The platform of POODL, train the Deep Learning (DL) Model with Labeled Data.
Users can make the trained DL Model detect and predict for the unlabeled.
How to use POODL Web-based train the DL Model
POODL Web application works on Cloud, user can start POODL operation making DL model with using image data provided by inspection equipment on cloud-based computing. And if the user likes to apply a useful DL model to the target, it is possible to install an on-premise target and target connecting to Web API base.
The image data generated by the inspection equipment is uploaded to POODL by an SDK or manual operation. Perform annotation work on image data and create a learning model based on it. Use the local environment after verifying whether the best model is available.
・ Has a tuning function that can create learning models required for work at every site of every company, such as multiple image learning algorithms and pre-processing of image data does not affect the image itself.
・ Has a function to cache large amounts of image data efficiently. By using hundreds of thousands of image data to create a learning model would be a highly accurate learning model.
・ An easy-to-use GUI for efficient annotation work is provided. In the future, we also plan to support workability improvement using semi-supervised learning.
Apply for the target
There is a several methods of applying to the target, a method of linking APIs via the Internet, which is used from a Web browser. The Web API adopts a microservices architecture and it brings adds various functions and works with third-party products. In addition, for offline it is used locally, it can be incorporated into existing inspection equipment.
Use cases of POODL
AI is used to automate and improve the variability of visual inspection by human in the quality inspection process of printed matter.
Although the development and introduction of image inspection devices have been advanced by machine vision technology, such as cameras and sensing technologies, visual inspection is still required at the site. The reason is that there is a variation in the quality of the printed matter and the limit of the conventional inspection performance, so it is impossible to perform a high-level human visual judgment with an inspection equipment of the conventional image processing technology. For example, noise and subtle color changes in the printed matter that occur during the printing process have reached the limit of their detection capability with conventional camera sensing technology. In terms of uniformity of judgment and accuracy of human visual observation using DL model will be increase productivity much more than human eyes-based inspection.