Semi-Supervised Concealed Object Detection Via Mixed Pseudo Labels

Deep neural networks have been widely applied and extensively studied in millimeter-wave image concealed object detection, achieving remarkable success. However, deep supervised neural networks heavily rely on a large amount of high-quality labeled data, the acquisition of which requires significant human effort and time, making it challenging to obtain. In contrast, unlabeled data is relatively easy to collect. Considering this limitation, We introduce a semi-supervised object detection model—Mixed Pseudo Labels (MixPL), which is built upon the teacher-student paradigm and incorporates two novel methods, Pseudo Mixup and Pseudo Mosaic—and experimentally evaluate its performance on millimeter-wave image datasets for concealed object detection in security screening. Compared to fully supervised learning, MixPL significantly improves model performance across different proportions of labeled data. Notably, under the 20% labeled data setting, MixPL achieves a remarkable improvement of 45.9%.