Coal Flotation Froth Image Datasets with Controlled Data Leakage

Six image datasets derived from laboratory flotation videos, designed to study the impact of data leakage on deep learning model reliability. The datasets were created by Chunlong Zhang and published on figshare in May 2026. They were used to train and test five deep learning models, including Inception-V3 and Vision Transformer, for flotation concentrate ash recognition.

Use Cases

• Studying the effect of sequence interval-induced data leakage on model accuracy based on the described experimental setup.
• Benchmarking the sensitivity of different deep learning architectures to data leakage based on the five tested models.
• Developing or testing strategies to mitigate data leakage in video-derived image datasets based on the proposed dynamic sampling method.

Strengths

• Includes six datasets with varying, quantified levels of data leakage, allowing for controlled experiments.
• Results include a specific metric showing the proposed mitigation strategy reduced performance inflation by an average of 81.36%.
• Published under a permissive CC-BY-4.0 license, facilitating reuse and sharing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data is sourced from a specific laboratory setting, which may limit generalizability to other industrial environments.

Provenance

Source

Chunlong Zhang via figshare.

Collection Method

Created from laboratory flotation videos, with sequence intervals determining the degree of data leakage between training and test sets.

Freshness

Last updated 2026-05-21 08:48:53; freshness should be verified.