POODL (deep learning platform for the printing industry)

https://poodl.cloud/

Transcript

1. /ˈpuː.dəl/ プードル Platform, train the DL Model with Labeled Data. Use the trained DL Model detect and predict for the unlabeled image data
2. Content proof In-line inspection Verification Final check Pre-press proof Off-line inspection Defect classification Design Printing Verification Packaging
3. Defect classification Verification Printing Packaging Not OK / Failed
4. Product Lineup • POODL Platform • Cloud version • On-premise version • Developer Utilities • POODL Client • POODL CLI • POODL SDK • API documentation • Developer Support
5. Basic Functions • Storage • Upload images and store them on a per-dataset basis • Allows you to split, merge, and edit datasets • Annotation • A simple user interface that allows you to focus on classification tasks • Ability to work securely with authority functions • Data Augmentation • Amplification of data that does not loss the characteristics of the inspection image • Learning • Process to train model without expertise • Analytical • Analyze the accuracy of a trained model • Download the appropriate learning model • Inference • Install into inspection equipment for fast off-line processing Storage Preprocessing Annotation Training Analytical
6. Storage Preprocessing Annotation Training Analytical Machine Operator Manager
7. Data in factory Production run Application / Machine is a FACTORY to make trained model Custom-ordered system Algorithm, extracted knowledge in image data
8. Features Web Application Authority Management Original DL Engine Efficient Annotation Scalable File Manager Linkage with External Devices
9. Web Application • By adopting container technology, the latest application can be used in both cloud and on-premises version • the functional unit services are able to flexibly deal with customization requests
10. Cloud Version • DB separated multi tenant • Scaling computing resources against requests • Low-cost deployments Features Web Browser Google Chrome, Firefox Network Cloud: >= 10 Mbps (download, upload) On-premises: >= 100 Mbps RAM >= 4GB Client System Requirements Application Foundation Separation of application and data, multi- tenant with DB division realized reliable data separation and scalability. A社 Computing Resources B社 C社 Authorization Service
11. On-premises version • Ensuring of security by self-management • Exclusive use of computing resources • Capitalization as production equipment • Independent on stability of internet line Features OS Linux (Red Hat Linux Series, Ubuntu etc.) Processing Unit GPU: NVIDIA GPU (CUDA 9.0) RAM >= 16GB Storage HDD (>=2TB), SSD for cache(>=400GB) Additional outer storage for backup Network >= 100 Mbps Application Environment Virtualizing: Docker System Requirements Manager Administrator Operator Machine Image Data knowledge Manage Collaborate through LAN ※ only for distribution in Japan
12. Authority Management • Automatically share what you work within the organization • Account privilege management is limited to administrators only • Controls authority group by group according to roles such as annotation work only, image upload only Organization Manager Group Operator Group Machine Group
13. Original DL Engine • Original DL Engine “DLC-Titan” based on TensorFlowTM • Learning of large-scale data set of several million scale • A network model suitable developed for print quality inspection • Prevent a faulty product from being put on the market by biased numerical optimization method • General-purpose trained model for application to small lot multi-product • It is also possible to any deep learning engine （It may be necessary to develop another API.） ※TensorFlowTM is developed by Google LLC, is an open-source software library for dataflow and differentiable programming across a range of tasks.
14. Efficient Annotation Tools • By specializing in learning of printed inspection images and narrowing down to necessary functions, you can concentrate on image classification work and efficiently perform annotation work. • The progress of work can be checked in real time by the administrator.
15. Scalable File Manager • Even when editing, copying, and concatenation operations of data sets do not excessively consume storage capacity Reference as virtual data Linked Real Object RDB ＋ Strorage Collect / Annotation Preprocess / Augmentation Training Inference / Analyze Light operating experience even if massive images
16. Linkage with External Devices • Provide API, SDK, CLI for direct data communication with existing facilities such as inspection machine. Inspection Machines File Servers Manager PCs ERP Systems
17. Developer support, Customization • POODL Client • POODL CLI • POODL SDK • API documentation We are also developing applications using POODL, implementing SDK into inspection machine, and providing support for linkage with other systems. Cloud Computing Service AWS (EC2, Batch, ECS, S3, RDS) Container / Virtualization Docker Web Application (Back-end) PHP(Laravel Framework) GUI (Front-end) Vue.js (element-ui) Database MySQL Machine Learning / Data Analyze AWS Batch, AWS ECS, Python (Tensorflow, Keras, numpy, pandas) Application Environment C#.NET (Windows Forms, WPF), Python, Node.js Image Processing OpenCV, HALCON Technology Base
18. Taktpixel Co., Ltd. contact@taktpixel.co.jp https://taktpixel.co.jp/#contact