Quality assessment for microscopic images of manufactured IC

• Main Author: Huang, Zehao
• Other Authors: Gwee Bah Hwee
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering; Quality assessment; Microscopic IC images; No-reference IQA
• Online Access: https://hdl.handle.net/10356/182000
• Institution: Nanyang Technological University

Microscopic IC image analysis is a critical method for ensuring hardware security tasks, including Trojan detection, IP infringement detection, and integrity verification. However, the effectiveness of these tasks heavily relies on the quality of the acquired microscopic images. Current IC image quality assessment (IQA) methods have limitations when dealing with complex microscopic IC images, particularly in detecting subtle distortions and addressing adversarial attacks. These limitations motivate the need for more precise and robust assessment methods to enhance hardware security tasks. This research focuses on No-reference IQA (NR-IQA) techniques, utilizing deep learning, including attention mechanisms and hyper networks, to evaluate the quality of microscopic IC images. Specifically, a multidimensional attention network for reference-free image quality assessment (MANIQA) and a DefenceIQA approach incorporating adversarial defence mechanisms are integrated. The use of attention mechanisms allows the model to capture both global and local distortions effectively, while the adversarial defenses enhance robustness against attacks. Experimental results demonstrate that our method achieves a Spearman's Rank Correlation Coefficient (SRCC) of 0.96 and a Pearson Linear Correlation Coefficient (PLCC) of 0.91, where SRCC measures the rank correlation between predicted and true quality scores, and PLCC measures the linear correlation. Compared to traditional IQA methods, our approach demonstrates an accuracy improvement of 15%, ensuring more reliable IC image quality assessments for subsequent hardware security tasks. The proposed model effectively predicts the quality of IC images and categorizes them into five quality grades, providing a reliable assessment basis for hardware security tasks. These findings indicate that our approach has significant potential for broader application in various IC image analysis scenarios.