Researchers at the UNIST Graduate School of Artificial Intelligence in South Korea have developed an advanced AI model «BF-STVSR» capable of simultaneously improving video resolution and frame rate.
The researchers, led by Professor Jaejun Yu, first author Eunjin Kim and co-authors Hyunjin Kim presented an improved «BF-STVSR» model that effectively increases both video resolution and frame rate simultaneously. Traditional methods for improving video quality with the help of AI usually involve increasing the resolution and frame rate separately from each other.
To a large extent AI models rely on pre-trained optical flow prediction networks. Optical flow calculates the direction and speed of an object in the video to generate intermediate frames. However, this method involves complex computations and is vulnerable to the accumulation of numerous errors, which limits both the speed and quality of video enhancement.
«BF-STVSR» uses signal processing methods adapted to for a specific video. This allows the model to learn bidirectional movement between frames regardless of external optical flow networks. Вy inferring the contours of the object and the flow of motion, the model effectively increases both resolution and frame rate simultaneously, resulting in a more natural and consistent video reconstruction.
Applying this AI model to low-resolution and low-frame rate video demonstrated superior performance compared to existing models, as evidenced by higher peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) scores. These metrics demonstrate that even fast-moving videos retain clear, undistorted human figures and other details.
«This technology has a wide range of applications, from recovering CCTV footage or black box recordings captured on low-cost devices to rapidly enhancing compressed streaming videos of high-quality media content. It can also benefit fields such as medical imaging and virtual reality», — explains Professor Jaejun Yu.
The results are presented on the preprint server arXiv
Source: TechXplore
Spelling error report
The following text will be sent to our editors: