双层卷积神经网络去噪模块驱动的图像去模糊方法

Image deblurring method driven by double layer convolution neural network denoising module

针对目前深度神经网络图像去噪模型处理不同噪声水平的不灵活性问题,给出一种双层卷积深度神经网络去噪模型驱动的图像去模糊方法。首先,通过扩张卷积网络来增加网络的宽度,增强去噪网络的学习能力;其次,采用批再归一化,解决小批量问题,加速训练网络的收敛;然后,利用噪声水平范围为[0,75]的先验信息,通过端到端的训练,使去噪器和反向投影(back projection,BP)模块同时优化,达到很好的图像去模糊效果;最后,采用训练好的网络对模糊图像进行测试,并与块似然对数期望(expected patch log likelihood,EPLL)、迭代解耦去模糊三维块区配(iterative decoupled deblurring BM3D,IDDBM3D)、非局部中心化稀疏表示(nonlocally centralized sparse representation,NCSR)、记忆网络(memory network,MemNet)、深度去噪卷积神经网络(deep denoiser convolutional neural network,DDCNN)及去噪先验深度神经网络(denoising prior deep neural network,DPDNN)等当前主流的基于模型的去模糊方法进行了比较。实验结果表明,该方法复原的图像具有较高的峰值信噪比(peak signal to noise ratio,PSNR)和结构相似性图像度量(structural similarity image measurement,SSIM)值。无论在主观视觉效果还是客观质量评价指标方面,该方法均优于当前主流的图像去模糊方法。

To solve this problem for inflexible of noise levels for deep convolution neural network for image denoising,an image deblurring method driven by a double deep convolution neural network for image denoising is proposed.The learning capability of denoising network is reinforced via using dilated convolution to broaden network width.Then,the small batch problems are solved by batch renormalization to accelerate the convergence of network training.The prior information of noise levels in[0,75]is used in end-to-end training to optimize the denoiser and BP modules and to obtain good image deblurring effect.Finally,the trained network is utilized for testing the blurring images,with comparison to the prevailing methods based on models like the expected path log likelihood(EPLL),iterative decoupled deblurring BM3D(IDDBM3D),nonlocally centralized sparse representation(NCSR),memory network(MemNet),deep denoiser convolutional neural network(DDCNN)and denoising prior deep neural network(DPDNN).Experimental results reveal that the images restored by this method enjoy a high peak signal to noise ratio(PSNR)and a high structural similarity image measurement(SSIM)value.This method excels all the prevalent image deblurring methods in terms of visual comparison and quality evaluation.