Classification algorithm analysis for texture detection in block-based hybrid video coding

Abstract

This study presents a comprehensive comparison of various supervised classification algorithms for texture detection in the context of block-based hybrid video coding. To accomplish this, a dataset of images extracted directly from video encoder block partitions was created and manually classified according to their texture levels. The study utilizes the Mean Directional Variance (MDV) algorithm to extract orientation information from each block in the form of average variances for specific rational slopes. This vector of variances is then processed to obtain a set of descriptive statistics that serve as input elements for training and evaluating four popular supervised learning models: Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Supervised Neural Networks (SNN). The objective is to identify the most effective algorithm for accurately classifying texture levels and utilizing this information in perceptual video coding.