EDGE SYNTHESIS BLOCK: A BUILDING UNIT FOR REAL-TIME SINGLE IMAGE SUPER RESOLUTION

This paper proposes a novel edge synthesis block as the basic building unit of model architectures for single image super resolution task. The edge synthesis block consists of two parallel branches of convolution layers designed to enhance the horizontal and vertical edges in an image. These building units can be stacked flexibly to build complex or light models depending on the desirable latency vs visual quality trade-off. The proposed building block has been tested in mobile phones for different image super resolution tasks such as pinch-zooming of gallery images, and image enhancement in a video call scenario. The edge-synthesis-block-based model architectures are not only very low complex in terms of multiply-accumulate (MAC) operations, but also are comparable to the state-of-the-art super resolution techniques in terms of visual quality.