用于去除随机脉冲噪声的两阶段盲卷积降噪模型

Learning a Two-Stage Blind Convolutional Denoising Model for the Removal of Random-Valued Impulse Noise

相对于经典的采用逐点检测与复原方式实现的开关型随机脉冲噪声(Random-Valued Impulse Noise,RVIN)降噪算法,基于深度卷积神经网络构建的非开关型RVIN降噪模型在降噪效果和执行效率上均有显著优势,但也存在着固有的数据依赖缺陷,不能在降噪效果和易用性两个方面同时获得最佳性能.为此,以DnCNN(Denoising Convolutional Neural Network)深度降噪网络模型架构为设计基础,提出了一种新的用于去除RVIN噪声的两阶段盲卷积降噪(Two-stage Blind Convolutional Denoising,TBCD)模型.在第一阶段,针对给定的受0~90%范围内某个比例RVIN噪声干扰的噪声图像,利用DnCNN-B(DnCNN for Blind Denoising Task)盲降噪模型完成初步降噪.同时,利用噪声检测模型预测出噪声图像相应的噪声标签,然后将噪声图像与噪声标签矩阵按位相乘生成稀疏采样图像(Sparse Sampling Image,SSI).在第二阶段,为了进一步提高DnCNN-B盲降噪模型所复原的初步降噪图像的质量,将其与SSI图像连接(concatenate)后再次输入到预先训练好的双通道图像质量提升模型中获得残差图像,之后将初步降噪图像减去残差图像得到最终的降噪图像.与现有的开关型RVIN降噪算法相比,所提出的非开关型TBCD模型在各种噪声比例条件下获得的峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)指标平均值能高出5~8 dB,展现出深度降噪网络模型的显著优势.而与盲降噪版的DnCNN-B和非盲降噪版的DnCNN-S(DnCNN with Known Specific Noise Ratio)两个深度降噪模型相比,TBCD模型复原图像的PSNR平均值比DnCNN-B高0.5 dB,仅比理想实现条件下(需给定准确的噪声比例值并调用专门训练的降噪模型)的DnCNN-S模型低0.3dB.这表明采用两阶段策略实现的TBCD盲卷积降噪模型在降噪效果和易用性两个方面都获得了最佳性能,很好地解决了深度RVIN降噪模型所存在的数据依赖问题.

The traditional switching random-valued impulse noise(RVIN)removal methods normally identify whether the center pixel of a local window is noise or not by comparing its local image statistic with a preset threshold,based on which the identified noise candidates were then suppressed by some noise reduction process.These methods show good denoising effect but at the cost of computational complexity.The deep convolutional neural network(CNN)based non-switching RVIN denoising models have significant advantages in both denoising effect and execution efficiency,in comparison with the classical switching ones that detect and restore RVIN noise pixel-by-pixel.Nonetheless,they cannot obtain best performance with respect to denoising effect and flexibility at the same time due to inherent data dependency.To this end,a novel two-stage blind convolutional denoising(TBCD)model on the basis of DnCNN(Denoising Convolutional Neural Network)architecture for the removal of RVIN was proposed in this paper.At the first stage,given a noisy image corrupted by RVIN with any noise ratio in the interval[0%,90%],the DnCNN-B(DnCNN for Blind Denoising Task)model was first applied to removal RVIN noise preliminarily.Meanwhile,a pre-trained noise detection model was exploited to obtain the corresponding noise label matrix of the given noisy image.And then a sparse sampling image was generated by taking dot products between the noisy image and its corresponding noise label matrix.At the second stage,to further improve the quality of the preliminary image restored by the DnCNN-B denoising model at the first stage,the preliminary image and the sparse sampling image were concatenated and fed into a pre-trained dual-channel image quality boosting model to generate a residual image.The final restored latent image can be obtained by subtracting the preliminary restored image from the residual image.The denoising performance of the proposed method was compared with the state-of-the-art switching ones(i.e.,PSMF,ROLD-EPR,ASWM,ROR-NLM,MLP-EPR,WCSR,and ALOHA),two CNN-based ones(i.e.,DnCNN-B and DnCNN-S),and three regularization-based Gaussian-impulse noise removal methods(i.e.,WENSR,WJSR,and LSM-NLR).Experimental results show that the proposed denoising method outperforms all the compared state-of-the-art ones with respect to denoising effect and execution efficiency.Specifically,compared with the existing switching RVIN removal methods,the proposed non-switching TBCD model outperforms them by 5~8dB in terms of peak signal-to-noise ratio(PSNR)across different noise ratios,demonstrating remarkable advantages of the deep CNN-based RVIN denoising model.Compared with the CNN-based DnCNN-B and DnCNN-S(DnCNN with Known Specific Noise Ratio)denoising models,the PSNR performance of TBCD model is about 0.5 dB higher than that of DnCNN-B and 0.3 dB worse than that of DnCNN-S,respectively.Note that the performance of DnCNN-S can only be achieved under ideal circumstances,i.e.,the noise ratio of the given noisy image must be accurately measured in advance and the noisy image should also be denoised with the specific pre-trained denoising model accordingly.This indicates that the TBCD model implemented by our two-stage strategy takes competitive advantage in terms of both denoising performance and flexibility,solving the data dependence problem of the non-switching deep RVIN denoising model well.In addition,compared with the mixed Gaussian-impulse noise removal methods,the proposed one exceeds them by at least 5 dB for the case of pure RVIN,and is also more effective than them for the removal of mixed Gaussian-impulse noise.In summary,the proposed denoising method exhibits a competitive performance and shows strong attraction for practical image processing applications.