IVDETR: A Multi-Source Image Pig Detection Method Based on Cross-Domain Feature Fusion

In large-scale centralised pig farming, object detection algorithms can accurately track and monitor each pig. However, a significant proportion of current detection methods are based on visible images, and the colour differences between different pig breeds limit the applicability of the models. Furthermore, complex background interference, low illumination conditions, or high illumination intensity noise in visible scenes can likewise lead to reduced accuracy during detection. Therefore, we propose a cross-domain encoder based on the self-attention and design the IVDETR model, which integrates features from infrared and visible light images for pig object detection. In order to facilitate the training and evaluation of the model, we constructed a comprehensive, standardised infrared-visible light dataset for the first time. The experimental outcomes demonstrated that the detection accuracy was increased from 0.939 to 0.998 in a real pig breeding environment.