CNN with spatio-temporal information for fast suspicious object detection and recognition in THz security images

Abstract

Abstract Owing to low radiation required for human body, complete detection of hazardous materials, noncontact privacy protection, and fast pass capability, terahertz (THz) security screening cameras (TSSCs) are widely deployed in public places. However, the existing TSSCs only report about the rough locations of suspicious objects, while directing the complicated detailed recognition work to security inspectors, thus leading to low recognition efficiency. In this paper, we aim to introduce the artificial intelligence technique, called convolutional neural network (CNN), with the spatio-temporal information of THz security image sequences, to achieve automatic object detection and recognition. Our method is composed of rough detection and detailed recognition. In rough detection, benefiting from the inherent alignment of human parts, the sparse and low-rank decomposition (SLD) is used to excavate the spatio-temporal context information. Specifically, the low-rank part representing the static background is regarded as the human body, while the sparse part representing the displaced target is regarded as the suspicious object. Then, by considering the shape knowledge and performing morphological processing, noise interference was reduced and the rough locations of suspicious objects were determined. In detailed recognition, supervised training was first conducted based on Faster R-CNN model with large-scale object labels. The trained Faster R-CNN could extract high-level semantic features of each anchor, and thus predict the class attribute of each object. Notably, with the results of rough detection, only anchors in the candidate domain are computed. Compared with the conventional full domain computation, this narrow-band approach not only reduces the computational complexity but also decreases the false positives caused by anchors in the background. Extensive experiments were conducted on THz security images, and the results prove that our method achieves high performance with respect to both accuracy and efficiency.