This bachelor's thesis uses neural networks to compress video. Due to improvements in deep learning, a new idea appeared. Neural networks can learn to compress image and video data using large training sets and appropriate architecture. In the thesis, we used convolutional autoencoders that can transform input data into smaller latent space. We present two approaches to compression. The first one is designed to compress images, while the second is improved to compress video material. It is based on the classic approach of predicting movement in a scene and has error correction. We described used architectures and processes of learning and testing. We focused more on a quantization operation which is an important element for controlling compression ratio and quality. We evaluated the first approach and compared it with the JPEG image compression format. We chose two different configurations for the second approach, tested them using multiple parameters, and compared results with performances of standard codecs. Although both approaches are capable of efficient compression, they can not compete with today's standards. Because of this, we also mentioned some novelties that could significantly improve performance.
|
