基于高阶交互的渐进式真实图像去噪网络

Progressive Real-word Image Denoising Network Based on Higher-order Interactions

针对真实图像噪声水平未知、噪声分布复杂等特点,提出一种多尺度渐进式去噪算法,进一步提升针对真实图像去噪算法的泛化性和鲁棒性,有效去除噪声并保留纹理细节,在保障空间精度的同时丰富语义信息。该模型整体遵循多尺度结构,并行的3条卷积流表现为3个子网,每个子网采用单一尺度通道,子网间进行拼接,保证信息最大限度地传递。最后子网的输出通过改进深度卷积块实现多尺度特征再利用,并与该子网原始特征进行融合以达到更高效的去噪。通过实验证明,该算法在SIDD、DND、PolyU数据集上的峰值信噪比分别达到38.69、39.12、37.24 dB,另外结构相似性、图像通用质量指标和视觉信息保真度指标表现优异,可以验证该算法不论在定量或定性分析上都有良好的表现,达到较高的性能。

For the characteristics of real-world image with unknown noise level and complex noise distribution,a multi-scale progressive denoising algorithm was proposed to further improve the generalization and robustness of the denoising algorithm for real-world images,effectively removing the noise and preserving the texture details,and enriching the semantic information while ensuring the spatial accuracy.The model as a whole is followed a multi-scale structure,where the three convolutional flows are represented in parallel as three sub-networks,each with a single-scale channel,and concatenation was performed between the sub-networks to ensure maximum information transfer.The output of the final sub-network was achieved by refining the deep convolutional blocks for multi-scale feature reuse,and fused with the original features of this sub-network for more efficient denoising.Through experiments,it is proved that the peak signal-to-noise ratio of this algorithm on SIDD,DND,and PolyU datasets reaches 38.69,39.12,and 37.24 dB,respectively.In addition,the performance of the algorithm in terms of structural similarity,general image quality indicators,and visual information fidelity indicators is excellent,which can verify that the algorithm has achieved high performance in both quantitative and qualitative analysis.