Texture Image Dataset with Text of Defect Aspect
– Overview
This image dataset was created for the purpose of estimating the state of a defect from a defect image. The images are classified into a normal class and three types of defect classes. The accompanying CSV file contains pairs of image filenames and the corresponding text descriptions of the defect appearance. By using this dataset, it is possible to evaluate scratches, foreign objects, and discoloration within an image using a VLM（Vision Language Model） based on defined aspects. The basic usage assumes verifying the zero-shot performance on texture-based defect images. For example, zero-shot performance can be evaluated by using four out of the five texture classes for training and the remaining one class for inference.
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/texture-with-text
＜Reference＞
栗田大樹，村上尚生，平松直人，田上鈴奈，小林大起，秋月秀一，橋本学，CLIPモデルを用いた未学習素地の欠陥様相推定手法の提案，動的画像処理実利用化ワークショップ2025（DIA2025），IS3-13，pp.460-464，きらめきみなと館，2025/03/06.
– Released Data
The images in this dataset were captured (width 1920, height 1080, 3 channels), cropped to RGB images (width 1080, height 1080, 3 channels), and then resized to RGB images (width 224, height 224, 3 channels) for release. All images are saved in PNG format.
Download：Texture Image Dataset with Text of Defect Aspect (540MB)
■ Aluminum Plate Images (1296 images, 114MB)
Breakdown : Scratch: 324 images・29MB, Discoloration: 324 images・29MB, Foreign Object: 324 images・29MB, Normal: 324 images・28MB
■ Cloth Images (1296 images, 160MB)
Breakdown : Scratch: 324 images・40MB, Discoloration: 324 images・40MB, Foreign Object: 324 images・40MB, Normal: 324 images・40MB
■ Leather Images (1296 images, 105MB)
Breakdown : Scratch: 324 images・26MB, Discoloration: 324 images・26MB, Foreign Object: 324 images・26MB, Normal: 324 images・26MB
■ Rubber Plate Images（1296 images，60MB）
Breakdown : Scratch: 324 images・16MB, Discoloration: 324 images・16MB, Foreign Object: 324 images・15MB, Normal: 324 images・13MB
■ Wood Image（1296 images，100MB）
Breakdown : Scratch: 324 images・24MB, Discoloration: 324 images・24MB, Foreign Object: 324 images・24MB, Normal: 324 images・28MB
– Directory Structure
datasets
dataset
|--aluminum
| |--GT.csv
| |--scratch
| | |--aluminum_scratch1.png
| | |--aluminum_scratch2.png
| |--discoloration
| | |--aluminum_discoloration1.png
| | |--aluminum_discoloration2.png
| | --foreign object
| | |--aluminum_foreign object1.png
| | |--aluminum_foreign object2.png
| | --normal
| |--aluminum_normal1.png
| |--aluminum_normal2.png
|--cloth
The dataset is divided into five categories: Aluminum Plate Image(aluminum), Cloth Image(cloth), Synthetic Leather Image(leather), Rubber Plate Image(rubber), and Wood Image(wood). Each category includes four types: three defect classes (Scratch, Discoloration, Foreign Object) and one Normal class. Each class contains 324 images. The GT.csv file also contains, row by row, the image filename and its corresponding text. Specifically, normal image filenames are paired with the text Normal, and defective image filenames are paired with text detailing the four defect types and appearances.