Abstract
Recent random-forest (RF)-based image super-resolution approaches inheritsome properties from dictionary-learning-based algorithms, but theeffectiveness of the properties in RF is overlooked in the literature. In thispaper, we present a novel feature-augmented random forest (FARF) for imagesuper-resolution, where the conventional gradient-based features are augmentedwith gradient magnitudes and different feature recipes are formulated ondifferent stages in an RF. The advantages of our method are that, firstly, thedictionary-learning-based features are enhanced by adding gradient magnitudes,based on the observation that the non-linear gradient magnitude are with highlydiscriminative property. Secondly, generalized locality-sensitive hashing (LSH)is used to replace principal component analysis (PCA) for featuredimensionality reduction and original high-dimensional features are employed,instead of the compressed ones, for the leaf-nodes' regressors, sinceregressors can benefit from higher dimensional features. Thisoriginal-compressed coupled feature sets scheme unifies the unsupervised LSHevaluation on both image super-resolution and content-based image retrieval(CBIR). Finally, we present a generalized weighted ridge regression (GWRR)model for the leaf-nodes' regressors. Experiment results on several publicbenchmark datasets show that our FARF method can achieve an average gain ofabout 0.3 dB, compared to traditional RF-based methods. Furthermore, afine-tuned FARF model can compare to or (in many cases) outperform some recentstateof-the-art deep-learning-based algorithms.