Archives

Parts in partitioned case dataset

– Overview

　This dataset was created for the purpose of evaluating the performance of visual inspection algorithms based on image processing. It consists of normal and logical anomaly images of multiple components arranged in partitioned block (partitioned-parts). By using this dataset, you can train inspection algorithms to classify normal and logically anomalous images of partitioned-parts, as well as perform quantitative evaluations of inspection performance on such parts. As a basic usage example, we assume that the “train” directory is used for training data and the “test” directory is used for testing data. The partitioned-parts include four types: individual bolts, individual nuts, bolt–nut pairs, and individual connectors.

　When you use this dataset in publications such as conference papers or journal articles, please cite the following URL or reference.

　URL:http://isl.sist.chukyo-u.ac.jp/archives/partitioned-image

　＜Reference＞　Tomohiro Yamada, Naoki Murakami, Naoto Hiramatsu, Hiroki Kobayashi, Shuichi Akizuki, Manabu Hashimoto, Logical Anomaly Detection based on Relative Similarity Analysis of Region Segments, International Workshop on Advanced Image Technology 2026 (IWAIT2026), vol.xx, pp.xx-xx, Kaohsiung, Taiwan, 2026/01/xx.

– How to collect data

This dataset was constructed using a specialized imaging device.
The overall appearance of the imaging environment is shown in the figure below.

– Released Data

This dataset consists of images captured by a specialized imaging device with a resolution of 1920 x 1080 pixels and 3-color channels. All images are provided in PNG format.

Download : Parts in partitioned case dataset（4.1GB） [1] [2] [3] [4] [5]

■bolts

Train (200 images, 681.7MB)
Test (Normal: 50 images, Anomaly: 50 images, 340.6MB)
Type of included anomalies: Count error, Type error

■nuts

Train (200 images, 669.6MB)
Test (Normal: 50 images, Anomaly: 50 images, 334.8MB)
Type of included anomalies: Count error, Type error

■bolts&nuts

Train (200 images, 693.2MB)
Test (Normal: 50 images, Anomaly: 50 images, 346.6MB)
Type of included anomalies: Count error, Type error, Combination error

■connectors

Train (200 images, 681.5MB)
Test (Normal: 50 images, Anomaly: 50 images, 341.7MB)
Type of included anomalies: Count error, Type error

- How to unzip the file
Please replace with the name of the file you downloaded, then run the following:

1. Combine the tar.gz parts

cat <dataset>_part_*.tar.gz > <dataset>.tar.gz

2. Extract the tar.gz file

tar -xvzf <dataset>.tar.gz

- Directory Structure dataset

dataset
|--bolts
|  |--train
|  |  |--good
|  |      |--001.png
|  |      |--002.png
|  |
|  |--test
|     |--good
|     |   |--001.png
|     |   |--002.png
|     |--logical_anomalies
|         |--001.png
|         |--002.png
| 
|--nuts
|--bolts_nuts
|--connectors

　This dataset consists of four categories: bolts only, nuts only, bolts and nuts, and connectors only. Each category includes two directories: training data (train) and test data (test). The directory “train” contains only normal images (good), and the directory “test” contains both normal images (good) and logically anomalous images (logical_anomalies).

▲Topへ