HSI-AgriFoodAnomaly: A Hyperspectral Dataset for Anomaly Detection in the Agri-Food  Industrial Inspection

Ensuring food safety in agri-food production requires rapid and accurate anomaly detection. Hyperspectral imaging (HSI) is a promising modality for this task due to its rich spectral information; however, progress has been limited by the absence of realistic, publicly available datasets. This work introduces \textit{HSI-AgriFoodAnomaly}, the first open-access hyperspectral benchmark specifically designed for foreign objects anomaly detection in conveyor-based agri-food inspection. The dataset comprises 147 pixel-wise annotated HSI-cubes acquired under industrial-like conditions, covering the spectral range from 400 nm to 1000 nm and capturing 300 contiguous spectral bands, featuring a diverse set of anomaly categories including plastics, textiles, metals, glass, and paper. It supports multiple vision tasks such as binary classification, object detection, and semantic segmentation. To validate the dataset’s utility, we propose a baseline detection framework based on 2D convolutional neural networks adapted to full-spectrum HSI-cubes (300 bands). Experimental results demonstrate that models trained on HSI-cubes consistently outperform their RGB images, particularly in detecting subtle or previously unseen anomalies. The proposed dataset and baseline provide a solid foundation for advancing HSI-based food quality inspection systems. The dataset and source code are publicly available at: https://github.com/lsllabisen/HSI-AgriFoodAnomaly-Dataset.git.