Object Detection of Tobacco-related Information Based on Visual Features

The forms of tobacco-related crimes are changeable and highly concealed, resulting in a waste of public security police resources. Tobacco-related images are unstructured, high-dimensional, and diverse. Therefore, this paper establishes a database of tobacco-related cases, and implements a target detection system for tobacco-related information. The system is divided into two parts: software and algorithm. The software part realizes the transmission of unformatted data, and detects the massive tobacco-related data on the cloud server to complete end-to-end case tracking. In the algorithm part, the detection and recognition of tobacco-related images, license plates, and text are improved. We propose Yolox-Cr, LPRNet-St, DBNet-lite corresponding tobacco-related target detection and recognition algorithms. In the three self-built datasets, the recognition accuracy of each algorithm for the corresponding tobacco-related targets exceeds 85%. The intelligent object detection system for smoking-related cases has been successfully applied to the analysis and judgment of smoking-related cases by a provincial public security department, improving the efficiency of law enforcement.