Data Compression Frameworks using RNNs and Deep Generative Models

Abstract

Data Compression is an essential technique for reducing the size of the digital data and enabling more efficient storage and transmission. Traditional data compression tech niques have been based on signal processing and statistical modeling, but recent advances in deep learning have opened up new possibilities for developing more effective and ef ficient compression algorithms. This thesis explores the use of deep learning for data compression, focusing on text and video compression using recurrent neural networks (RNNs) and variational autoencoders (VAEs). The main goal of the research is to de velop a new compression framework that can achieve higher compression rates while maintaining good reconstruction quality. The first part of the thesis presents a comprehensive review of the literature on data compression and deep learning, covering the main concepts, techniques and applications. This includes a detailed overview of the key deep learning models and architectures used for data compression, such as RNNs, and generative models like VAEs. The second part of the thesis presents the proposed compression algorithms. The RNN-based approach to compress text uses the concepts of probability predictor and arithmetic coder. The VAE based algorithm uses a probabilistic approach to model the data distribution and generate compressed representations. The third part of the thesis evaluates the performance of the proposed algorithms using a range of standard benchmarks and datasets. The results show that the proposed algorithm can achieve compression rates which match the performance of exisiting codecs, while maintaining decent reconstruction quality. The proposed framework addresses the challenges associated with video compression by designing an innovative deep learning architecture that captures temporal dependencies in videos to achieve higher compression ratios without compromising visual quality. A large-scale video dataset is curated to train the framework, ensuring its generalizability to various video content, resolutions, and encoding formats. Evaluating the perceptual quality of the reconstructed videos is crucial, and this thesis uses evaluation metrics like compression rate, PSNR and conducts subjective and objective assessments. A com parative analysis is performed against existing video compression standards to assess the framework’s performance and effectiveness. Real-time processing and deployment scenarios are considered, with a focus on optimizing computational requirements and limitations. Overall, the research presented demonstrates the potential of deep learning for data com pression and provides new insights into the design and optimization of compression algo rithms using RNNs and VAEs. The findings are expected to have significant implications for various applications of data compression, including video, text, network compression and data storage where efficient compression is vital for seamless user experiences and resource optimization. The developed framework enables efficient storage, transmission, and processing of video data while preserving a decent-quality visual information.