Sparse representation super-resolution method for enhancement analysis in video forensics

Abstract

The enhancement analysis in video forensics is used to enhance the clarity of video frames of a video exhibit. The enhanced version of these video frames is important as to assist law enforcement agency for investigation or to be tended as evidence in court. The most significant problem observed in the analysis is the enhancement of objects under probe in video. In many cases, the probes appeared to be in low-resolution and degraded with noise, lens blur and compression artifacts. The enhancement of these low quality probes via conventional method of denoising and resizing has proven to further degrade the quality of the prober The objective of this paper is to propose an enhancement analysis algorithm based on super-resolution. Hence, we present an solution which is a single-frame solution for super-resolution. For that purpose, our proposed method incorporates sparse coding with Non-Negative Matrix Factorization in order to improve hallucination of probes in video. Sparse coding is employed in learning a localized part-based subspace which synthesizes higher resolution with respect to overcomplete patch dictionaries. We test our proposed method and compare with state-of-the-art methods namely resampling and super-resolution method, by enhancing probes in exhibit videos. We measure the image quality using peak-signal-to-noise-ratio. The result shows that our proposed method outperforms state-of the-art methods after enhancing probes in exhibit videos.