Efficiency Across Datasets: Insights into Video Compression Techniques

In recent times, there has been notable progress in deep learning-based image compression, showcasing enhanced coding efficiency and subjective quality. However, comparatively less attention has been devoted to video compression employing deep neural networks. This paper introduces DeepLVCnet is a deep predictive video compression network that operates from end to end. With multiple framework theories, multiple scale arrangements, and a temporal context-adaptive entropy model, the method makes use of mode-selective uni-and bi-directional predictions. DeepLVCnet jointly compresses motion information and residual data through feature transformation layers within the multi-scale structure. A mode-selective architecture with uni-directional and bidirectional predictive modes is incorporated into DeepLVCnet, in contrast to earlier deep learning-based video compression techniques limited to P-frame or B-frame usage. Furthermore, utilizing temporal context information from reference frames during current frame coding, a unique temporal-context-adaptive entropy model is developed. This model is designed to enable parallel processing, offering computational and architectural advantages over autoregressive entropy models used in some CNN-based video compression methods. The overall objective is to enhance coding efficiency and subjective video enlargement quality through these innovations.