Spherical Rotation for Omnidirectional Video Coding Based on Texture Analysis

As the core of virtual reality and augmented reality technologies， omnidirectional video offers a 360° full-coverage perspective. However， this immersive experience generates massive data volumes， posing severe challenges in terms of storage and transmission. Therefore， optimizing the coding efficiency of omnidirectional video， while balancing the associated coding complexity， is crucial for advancing its widespread application in fields such as healthcare and education. To adapt to existing coding frameworks， omnidirectional video must be projected from a three-dimensional sphere onto a two-dimensional plane， which inevitably leads to oversampling and discontinuous boundaries， introducing distortion that degrades the coding efficiency. Rotating the spherical domain can mitigate such distortion by optimizing the texture distribution and thereby improving the coding performance. Based on this， in this work， we conduct an in-depth study on spherical-domain rotation for omnidirectional video. Through a systematic comparison of the encoding-rate-distortion performance， the equiangular cubemap is identified as the basic projection format for the subsequent experiments. Exhaustive spherical-rotation traversal experiments are performed to analyze the correlation between rotation angles and encoding performance. The results confirm the feasibility of enhancing the encoding performance through spherical rotation and establish the existence of an optimal rotation angle. Based on these findings， this work proposes an efficient spherical-rotation algorithm leveraging texture characteristics. The algorithm first extracts texture features from projection planes using the Scharr gradient operator. Subsequently， it determines the dominant texture orientation through hierarchical screening. This mechanism operates in two stages： The first stage employs dynamic thresholding to filter out flat blocks and exclude entirely flat planes， based on the proportion of valid blocks. The second stage introduces histogram envelope analysis to eliminate planes with disordered texture orientations by examining the peak-valley characteristics of the gradient-angle distribution. Finally， the dominant orientation is determined based on the weighted gradient-angle statistics of the remaining valid planes. Experimental results show that， by rotating the identified dominant texture orientation to the horizontal direction， the proposed algorithm achieves up to 3.15% bitrate savings for specific sequences， with an average bitrate saving of 0.56%， while introducing almost no additional encoding complexity. Furthermore， when combined with an optimal rotation-axis-prediction algorithm， the average bitrate saving reaches 0.66%.