Texture-Based Video Partitioning via CNN for RDO Efficiency

The patent applies preliminary action by using a Convolutional Neural Network to pre-analyze image data and generate partition probability maps before the actual encoding process. The CNN predicts likely partition boundaries and regions of interest in advance, allowing the RDO process to focus computational resources only on high-probability partition locations rather than evaluating all possible partitions exhaustively.

The patent replaces the traditional mechanical RDO optimization process with a neural network-based probabilistic model. Instead of computationally evaluating rate-distortion costs for numerous partition configurations, the system uses the CNN's learned probability distributions to directly guide partition decisions, substituting iterative optimization with a single-pass probabilistic assessment.

The CNN performs preliminary analysis of the image content to identify probable partition structures before encoding begins. By pre-computing partition probability maps that highlight likely boundaries and uniform regions, the system eliminates the need for time-consuming exhaustive RDO searches during the actual encoding phase.

The patent applies skipping by allowing the encoder to bypass exhaustive RDO evaluation for regions where the CNN predicts low partition probability or uniform characteristics. The system rushes through these low-complexity regions using the pre-computed probability guidance, reserving full RDO processing only for complex, high-uncertainty regions that truly benefit from optimization.