Taktpixel Technical Whitepaper

Taktpixel Technical Whitepaper

Taktpixel Co., Ltd. provides deep learning model support platform for the printing industry called "POODL" and, researches and develops peripheral technologies. This document shows that deep learning and machine learning technology adopted printing inspection site.
https://poodl.cloud/
https://taktpixel.co.jp/

4a8d16c35a4d9b2f6630cb4bdaa7ad35?s=128

Taktpixel Co., Ltd.

March 28, 2019
Tweet

Transcript

  1. 1 2019/03/27 Whitepaper of Taktpixel’s AI development Be in AI

    Solution to the Production Taktpixel Co.,Ltd
  2. Our Development for AI technology at glance 2 Development for

    cloud based software Deep Learning Novel detection by Encoder- Decoder model Image processing Inspection of printing, High speed processing Printing manufa- cturing Refer to another paper Deep learning engine for judgement process of printing image data ・Heatmap highlight ・Cloud based architecture ・Setting big data Other Development of technical theme ・Inspection, learning using non-defective products image with Encoder-Decoder model ・Automatic generation of image processing filter (Direct transfer from digital image data to scanning data )
  3. Deep learning engine for judgement process of printing image data

    • TensorFlow(+Keras) based engine • Available to use Python language • Basic functionalities • Generate image file list • Data augmentation (Pre-processing) • Data dividing • Cache function input image data • Deep leaning and generate DL models • Inference • Heatmap highlight (Saliency map) • Web Server (AWS virtual machine environment support)
  4. Feature of Printing image data Technical Issue ・Quantitative evaluation ・Need

    to input multi images ・changes of back ground colors Know-how is becoming common in the classification of photographs (3 channel color image) such as dogs and cats Inspection • Get further improvement the accuracy of the results of automatic inspection equipment • Analyze the result and apply it to production control Deep learning engine for judgement process of printing image data → Re-consider class columns → Re-design models → Learn big data
  5. Example images Pass / Failed classification Defective Variety Reason Example

    for Images and Classifications × Failed × Failed × Failed Pass Pass Hair waste Ink waste Dirt Foreign Substance Color shift Adhesion of the hair dust is treating as a defective product Treat ink waste as defective Treat dirt as defective Treat object that is able to get rid of as non-defective Treat micro misregistration as non-defective Deep learning engine for judgement process of printing image data
  6. Design for Deep Learning Model Input layer Middle layer Output

    layer [b, x, y, c] b: Batch size x: Width of image y: Height of image c: Number of Channels Optimizing pre-processing in order to set the number of image channels to 4 or more Make the network structure of appropriate size (It is not good to be too large) Taking Global Average Pooling layer instead of Fully Connected layer Terms ・Start learning with random initial values. Transfer learning from ImageNet is not performed. ・Set Learning Rate in Stepwise. Optimizer does not affect accuracy very much. Batch Normalize layer is more important than Dropout layer. Incorporate the Residual layer 11,283,478 26,073,878 23,593,174 50,474,518 134,360,598 55,784,214 139,670,294 21,818,390 54,313,942 0 20,000,000 40,000,000 60,000,000 80,000,000 100,000,000 120,000,000 140,000,000 160,000,000 Parameter size Fall into qualitative classification issue Good balance between accuracy and computational Deep learning engine for judgement process of printing image data
  7. Learning of large data set Reading image data Definition of

    Network structure type Initial weight setting It is impossible to learn image data that exceed the RAM capacity in conventional methodology of one time reading at pre-processing . With the mechanism of storing image data in storage and reading each batch, it is possible to perform stable learning processing ,even if it was tens of thousands of data sets. Reading image data & Image loading and cache creation Definition of Network structure type Initial weight setting Cycle of epoch Cycle of batch learning Inference, slope, weight calculation Cycle of epoch Cycle of batch learning Inference, slope, weight calculation Cache read Deep learning engine for judgement process of printing image data
  8. Amazon S3 Inspection System Manager Endusernetwork AWS Batch Managed ECS

    Amazon ECR Generation container status/ notification of completion Amazon SES Learning/Inference, Script(Python) AWS Lambda AWS Code Build AWS Code Pipeline 10GB~1TB Amazon ECS, Fargate Managing Application Amazon RDS Amazon ECS, Fargate Certification Application Amazon RDS Learning processing spot, instance{p2, p3), Storing user data Amazon ECS, Fargate GUI AWS Cloud Architecture for cloud based machine learning Deep learning engine for judgement process of printing image data
  9. Heatmap highlight (Saliency MAP) Interpretable Explanations of Black Boxes by

    Meaningful Perturbation : http://openaccess.thecvf.com/content_ICCV_2017/papers/Fong_Interpretable_Explanations_of_ICCV_2017_paper.pdf [1610.02391] Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization : https://arxiv.org/abs/1610.02391 [1710.11063] Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks : https://arxiv.org/abs/1710.11063 Regarding of results inferred by the deep learning model, it is difficult to comprehend calculation results, so researches approaching to explainable deep learning (Interpretable DL) are actively conducted. After classification processing, it is possible to indirectly analyze whether the model has acquired generalization ability by describing heatmap highlight. With implemented using the more stable mask method, we are able to succeed not to be restricted by the neural network structure. $ dlc-titan predict-ss --model-network ${SOURCE}/network.json --model-weight ${SOURCE}/trained.h5 --image-input-model ${F1} ${F2} -- cam-type Grad-CAM --image-masking-stride 1,1 --grid 16,16 --image-target-size 112,112,3 --last-conv-layer prediction --output- dir ./predict-ss/output/ --output-heatmap ./predict-ss/heatmap/${F2##*/} --verbose Available in “DLC-Titan” Deep learning engine for judgement process of printing image data
  10. Novel detection by Encoder-Decoder model Preparation of non-defective product data

    (100 images) Pre-processing Adding noise Input layer Compression feature Output layer Encoder Decoder In terms of hyperparameters and network model structure and how to add noise, it is necessary to adjust according to the features of printing Other Development of technical theme
  11. Novel detection by Encoder-Decoder model • No need to correct

    defective product data • As it is not a difference with the reference image, able to detect ambiguous • Defects is able be detected without alignment • No need to make complicated settings such as pointing at the frame of the inspection area Inspection target Non-defective product image Difference Image Algorism Other Development of technical theme
  12. Automatic generation of image processing filter 12 ? Color adjustment

    Blur Contour extraction Distortion ・Genetic algorism ・Deep learning (Encoder-Decoder) Convertor Define the relationship between the objects to be compared Regarding color conversion, an algorithm for generating an approximation conversion matrix of CMYK to RGB four-dimensional to three-dimensional color space is very useful Other Development of technical theme
  13. Direct transfer from digital to scanning image 13 Color ,noise,

    blur, there are various differences between design digital image data and scanning image data of the printing. Designed digital image Scanning image of printing products Deep learning Genetic algorism Other Development of technical theme
  14. contact@taktpixel.co.jp https://taktpixel.co.jp/en/#contact Contact info