In response to the low recognition rate of smuggled dangerous goods in real security check scenarios and the shortcomings of traditional image classification and existing deep learning algorithms in the field of security check, effective recognition algorithms are proposed to improve security check efficiency and accuracy, and ensure passenger safety. This article uses deep convolutional neural networks to propose a deep learning network-based algorithm for identifying dangerous goods in human body smuggling and security checks. By combining infrared images with VI, using image segmentation and unsupervised registration techniques, a self-encoding deep learning network is used as the registration network skeleton for human contour registration learning. The network is optimized by measuring the similarity of the registered images and minimizing the similarity cost. The experimental results show that, compared with other algorithms on the public dataset SIXray, our algorithm achieves an average accuracy (mAP) of 93.21%, which is higher than other compared algorithms, especially 38.7 percentage points higher than Inception V3, with small increases in parameter quantity and model size and almost unchanged detection time. The algorithm in this article performs excellently in terms of parameter quantity, model size, detection time, and average accuracy, and has higher feasibility and practicality, providing efficient and accurate solutions for security inspection work.
