Nowadays, consumer mobile phones come in with ultra-high resolution cameras, and an incredible number of high-resolution images and videos are created every day. The images and videos have to be downscaled with very large factors to be displayed on general screens.
Deep learning-based downscaling methods show superior performance only for some predetermined integer factors such as 2, 3 and 4. For arbitrary factors, the latest image downscaling algorithms preserve edges and fine details but still suffer from noise amplification. They make undesirable artifacts especially when a downscaling factor is very large.
Kim Su Hyon, a researcher at the Faculty of Information Science and Technology, has proposed an algorithm referred to as NDPID (Noise-free DPID or New DPID) for downscaling ultra-high resolution images to a thumbnail size in real-time without amplifying noise. The proposed algorithm is based on inverse joint bilateral filtering using an area pixel model and moving average.
Unlike the DPID, which employs a rectangular function (box filtering) as the spatial kernel, the NDPID uses two-step 1D APID (Area Pixel model based Image Downscaling) filter. The main reason for employing this 1D spatial kernel is to decompose the proposed downscaling algorithm into two subsequent processes each of which performs capturing pixels’ distinctness for their weights and smoothing of the weights. By these two processes, the algorithm alleviates an isolated noise pixel twice but a thin line (important detail) only once. Consequently, the lines and edges survive while the NDPID alleviates the isolated noise pixel in both horizontal and vertical smoothing processes.
The proposed algorithm is much faster than state-of-the-art downscalers and is free from the restraints of predetermined integer downscaling factors. The experimental results show that the proposed algorithm is about 7.37% faster on average than the DPID, the fastest detail-preserving image downscaler in use. GPU implementation of the algorithm downscales a 2K video to 128-pixel width without temporal artifacts at the speed of 116 frames per second. Moreover, the PSNR and SSIM scores achieved by his method were respectively 35.9% and 16.5% higher on average than the highest values scored by the existing methods when downscaling images contaminated by 5% salt and pepper noise.
If further information is needed, please refer to his paper “A New Rapid and Detail-Preserving Image Downscaling Without Noise Amplification” in “IEEE Access” (SCI).
© 2021 KumChaek University of Technology