中国科学院云南天文台机构知识库

多种退化类型混合的图像比单一类型的退化图像降质更严重， 很难建立精确模型对其复原， 研究端到端的神经网络算法是复原的关键. 现有的基于操作选择注意力网络的算法（operation-wise attention network， OWAN）虽然有一定的性能提升， 但是其网络过于复杂， 运行较慢， 复原图像缺乏高频细节， 整体效果也有提升的空间. 针对这些问题， 提出一种基于层级特征融合的自适应复原算法. 该算法直接融合不同感受野分支的特征， 增强复原图像的结构; 用注意力机制对不同层级的特征进行动态融合， 增加模型的自适应性， 降低了模型冗余; 另外， 结合L₁损失和感知损失， 增强了复原图像的视觉感知效果. 在DIV2K， BSD500等数据集上的实验结果表明， 该算法无论是在峰值信噪比和结构相似性上的定量分析， 还是在主观视觉质量方面， 均优于OWAN算法， 充分证明了该算法的有效性.

The degradation of mixed degraded images is more serious than that of single degradation types, and it is difficult to restore them by precise modeling. The key to restore mixed degraded images is to study the
end-to-end neural network algorithm. Although the existing operation-wise attention network (OWAN) algorithm has a certain performance improvement, its network is too complex, it runs slowly, the restored image lacks high-frequency details, and the overall effect also has room for improvement. To solve these problems, an adaptive restoration algorithm based on hierarchical feature fusion is proposed. The algorithm directly fuses the features of different receptive field branches to enhance the structure of the restored image. The attention mechanism is used to dynamically fuse the features of different hierarchies to increase the adaptability and reduce the redundancy of the model. In addition, combining the 1 L loss and perception loss, the visual perception effect of the restored image is enhanced. Experimental results on DIV2K, BSD500 and other data sets show that the proposed algorithm is better than the OWAN algorithm in terms of quantitative analysis of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), as well as subjective visual quality.