Deep CNN Based Video Compression With Lung Ultrasound Sample

10.6180/jase.202303_26(3).0002

Helen. K. Joy；Manjunath R Kounte

CNN ； Motion estimation ； COVID-19 ； P-frame ； B-frame

淡江理工學刊

26卷3期（2023 / 03 / 01）

313 - 321

英文

Video compression and transmission is an ever-growing area of research with continuous development in both software and hardware domain, especially when it comes to medical field. Lung ultra sound (LUS) is identified as one of the best, inexpensive and harmless option to identify various lung disorders including COVID-19. The paper proposes a model to compress and transfer the LUS sample with high quality and less encoding time than the existing models. Deep convolutional neural network is exploited to work on this, as it focusses on content, more than pixels. Here two deep convolutional neural networks, ie, P(prediction)-net and B(bi-directional)-net model are proposed that takes the input as Prediction, Bidirectional frame of existing Group of Pictures and learn. The network is trained with data set of lung ultrasound sample. The trained network is validated to predict the P, B frame from the GOP. The result is evaluated with 23 raw videos and compared with existing video compression techniques. This also shows that deep learning methods might be a worthwhile endeavor not only for COVID-19, but also in general for lung pathologies. The graph shows that the model outperforms the replacement of block-based prediction algorithm in existing video compression with P-net, B-net for lower bit rates.