CorkHSI

This data contains hyperspectral images, in near-infrared wavelength range, taken from cork sheets. It is mainly collected for the purpose of semi-supervised anomaly detection and segmentation of chemical and physical defects in cork sheets.

This is the official dataset that is collected and used in our publication in ICPR2026 titled CorkHSI: Hyperspectral Anomaly Detection in Corks Using an Autoencoder with a Novel Spectral–Spatial Loss Optimization.

Anomaly Detection and Segmentation in Cork Sheets

Cork is a natural, renewable, and biodegradable material that is widely used in various industries, including wine bottle stoppers, flooring, and insulation. However, during the production process, cork sheets can develop various defects that can affect their quality and performance. These defects can be caused by various factors, such as insect attacks, mechanical damage, and chemical contamination. Detecting and segmenting these defects is crucial for ensuring the quality of cork products. Hyperspectral imaging (HSI) is a powerful technique that can capture detailed spectral information from materials, making it suitable for detecting subtle defects in cork sheets. Our CorkHSI dataset aims to facilitate research in this area by providing a comprehensive collection of hyperspectral images of cork sheets with various defects. It can be used to develop and evaluate semi-supervised anomaly detection and segmentation algorithms for cork sheets as it contains the labels for anomalous samples as well as pixel-level ground truth masks of defective regions which can be used for anomaly segmentation tasks.

Data Colelction

The CorkHSI dataset was collected using a SPECIM FX17 near-infrared hyperspectral line-scan camera, which captures each cork sheet as a hyperspectral cube with 224 spectral bands. During data acquisition, the camera was mounted on an adjustable support, and the cork samples moved under the camera on a laboratory scanner while two halogen light sources were positioned on both sides to provide uniform illumination and reduce shadows and reflections. For each sample, dark and white reference measurements were also captured: the dark reference was recorded with the camera shutter closed to measure sensor noise, while the white reference was captured from a white calibration target under the same lighting conditions. These references are used to normalize the raw hyperspectral data into reflectance values. The dataset contains normal cork samples and defective samples with real production-related anomalies, including scratches, cracks, holes, water, oil, glue, insects, sticks, plastic, and other foreign materials, making it suitable for evaluating hyperspectral anomaly detection methods for both physical and chemical defects. Here is the setup we used for data colelction:

Data Sample

Here is a sample of the data in specific wavelengths containing different types of defects.

Dataset Structure

This dataset contains train and test collections in which training samples are only normal for defective-free samples while test collection contains both normal and anomalous samples.

Anomalous samples are categorized into 10 different defect types as Crack, Exmaterial, Hole, Glue, Insect, Oil, Plastic, Scratch, Stick, and Water.

CorkHSI/
[Hyperspectral Images of Cork Sheets]
    ├─► Test/
    │   ├─► Normal/
    │   ├─► Crack/
    │   ├─► Exmaterial/
    │   ├─► Hole/
    │   ├─► Glue/
    │   ├─► Insect/
    │   ├─► Oil/
    │   ├─► Plastic/
    │   ├─► Scratch/
    │   ├─► Stick/
    │   ├─► Water/
    ├─►Train/
    │   ├─► Normal/
    └─►Anomaly_Map_Ground_Truth/

The number of samples per each group is summarized in the following table:

Set	Group	No. of Samples
Train	Normal	58
Test	Normal	14
Test	Crack	11
Test	Exmaterial	3
Test	Hole	14
Test	Glue	7
Test	Insect	9
Test	Oil	10
Test	Plastic	8
Test	Scratch	11
Test	Stick	5
Test	Water	11

Sample Descriptions

Each sample is stored in a separate folder containing the hyperspectral image in .raw format, dark and white reference .raw data for normalizing the image, hyperspectral metadata in .hdr format, and also an object mask in .png format which represents the cork sheet area in the image. As an example each sample folder contains the following files:

1/
[A sample folder]
    ├─► 1.hdr
    ├─► 1.raw
    ├─► DARKREF_1.hdr
    ├─► DARKREF_1.raw
    ├─► WHITEREF_1.hdr
    ├─► WHITEREF_1.raw
    └─► 1.png

License

This dataset is licensed under the Attribution–NonCommercial 4.0 International License (CC BY-NC 4.0).