基于对偶学习的密集连接超分辨率网络

Dense Connected Super-Resolution Networks Based on Dual Learning

摘要: 深度学习的方法大大提升了超分辨率技术的性能，但是由于超分辨率重建本身是一个不适定的工作，导致生成的效果不稳定，且对于不同通道的平均化处理也会导致图像的特征难以凸显。本文使用了对偶学习的策略，通过引入SR图像下采样的子任务，减小解空间，进一步约束超分图像的生成过程；同时使用了密集连接的残差通道注意力模块，使图像的特征有所侧重的同时，又能较好地传递。最后通过与经典算法的对比和消融实验，验证了本方法具有较好的生成效果，且我们的改进也切实有效。

Abstract: The deep learning method greatly improves the performance of super-resolution. However, since the super-resolution reconstruction itself is an illposed work, the generated effect is unstable, and the averaging processing for different channels will also cause the image features to be difficult to highlight. In this paper, the dual learning strategy is used to reduce the solution space and further constrain the SR image generation process by introducing the sub task of SR image downsampling; at the same time, the Residual Dense Channel Attention Blocks are used, so that the image features are focused and can be better transmitted. Finally, through the comparison with the classical algorithm and the ablation experiment, it is verified that this method has a better generation effect, and our improvement is also effective.