去噪正则化与FFDNet结合的相位恢复算法

Phase Retrieval Algorithm Based on RED Combined with FFDNet

相位恢复是图像逆问题的一种,通过图像信号的幅值,恢复出采样过程中缺失的相位信息。目前相位恢复算法使用稀疏先验以及传统去噪器先验,存在特征表征不充分的问题。prDeep算法使用DnCNN卷积神经网络作为去噪器先验,结合去噪正则化模型提升了恢复效果,但是仍存在对复杂噪声鲁棒性较差的问题。针对prDeep算法对复杂噪声鲁棒不足的问题,提出了用FFDNet作为去噪器先验与去噪正则化模型相结合的算法。该算法利用FFDNet网络对噪声的自适应性,使用去噪正则化(RED)构建优化模型,解决了复杂正则化模型求导繁琐的问题。在保证卷积神经网络对特征表征能力的同时,提高了对复杂噪声的鲁棒性以及算法的迭代效率。仿真实验结果表明,该算法在不同噪声等级下,恢复图像信噪比和迭代效率均有所提升。

Phase retrieval is a kind of image inverse problem.Through the amplitude of the image signal,the missing phase information in the sampling process can be recovered.Current phase retrieval algorithms use sparse priors and traditional denoiser priors,which have the problem of insufficient feature representation.The prDeep algorithm uses the DnCNN as the denoiser prior,combined with the denoising regularization model to improve the recovery effect,but it still has the problem of poor robustness to complex noise.Aiming at the problem of insufficient robustness of the prDeep algorithm to complex noise,an algorithm combining FFDNet as the denoiser prior and the denoising regularization model is proposed.According to FFDNet network’s adaptability to noise,the proposed algorithm uses regularization by denoising(RED)to construct an optimization model,which solves the cumbersome problem of complex regularization model derivation.While ensuring the ability of convolutional neural network to characterize features,it also improves the robustness to complex noise and the iterative efficiency of the algorithm.The simulation experiment shows that the proposed algorithm can improve the SNR and iterative efficiency of the restored image under different noise levels.