Single image super-resolution via deep progressive multi-scale fusion networks

Deep convolutional neural network-based single-image super-resolution (SR) models typically process either upsampled full-resolution or original low-resolution features, which suffer from context lack and spatially imprecision, respectively. To solve this, we propose a novel progressive SR network to preserve spatial precision through the original resolution and to receive rich contextual information from low-to-high resolution representations. Our proposed progressive, selective scale fusion network includes four key points: (a) parallel multi-scale convolution branches to extract multi-scale features, (b) information exchange across the multi-scale branches, (c) attention mechanism-based multi-scale feature fusion, and (d) gradual aggregation of multi-scale streams from low-to-high resolutions. The proposed method learns hierarchical features that aggregate contextual information from different resolution streams while maintaining high-resolution spatial details. Both quantitative and qualitative experiments on benchmark and real-world datasets show that our method offers a favorable performance against state-of-the-art methods for SR tasks with different scaling factors.