Face Detection in Blurred Surveillance Videos for Crime Investigation

Abstract

Abstract Face detection itself acts as a significant problem in low-resolution surveillance video, primarily due to out-of-focus blur. Real-time and forensic analysis for Law enforcement desires superior face recognition system for person identification in surveillance applications. However, face detection influences face recognition in critical situations. If faces on the blacklist are not detected, recognition fails, which increases the burden and is a tremendous challenge for police authority. This motivates us to present the deblurring algorithm to improve the face detection rate and reduce the false-positive rate. Hence, this paper focuses on removing blur by applying a blind deconvolution algorithm, suppressing the ringing artifact in surveillance video by adopting Discrete Wavelet Transform (DWT). Finally, after this preliminary work, faces are detected by Yolo v2. The proposed framework works well in the sparsely crowded scenario and improves face detection rate compared to the Lucy-Richardson-Yolo v2 framework, which suffers from the problem mentioned earlier. Experiments and evaluations on the surveillance dataset show that the proposed blind deconvolution with ringing suppression-based solution outperforms the state-of-the-art methods in detecting both frontal and profile faces.