Challenge
In semiconductor manufacturing, product quality is assessed mainly by images taken by microscopes (optical, TEM, SEM…).
One of the most commonly used assessment methods is measurement.
For example, measuring the thickness of a layer of material, the size or area of a part, the angle or curvature, etc. The challenges include a limited number of sample data (particularly cross-sectional images) and heavy noise.
Solution
Traditional image processing methods have the advantages of simplicity, and fast implementation, but the disadvantage of lack of generality, often having to adjust many parameters to run well.
In contrast, Deep Learning-based methods have more general processing capabilities, but require a long model training process and require powerful hardware.
Our approach is to flexibly combine both traditional image-processing methods and Deep Learning-based methods to take advantage of both.
Outcome
Our results are considered as good as manual measurements conducted so far by experienced staff.
Our automatic measurement solution was integrated into the client’s quality control system and significantly improved efficiency.
