UltraChip: Improving Unsupervised Ultrasonic Image Anomaly Detection via Frequency-Spatial Feature Filtering and Gaussian Mixture Modeling
Wenjing Zhang1, Ke Lu1,2, Jinbao Wang3, Hao Liang1,4, Can Gao3, Jian Xue1
1School of Engineering Science, University of Chinese Academy of Sciences, Beijing 100049, China
2Peng Cheng Laboratory, Shenzhen 518055, China
3Shenzhen University, Shenzhen 518060, China
4Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China
INTRODUCTION
UltraChip is an industrial ultrasonic C-scan dataset designed for chip-package defect inspection in non-destructive testing (NDT). All samples are collected from real-world environment, capturing practical challenges such as strong noise, subtle defect appearance, and process-dependent variations. The dataset is intended for research on unsupervised/weakly-supervised anomaly detection, defect localization, and robustness evaluation.
UltraChip is acquired using a high-precision scanning acoustic microscope system, where ultrasonic echoes are sampled and digitized to form C-scan grayscale images with fine structural details. The dataset contains approximately 8,000 ultrasonic images and covers three representative defect types: Hole, Crack, and Layer. For precise localization evaluation, every defective test image is provided with pixel-level ground-truth masks.
To reflect structural differences introduced by various packaging processes, UltraChip is further grouped into three structural categories: Bonding, Two-Lead, and Four-Lead, enabling benchmarking under diverse normal conditions and process-dependent appearance shifts.

Some sample images in UltraChip
STRUCTURE
The UltraChip dataset is organized in a hierarchical directory layout and is compatible with common unsupervised anomaly detection pipelines. Each category contains train/test/ground_truth splits.
- train/good/: normal-only training samples (Good).
- test/good/: normal test samples.
- test/hole/, test/crack/, test/layer/: defective test samples.
- ground_truth/: pixel-level masks for defective test samples (defect folders match defect types).
Tabel 1 summarize the number of samples in the training set (normal only) and the test set (good + defects) for each structural category:
Table 1: Detail Information of UltraChip
|
Category
|
Train (Good)
|
Test (Hole)
|
Test (Crack)
|
Test (Layer)
|
Test (Good)
|
Test Total
|
Total
|
| Bonding |
946 |
277 |
27 |
9 |
309 |
622 |
1,568 |
| Four-Lead |
2,787 |
394 |
5 |
1 |
405 |
805 |
3,592 |
| Two-Lead |
2,134 |
408 |
5 |
0 |
406 |
819 |
2,953 |
| Overall |
5,867 |
1,079 |
37 |
10 |
1,120 |
2,246 |
8,113 |
Recommended Directory Layout
UltraChip/
Bonding/
train/good/
test/good/
test/hole/
test/crack/
test/layer/
ground_truth/hole/
ground_truth/crack/
ground_truth/layer/
Four-Lead/
...
Two-Lead/
...
OBTAINING THE DATA
We make UltraChip dataset available for academic research purposes. To access the dataset for research (noncommercial use), please send an
email to luk@ucas.ac.cn with a
signed AGREEMENT.
Please note that
any commercial use of the dataset is prohibited.
PUBLICATIONS
- Wenjing Zhang, Ke Lu, Jinbao Wang, Hao Liang, Can Gao and Jian Xue. Improving Unsupervised Ultrasonic Image Anomaly Detection via Frequency-Spatial Feature Filtering and Gaussian Mixture Modeling. IEEE Transactions on Image Processing, 2026. DOI: 10.1109/TIP.2026.3659292
Copyright (C) 2026 University of Chinese Academy of Sciences, All Rights Reserved.