Video Compression Method Using Vector Quantization

As the number of users viewing video content on the Internet increases, video traffic is growing rapidly. When viewing video over the Internet, poor communication conditions and degraded viewing quality cause video and audio to be halting during playback due to the increased time needed for processing the playback of video content. In order to suppress such degradation in viewing quality, many researchers have proposed video compression models that leverage machine learning as a method to transmit high-quality video using fewer bits. These studies use convolutional neural networks to reduce the dimensionality of the image data. They also use scalar quantization and vector quantization to convert continuous values into discrete values. Previous research works have proposed video compression models using only scalar quantization, but not vector quantization. In this paper, we propose a video compression method using vector quantization and evaluate the video compression model. The proposed method reduces the processing time of video compression by fine-tuning the next frame using the weights of the fine-tuned frame. Therefore, the proposed method can fine-tune more frames within a given time. From the evaluation of the video compression model using the proposed method, the image similarity is higher than that of a video compression model using a method without fine-tuning.