FastRWDnet: implementation of novel real-time deep video denoising utilizing optimized FastDVDnet

Every day, huge amounts of video are created and stored, yet many of these are inappropriate for deep learning or other automation-related tasks due to being noisy and poor quality. Although digital video processing has many uses, little research or software is produced. Until recently, video denoising with neural networks was a relatively unexplored field. FastRWDnet, the solution we present in this research, achieves comparable results to existing state-of-the-art FastDVDnet approach while requiring much less processing time. In comparison with other existing neural network denoisers, our technique has several advantageous qualities, including short run-times and the ability to handle a wide variety of noise levels with a single network model. In order to achieve faster segmentation, the novel technique of FastRWDnet uses modified bottleneck blocks of ENet for denoising purposes instead of UNet, which was employed in the FastDVDnet. Due to the architecture’s qualities, it is possible to run this algorithm in a real-time implementation. It is suitable for practical denoising applications due to its combination of high denoising performance and minimal computational load. We have implemented our algorithm on a web architecture and successfully obtained real-time denoising output for both prerecorded video footage and real-time video streams.