Deep Convolutional Neural Network Driven Interpolation Filter for High Efficiency Video Coding
Helen K. Joy, Manjunath R Kounte · Mathematical Modelling and Engineering Problems · 2024
Research in video coding has gained significant importance in recent years, driven by the increasing demand for multimedia transmission.High Efficiency Video Coding (HEVC) has emerged as a prominent standard in this field.Interpolation is a crucial aspect of HEVC, particularly when using fixed half-pel interpolation filters derived from traditional signal processing techniques.In recent times, there has been an exploration of interpolation filters that are based on Convolutional Neural Networks (CNNs).Conventional signal processing techniques are used in traditional HEVC methods to employ fixed half-pel interpolation filters.Recent advancements have delved into the application of Convolutional Neural Networks (CNNs) to enhance interpolation performance.Our proposed method utilises a sophisticated CNN architecture specifically crafted to extract valuable features from low-resolution image patches and accurately predict high-resolution images.The network consists of multiple layers of CNN blocks, which utilise 1×1 and 3×3 convolutional kernels to enable efficient and thorough feature extraction through parallel processing.This architecture improves computational efficiency and greatly enhances prediction accuracy The suggested interpolation filter shows a 2.38% enhancement in bitrate savings, as evaluated by the BD-rate metric, specifically in the low delay P configuration.This highlights the potential of deep learning techniques in improving video coding efficiency.