Comparison of machine learning-based feature pooling strategies for colour image fidelity assessment

摘要

Current models of subjective image fidelity assessment rely on two steps: feature extraction, which is usually inspired by early human vision (the eye), and pooling, which on the contrary often fails to represent the complex mechanisms of late vision (the brain). This study evaluates the feasibility of using a shallow machine learning-based model to mimic the latter. We look at how three types of machine learning algorithms (shallow neural network, support vector regression and random forest) can improve feature pooling for three of the best-performing colour image fidelity assessment indices. We demonstrate that, with only three early vision features and a shallow machine learning unit, we can achieve a prediction accuracy that is comparable to that of more complex models, based for instance on deep learning.