Poster Session: Democratizing Deep Learning to Accelerate Image Based Defect Inspection
There is an exponential growth of enterprises adopting Artificial Intelligence to solve their business-critical problems. But they face high barriers in terms of resources and expertise required to operationalize the ML Model, which in fact has made the AI journey time consuming and challenging. In deep learning which a subfield of machine learning and AI, there is no handcrafting of features as the neural networks can automatically learn them. But for the developed model to be accurate, reliable, and robust, it has to be trained on huge amounts of data. Because of the dynamic nature of the technology and open research initiatives, for solving a specific problem, there are numerous network architectures available. There is also a need to explain why the model has come up with a prediction result during inference. For solving computer vision problems, we can bring in automation in the different phases like image annotation, image augmentation, network topology selection and training, evaluation, cross-validation, inference, and deployment. This leads to accelerated development and also ensures that errors are minimum. Data scientists and data analysts can focus on research and identifying the best network topologies, customization options, optimization of parameters, etc. In this proposal, we showcase how QuEST’s accelerator tool for deep learning can help to accelerate the development of a deep learning-based solution for image-based defect inspection.
The defects during semiconductor wafer manufacturing cause huge losses to the companies' yield. Identification and classification of the defects help to analyze the root cause of defect so that it can be resolved faster. Automatic optical inspection and images from scanning electron microscopes (SEM) can be used to examine defect patterns and identify root causes of die failures. These techniques help to achieve consistency and accuracy in detection, thereby improving product quality. The defect is identified by some features or patterns in the images. It is also necessary to identify whether the defect is a true defect or not and classify the type of the defect. Usually, a final verification is done by a human expert to judge the type of defect. Deep learning techniques using Convolutional Neural Networks can be used to detect and classify the type of manufacturing defect with better accuracy in less time.
We have developed an AI powered solution using Deep Neural Networks for semiconductor defect inspection problems. For this, we have used QuEST's proprietary Deep Learning Automation tool to accelerate the different phases involved in the development of the deep neural network. The developed model can localize and classify defect areas. A segmentation network to locate the area of defect by training the model on a weakly labeled simulated dataset representing typical manufacturing defects was built and could achieve a mean IoU of 0.9. Domain expert can control rate of overkill and underkill on inspection of defect size/area. An optimized model using OpenVINO was also developed. By using the accelerator tool for development of the model we could prove that there was a 30% increase in productivity and a cost reduction of 30%.