Sparse Weight Depthwise Separable with Convolution Neural Network for Video Encoding
International journal of intelligent engineering and systems · 2025
Video encoding, also known as video compression, plays a vital role in enabling the efficient transmission and storage of video content.Effective video compression is essential to produce higher-quality videos because certain video content generates substantial internet traffic that enhances the speed of video transmission.However, video encoding in Coding Unit (CU) partitioning minimizes video efficiency due to increased complexity, which results in higher compression times and a loss of quality.This research proposes Sparse Weight Depthwise Separable with Convolution Neural Network (SWDS-CNN) to encode videos without affecting the efficiency.In traditional CNN, SWDS is incorporated to minimize the number of parameters and computations, which leads to faster processing and lower memory usage.SWDS prunes redundant weights while preserving significant spatial information, which ensures effective feature extraction with minimal redundancy.Its incorporation into CNN ensures adaptability across various video encoding tasks by balancing efficiency and performance.The quadtree is applied in CU partitioning, which enables better adaptation to varying complexity within video frames.The SWDS-CNN achieves a better Bjontegaard Delta Bit Rate (BD-BR) of 1.24% and a Multi-Scale -Structural Similarity Index Measure (MS-SSIM) of -40.56% using JVT-VC and UVG datasets, compared to existing methods like CNN with joint texture recognition and hierarchical random-access coding.