Neural Networks for Image and Video Compression
Artem Gorodilov, Dmitriy Gavrilov, Dmitriy Schelkunov · 2018
The data transition has become one of the most actual problems nowadays. Classical methods of video and image compression give great results but refers huge computational resources. One of the improvement ways in video and still images processing and transmitting is using neural networks. In this field they can be used for pre-/post-processing, algorithm optimization, segmentation tasks. There is a lack of studies which pay attention to neural network application for reducing spatial and temporal redundancy in video encoding operations. In this work we propose new methods of preprocessing and postprocessing using convolutional neural networks along with classical algorithms of video compression. The main structure consists of downscale and upscale parts which allows to downsize a transmitted image or frame and reconstruct it in decoder side with high accuracy. The reconstruction process uses preprocessing before upscale operation and post-processing after image size restoration. The current algorithm works in conjunction with the classic video codec H.264/ A VC. The work's results show improvement of the transmitted image quality at a lower bit rate. The new method improves previous ones in term of compression rate.