RBDN: Residual Bottleneck Dense Network for Image Super-Resolution

Recent studies have shown that a super-resolution generative adversarial network (SRGAN) can significantly improve the quality of single-image super-resolution. However, existing SRGAN methods also certain drawbacks, such as an insufficient feature utilization, large number parameters. To further enhance visual quality, we thoroughly studied three key components SRGAN, i.e., architecture, loss, and perceptual propose DenseNet with Residual-in-Residual Bottleneck Block (RRBB), called residual bottleneck dense (RBDN), for First, to utilization features between various layers network, adopted cascading connection layers. At same time, reduce computational cost, added structure each layer, greatly reducing parameters accelerating convergence speed training process. Second, proposed RRBB, basic building unit, removes batch normalization (BN) layer employs ELU function opposite effects in absence BN. In addition, applied improved overall loss during model process stably train realism reconstructed high-resolution image. prove superiority our model, conducted comprehensive objective evaluation Peak Signal-to-Noise Ratio, structural similarity, learned image patch other indicators obtained from test sets, Set5, Set14, BSD100, recent state-of-the-art model. Finally, qualitative quantitative analyses results terms indicators, authenticity restored HR images, textural details, which show RBDN