O slideshow foi denunciado.
Utilizamos seu perfil e dados de atividades no LinkedIn para personalizar e exibir anúncios mais relevantes. Altere suas preferências de anúncios quando desejar.

基於少樣本深度學習之橡膠墊片檢測系統

基於少樣本深度學習之橡膠墊片檢測系統

  • Seja o primeiro a comentar

基於少樣本深度學習之橡膠墊片檢測系統

  1. 1. 何昭慶 HoChao@ntut.edu.tw 國立臺北科技大學製造科技研究所暨機械工程系
  2. 2. 致謝 Acknowledgement  臺北科技大學  蘇詠靖 顧問、房士豪 顧問  李柏桀、盧志嘉、陳奕帆、潘桓甯  科技部、產學計畫  結合機器視覺與深度學習之表面缺陷檢測系統  深度學習之外觀瑕疵檢測系統  Cell Phone Gasket智動化檢測 Intelligent Automated Inspection Laboratory
  3. 3. Method of Image based Inspection Deep Learning • Regular rule based tools fail • Deformable objects • Variable orientation • Cracks inspection • Scratches inspection • Injection moulding • Heterogeneous material • Medical imaging Rule-based processing • Code reading • Gauging • Print inspection • Fixed orientation • Dimensional measurements • Presence or absence checking • Print inspection • Rigid objects Intelligent Automated Inspection Laboratory
  4. 4. Image processor Semi-automatic annotation Dataset manager Device manager Challenges for AOI in defect inspection Intelligent Automated Inspection Laboratory Bigger whole image Bad labelling Lumen decay
  5. 5. Data driven model Deep General feature Specific feature Intelligent Automated Inspection Laboratory
  6. 6. Toolchain for AOI Intelligent Automated Inspection Laboratory 3rd party Our solutions
  7. 7. System architecture Intelligent Automated Inspection Laboratory • Data capture • Light control • Preprocessing • Post-processing • Display • Device management • Data management • Data analysis • Model management • Logging • MES interface • Model training • Inference
  8. 8. Back-End Linux Platform Database Linux Platform Front-End Viewer Transfer learning DB AI analyzer (NTUT-Net) 0 1 0 1 0 1 0 10 1 0 1 0 1 0 1 Data-driven model Intelligent Automated Inspection Laboratory
  9. 9. Offline Training Online Prediction TrainingTrainingData PredictionPredictionData Pixel annotation Extract image patches DNN model DNN model Sliding window extraction Softmax OK Defect … Weights s Padding P Intelligent Automated Inspection Laboratory
  10. 10. Silicone rubber: seamlessly sealing gaskets Provide the water-proofed protection • Quality control of molding silicone rubber gaskets production is very challenge • Rely on manual observation to classify the faults Challenges: 1. Dark and low reflective material surface of the silicone rubbers. 2. Defects are irregular shape. AI: • Ultimately a technology to better extract faint signal from noise. • Deep Learning is not magic, it cannot extract signal from nowhere C.-C. Ho, et al., ISPEMI 2018 Intelligent Automated Inspection Laboratory
  11. 11. Inner ring Deep Learning: labeling Thickness Notches Background 1 Background 4 Background 5 Background 2 Background 3 Intelligent Automated Inspection Laboratory
  12. 12. Blending 原始影像 特徵圖 Mask 缺陷位置 Training Intelligent Automated Inspection Laboratory
  13. 13. Thanks for your attention Q & A Intelligent Automated Inspection Laboratory

×