Lithographic Hotspot Detection Based on Dual-Branch Feature Fusion and Attention Augmentation

Abstract

With the rapid development of semiconductor technology and the continuous scaling down of circuit feature sizes, hotspot detection has become an increasingly critical task in the workflow. Although most deep learning models have been widely applied to this task, many still suffer from low accuracy and a high number of false alarms when dealing with class-imbalanced datasets. To address these issues, we propose a deep learning model based on dual-branch feature extraction and attention enhancement. This model consists of two parallel feature extraction branches. The first leverages a MetaDCNeXt architecture to capture spatial structural features of hotspots, while the second employs a multi-DenseConvSE architecture to extract semantic features through dense connections and channel attention. A positional convolutional block attention module is used to fuse the outputs from both branches, suppressing redundant information and emphasizing critical features. Experimental results on the ICCAD 2012 benchmark demonstrate that our model achieves the lowest average number of false alarms and the highest average F1 score, with its overall performance significantly surpassing all compared detectors.