Feature extraction for defect classification and yield enhancement in color filter and micro-lens manufacturing: An empirical study

Abstract

Yield improvement is an important issue in semiconductor-manufacturing supply chains, including color filter and micro-lens manufacturing. In the color filter and micro-lens processes, it is critical to quickly identify the defect pattern through the defect pictures and then take corrective actions to avoid greater yield loss. Until now, defect pattern recognition heavily relies on domain experts’ knowledge, which easily causes inconsistent classification results from person to person and unsatisfactory performance. In this study, a framework is proposed to achieve automatic defect detection and classification in color filter and micro-lens manufacturing to enhance the decision quality of pattern recognition. In particular, the proposed framework integrates Canny edge detection and classification and regression tree methodology. To validate the viability of the proposed framework in real settings, an empirical study was conducted in collaboration with a leading complementary metal oxide semiconductor image sensor foundry in Taiwan. The results not only showed the effectiveness of the proposed framework but also demonstrated the practical values.

Keywords:

• surface defect detection
• color filter and micro-lens processes
• classification and regression tree
• feature extraction
• yield enhancement

Acknowledgement

This research is supported by the Hsinchu Science Park (101A51), the National Science Council, Taiwan (NSC 100-2628-E-007-017-MY3; NSC 101-2811-E-007-004), the Advanced Manufacturing and Service Management Research Center of National Tsing Hua University (101N2074E1), the Toward World Class University Project from the Ministry of Education (102N2075E1), and the NSC Semiconductor Technologies Empowerment Partners (NSC STEP) Consortium (NSC 102-2622-E-007-013). The authors particularly appreciate the domain experts in VisEra Technologies Company Ltd. including Dr Ben Fun, GB Huang, and River Hung for their assistance in data collection and validation.