method and defect observation device
CN103676868A (en) * 2013-12-09 2014-03-26 华南理工大学 Automatic monitoring and intelligent analyzing system used in FPC manufacturing critical process
TW201732979A (en) * 2016-01-04 2017-09-16 克萊譚克公司 Optical die to database inspection
CN106290378A (en) * 2016-08-23 2017-01-04 东方晶源微电子科技（北京）有限公司 Defect classification method and defect inspecting system
CN107677679A (en) * 2017-09-22 2018-02-09 武汉精测电子技术股份有限公司 Sorting technique and device the defects of L0 pictures in a kind of AOI detection
CN107992900A (en) * 2017-12-18 2018-05-04 深圳市盛波光电科技有限公司 Sample acquiring method, training method, device, medium and the equipment of defects detection
TW201928308A (en) * 2017-12-25 2019-07-16 群光電子股份有限公司 Light source detection system and method thereof

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
人工智慧AOI應用於晶圓、電路板 - Manufacturing Defect Detection with Big Data : A Deep Learning Approach 2018.05.13 *
人工智慧AOI應用於晶圓、電路板 - Manufacturing Defect Detection with Big Data : A Deep Learning Approach 2018.05.13。

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI780580B (en) * 2021-01-19 2022-10-11 大陸商富泰華工業（深圳）有限公司 Image reinspection method, computer device, and storage medium

Also Published As

Publication number Publication date
CN109060817A (en) 2018-12-21
TW202004574A (en) 2020-01-16
CN109060817B (en) 2021-03-26

Similar Documents

Publication Publication Date Title
TWI703514B (en) Artificial intelligence recheck system and method thereof
US11449980B2 (en) System and method for combined automatic and manual inspection
EP3502966B1 (en) Data generation apparatus, data generation method, and data generation program
CN110659660B (en) Automatic optical detection classification equipment using deep learning system and training equipment thereof
CN110659662B (en) Flaw detection system and method using artificial intelligence
KR20190075707A (en) Method for sorting products using deep learning
US11315231B2 (en) Industrial image inspection method and system and computer readable recording medium
JP6795562B2 (en) Inspection equipment and machine learning method
CN114862832A (en) Method, device and equipment for optimizing defect detection model and storage medium
TWI694250B (en) Surface defect detection system and method thereof
KR102189951B1 (en) System and method for inspection of ship painting condition using image analysis
CN110290203A (en) Artificial intelligence cloud computing carries out Noninvasive production defect recognition and information communication to manufactured goods
WO2013001594A1 (en) Method for management of inspection of electronic substrate, apparatus for same, and apparatus for visual inspection of electronic substrate
JP2020042668A (en) Inspection apparatus and machine learning method
JP7440823B2 (en) Information processing device, information processing method and program
TW202230281A (en) Image reinspection method, computer device, and storage medium
JP2019158684A (en) Inspection system, identification system, and discriminator evaluation device
WO2019073615A1 (en) Worksite image assessment device and worksite image assessment program
TWI711936B (en) Equipment checking device and equipment checking system
JP2022088713A (en) Inspection system in article assembling process
US20230316499A1 (en) Autonomous penetrant testing
US20210055235A1 (en) Method to Automatically Inspect Parts Using X-Rays
CN115987380B (en) Test result analysis method and system of optical time domain reflectometer
CN117828499B (en) PCBA abnormal part determination method, system, storage medium and electronic equipment
CN117358622B (en) Method, device and system for detecting indoor and outdoor units of air conditioner