Learning-Based Quality Assessment for Image Super-Resolution

Image Super-Resolution (SR) techniques improve visual quality by enhancing the spatial resolution of images. Quality evaluation metrics play a critical role in comparing and optimizing SR algorithms, but current achieve only limited success, largely due to lack large-scale databases, which are essential for learning accurate robust metrics. In this work, we first build image database using novel semi-automatic labeling approach, allows us label large number images with manageable human workload. The resulting Semi-Automatic Ratings (SISAR), so far largest SR-IQA database, contains 12 600 100 natural scenes. We train an end-to-end Deep (DISQ) model employing two-stream Neural Networks (DNNs) feature extraction, followed fusion network prediction. Experimental results demonstrate that proposed method outperforms state-of-the-art achieves promising generalization performance cross-database tests. SISAR DISQ will be made publicly available facilitate reproducible research.