GENSS: Defect Classification Method on Extremely Small Datasets for Semiconductor Manufacturing

In order to detect defects on semiconductor components, machine learning (ML) and deep learning (DL) approach are introduced into vision inspection system where these ML/DL approach capable to identify object with complex appearance and eliminate the task requirement of placing PCB or electronic component on special fixture and jig. However, there is a small dataset issue occur when implementing ML/DL approach into vision inspection application due to collecting such data is difficult and expensive in industry application. A novel defect classification system using extremely small datasets is proposed, named GENSS, where the GEometric variant property is used to geNerate synthetic data, and then uses Structural Similarity and image hashing approaches to automatically select pre-trained models. Early experiments showed that the suggested method is a practical approach that can achieve high accuracy, at 85.71 %.