# IH-ViT: Vision Transformer-based Integrated Circuit Appear-ance Defect Detection

**Authors:** Xiaoibin Wang, Shuang Gao, Yuntao Zou, Jianlan Guo, Chu Wang  
**arXiv:** [2302.04521](https://arxiv.org/abs/2302.04521)  
**Date:** 2023-02-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.04521v1/

## Key points

- The algorithm revolutionized the field of IC defect detection
- Combines strengths of ViT and convolution operation for more accurate decisions
- Overcomes subtle differences in defect details that traditional algorithms struggle with
- Addresses imbalance of information across diverse datasets using dual-channel image segmentation technique
- Hybrid IH-ViT model achieved an impressive accuracy rate of 72.51%
- Outperformed standalone models like ResNet50 and ViT
- Promises to enhance productivity within IC packaging and testing companies

## Summary

The  algorithm has revolutionized the field of , addressing long-standing challenges such as low recognition rates and slow speeds. By combining the strengths of  and , this model can extract image features from both local and global perspectives, resulting in more accurate decisions regarding the class of defects present in ICs. One major hurdle in detecting these defects is the subtle differences in details that traditional algorithms struggle to identify. To overcome this limitation, the traditional ViT model was enhanced with an additional convolution operation within the batch processing stage, allowing for more precise detection of nuanced defects. Another obstacle faced in IC defect detection is the imbalance of information across diverse datasets from various sources. To address this issue, a dual-channel image segmentation technique was employed to further enhance accuracy. After rigorous testing, the hybrid IH-ViT model demonstrated significant improvements over existing models, achieving an impressive accuracy rate of 72.51%. This represents a substantial increase compared to standalone models like ResNet50 and ViT, highlighting the effectiveness of this proposed algorithm in quickly and accurately detecting IC appearance defects. Ultimately, this advancement holds great promise for enhancing productivity within IC packaging and testing companies by streamlining defect identification processes.