Efficient and Adaptive Semantic Segmentation of HBMs using Incremental Learning

Due to the increasing miniaturization of package interconnects, detecting multiple defects in buried structures is becoming more important and challenging. Recent deep learning image segmentation models have significantly improved defect detection accuracy and are now being integrated into inspectio...

Full description

Saved in:

Bibliographic Details
Published in:	2023 IEEE 25th Electronics Packaging Technology Conference (EPTC) pp. 938 - 943
Main Authors:	Chang, Richard, Thakur, Namrata, Wang, Jie, Li, Yurni, Chong, Ser Choong, Singh Pahwa, Ramanpreet
Format:	Conference Proceeding
Language:	English
Published:	IEEE 05-12-2023
Subjects:	Adaptation models Computational modeling Deep learning Limiting Semantic segmentation Training Training data
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Due to the increasing miniaturization of package interconnects, detecting multiple defects in buried structures is becoming more important and challenging. Recent deep learning image segmentation models have significantly improved defect detection accuracy and are now being integrated into inspection processes by major semiconductors companies. However, data preparation is very time-consuming and requires significant resources. Additionally, trained models are only applicable to the current product version and may be unable to adapt to new design when it becomes available, limiting their scalability and adaptability. Current deep learning models and processes are also still limited by their lack of flexibility and data-heavy approach. Current models are often tied to their training datasets, in turn, limiting their usability and consuming large storage resources for future retraining. In this paper, we introduce a new framework to significantly increase the model's flexibility and adaptability to new data. We propose a generic training strategy and a new loss function that reduced the training time by 60% and the required amount of data by 48%. It will then make the training process more efficient as well as save computational and storage resources. This will also increase the reliability and flexibility of segmentation models for in-line inspections on 3D scans. Applicability of deep learning models to multiple packaging conditions and parameters will also improve the adoption by the industry.
DOI:	10.1109/EPTC59621.2023.10457605