Deep Learning Training Dataset for Multimodal Images of Individual Dairy Cow Heads

The dataset contains 2,636 pairs of high-quality registered VIS (Visible) and IRT (Infrared Thermography) images, covering both daytime and nighttime scenes. The images are in PNG format, and the labels are in YOLO format. The data is divided into training, validation, and test sets, with each part containing infrared images, visible light images, and head annotation files. The dataset was collected in September 2024 at a dairy farm in Changping District, Beijing, and includes images of natural activities from 50 dairy cows.The data was captured using synchronized visible light cameras (1080p resolution) and infrared cameras (384×512 resolution, temperature measurement accuracy ±0.5°C). The cameras were mounted on a 1.5-meter high platform, and the shooting distance was 1.5-2 meters. The collected videos were optimized through image rectification, cropping, blur detection, and quality assessment to generate pairs of visible (VIS) and infrared thermography (IRT) images.Data augmentation techniques (such as rotation, translation, and scaling) were applied to the training set, with each image augmented three times to increase diversity. Low-quality and duplicate frames were removed using blur detection and SSIM evaluation. The infrared temperature measurement error is ±0.5°C, and the image registration error was minimized through orthogonal correction.This dataset provides high-quality support for cross-modal learning, image pairing, and precision agriculture research. It can be used for dairy cow health monitoring, disease early warning, and behavioral analysis.