Low-Resolution Face Image Reconstruction Using Multi-Stage FSRCNN to Improve Face Detection and Tracking Accuracy in CCTV Surveillance

Face detection and tracking under real-world condition remain challenging under different illumination, crowded scenes, partial occlusions and small or low-resolution face images. In traditional face tracking schemes, these factors often cause the false positive rate to be high and the accuracy to be low. Specifically, little or no detailed information is presented for small or distant faces, here the reliability of detection is diminished and non-face-object can provoke false alarms thus degrading the performance of a system in general. Such problems are not unclear and need a sophisticated solution to improve the resolution and detection performance in various scenarios. In this paper, a new face tracking system based on a cascade classifier, a two-step model of Fast Super-Resolution Convolutional Neural Network (FSRCNN) and DLib face validator is presented. The low-resolution facial parts are first enhanced by the FSRCNN to optimize the detection by the cascade classifier. The DLib face validator improves the approach by validating the discovered faces, and reducing false positives. The system was tested over a CCTV scenario video corpus of several challenging conditions represented by crowded environments, dynamic object and human faces of different sizes and locations. The performance analysis focused on performance metrics such as precision, recall, and false positive rate, which provided a comprehensive overview of the system's robustness. The results demonstrate a significant improvement in face detection accuracy, as high as 98% precision and very few false positive detections. The synergy between the FSRCNN method and the DLib validation was especially effective on small and far-away faces, which are normally difficult to perceive. Whilst their improvements on memory consumption were small, they proved effective for face detection in challenging conditions. The ability of the system to maintain high measurement accuracy while avoiding errors makes it well suited for use in surveillance, security and monitoring systems. In conclusion, this research highlights the effectiveness of combining super-resolution techniques with traditional face detection methods to address the limitations of existing systems. The future work will focus on increasing recall rate and constantly maturing the extraction system to work well in various realistic conditions, thus making it effective and general for different applications.