A Forensic Video Upscaling Colorizing and Denoising Framework for Crime Scene Investigation

Digital videos have been widely used as key evidence sources in Forensic crime scene investigations. Resolution is one of the most dominating parameter which affects the overall quality of the video. The main goal of this paper is to find an efficient forensic video analysis framework to assist the forensic crime scene investigation. A forensic video analysis framework (FVAF) that employs an efficient video enhancing deep learning model for increasing resolution of the low quality videos is used. The low resolution video is fed as input to the model. First, the video is pre-processed using fastai deep learning library. Large videos are cropped to manage runtime efficiently. Second, the video is rescaled for increasing the resolution by Spatial Resolution method. The framework successfully increases the resolution of the video from SD-standard definition Resolution type of 480p with Aspect Ratio 4:3 of Pixel size 640 × 480 to Full Ultra HD Resolution type of 8K or 4320p with Aspect Ratio 16∶9 of Pixel Size 7680 × 4320. The rescaled videos are submitted for colorization process. DeOldify deep learning model using Self-Attention Generative Adversarial Network and Two Time-Scale Update Rule is adopted by FVAF framework for colorizing the videos. Also, the colorized videos are trained and tested by various video enhance AI models model Gaia High Quality 4K rendering and Theia fine Tune detail. 4K not rendered and Theia Fine Tune Fidelity: 4K not rendered and video denoise AI models model Standard, clear, lowlight, severe noise and Raw. The upscaled and colorized video is also trained and tested using denoise video enhance AI and video denoise AI models. The results of each model are stored for comparison. From the stored results best video enhance AI model and the best video denoise AI models is selected. Lowlight AI model and Gaia high quality 4K rendering are used in this FVAF to produce high standard video for Forensic Analysis. We run this model using GPU to efficiently pre-process the video. By this framework, we increase the resolution of the video footages to further assist the forensic crime investigation.