Hybrid network with difficult–easy learning for concealed object detection in imbalanced terahertz image dataset

Terahertz imaging technology has been widely used in security inspections due to its ability to detect various concealed hazardous materials and the advantage of being harmless to the human body. However, limited by the terahertz imaging system, it is challenging to detect concealed objects due to hard samples and imbalanced categories caused by terahertz image quality. To solve these issues, we propose a hybrid network with difficult–easy learning (DEL) for concealed object detection in the imbalanced activated terahertz image dataset. Based on the one-stage framework YOLOv5m, a path aggregation hybrid structure (PAHS) is proposed to improve the performance of the proposed network while maintaining real-time detection. Specifically, PAHS with transformer block (TB) and a fine-tuned global context attention (GCA) are designed to fully exploit and fuse the multi-scale information by path aggregation, which improves the detection accuracy of low contrast and noise-interfered objects. To solve the problem of imbalanced categories in the activated terahertz dataset, a DELoss is developed to guide the network classification. Moreover, EIOU is adopted to boost the network training, and a modified B-Ocl loss is used to discriminate the positive and negative samples. Experiments are conducted on a public imbalanced activate terahertz image dataset. The experimental results illustrate that the proposed network achieves competitive performance compared with recently reported state-of-the-art detection methods. Moreover, the proposed method improves the balanced detection ability of different categories.