BYNET:IMAGE SUPER RESOLUTION WITH A BYPASS CONNECTION NETWORK
This paper proposes a deep residual network, ByNet, for the single image super resolution task. The main innovation is the introduction of two effective components, bypass connections and a feature scaling layer. Bypass connections are formed either by skip connections that jump multiple layers or by adding a convolution layer in such a jump. The ﬁnal feature scaling layer enables more robust convergence. Experiments on standard benchmarks show that the proposed method achieves state of the art results over multiple scales in terms of PSNR and structural similarity (SSIM).
Index Terms— super resolution, deep convolutional neural networks, residual learning, image enhancement
Research Areas : #Vision Program
Careers : Open Positions