Intelligent defect detection platform

• Responsible for responding to the needs of the front-end, using Python flask and Pytorch to conduct model training and provide the front-end with visualization functions and model-related data.
• take advantage of asyncio and thread to shorten the front-end response time.
• Utilize multi-process multi-GPU training successfully reduced the time to 3/4 of the original.
• Product different docker images and docker-compose for the project so that the projects do not affect each other and achieve one-click deployment.
• Use docker-compose to define training, database, cache, and prediction services, and split the container to make it easier for others to participate in the project.

Panel defect detection

• Use the Python Pytorch framework to write an object detection algorithm to automatically find out the defect location and type.
• According to requirements of the project , multi-process and multi-thread mechanisms are used to improve execution efficiency, successfully reduce the time to 1/2 and increase FPS.
• Customize different dockerfile and docker-compose for the project so that the projects do not affect each other.
• Product different docker images and docker-compose for the project so that the projects do not affect each other and achieve one-click deployment.

Mobile phone appearance defect detection

• In order to replace manual work, write Python code and then automatic detection is introduced
• Using the Keras framework, developed a semantic segmentation algorithm (CNN/U-net)

Panel defect classification

• In order to automate the production line, write Python code and then automatic detection was introduced. Responsible for using Tensorflow framework, developing deep learning algorithms (CNN/Xception), and increasing the detection rate from 50% to 90%.
• Use weight class and sampler mechanism to overcome the problem of imbalanced data. Use transfer learning to speed up the model convergence time.