Mixed Global and Local Attention Alleviates Domain Shift Between Terahertz Image Datasets

Terahertz imaging technology has a broad application prospect in the field of security inspection due to its strong penetrability and negligible radiation effect on people. With the development of machine learning in recent years, the application of machine learning methods to security inspection can save labor costs and ensure the accuracy of long-time security inspection. However, terahertz imagers produce different bottom noise at different temperatures, which leads to a significant domain offset problem between data sets collected at different temperatures. This domain shift makes the model trained on the source domain dataset to have a significant error when applied to the target domain dataset. Therefore, how to overcome the temperature-induced domain offset is an urgent problem for terahertz imaging techniques. In this paper, we innovatively propose to apply the attention mechanism to solve the problem of domain bias in object detection. We propose a lightweight attention mechanism Mixed global and local attention(MGLA) aimed at mitigating the domain shift between the target and source domains without affecting the source domain detection effect. MGLA can take into account the local information while fusing the channel information and spatial information and global information to improve the representation of the network. The experiment results show a 10% improvement in mAP0.5 and a 6.3% improvement in mAP50:95 compared to YOLOV8 baseline.