Research On Mixed Noise Removal Algorithm Based On Deep Convolutional Neural Network

Digital images are frequently corrupted by various noise during acquisition and transmission,which degrades image quality and affects subsequent image processing tasks.Noise removal is necessary for further analysis and processing.Most image denoising methods are developed under the assumption that the underlying data follows a Gaussian distribution with zero-mean.In practice,this assumption seems to be rarely valid.Images corrupted by Gaussian noise might be contaminated by impulse noise or Poisson noise simultaneously.The existing regularization based denoising algorithms for mixed Gaussian-impulse noise and mixed Poisson-Gaussian noise are flexible to handle different types and levels,of noise,which utilize handcrafted image priors to construct a regularization term.These methods take iterative procedure to solve the optimization problem,leading to high computational complexity.Recently,the deep convolutional neural network(DCNN)with strong nonlinear mapping ability offers new opportunities for the research of image denoising.The DCNN-based denoising models have significant advantages over the traditional methods in terms of both denoising effect and execution efficiency.However,the existing models are still non-blind,where the best denoising effect only can be obtained by training a specific denoising model at a fixed noise level,limiting the practical application.The denoising performance of the classic regularization based noise removal methods and the flexibility of the DCNN-based denoising models can be further promoted.To this end,an image classification mechanism is designed to classify the training images into several sub-classes,of which images are similar in distortion levels.Then,several DCNN-based denoisers are trained on each sub-class to remove mixed noise quickly and flexibly.In this paper,an image quality-aware denoising convolutioanl neural network(IQANet)was proposed for the removal of the mixed Gaussian-impulse noise.Specifically,according to the statistical distribution of the peak signal-to-noise ratio(PSNR)of a large number of noisy images,a noise blend mode classification dictionary(NBMCD)was constructed.On the basis of NBMCD,noise images were classified into several classes,and specific denoisers were trained for each class.To improve the practicability of the IQANet,a CNN-based image quality estimation model was exploited to estimate the PSNRs of the noisy images.For a given image to be denoised,The NBMCD was queried according to its PSNR estimated by the image quality estimation model,and its category can be determined.Then,the corresponding pre-trained denoiser was exploited to achieve efficient blind image denoising with high quality.In the case of the mixed Poisson-Gaussian noise,the distortion of the noisy images were measured by their noise level function(NLF).And a NLF-aware denoising convolutional neural network(NLFNet)was proposed.Concretely,an image information database(IID)was first bulided,which contains numerous noisy images with different noise levels,corresponding original images,and polynomial coefficients of NLF of each noisy image.Then the NLFs of the noisy images in the IID were clustered,and the noisy images are classified into several sub-classes accordingly.Each noisy image subset and its corresponding original image subset were used to train specific DCNN-based denoising models.Given an image to be denoised,its homogenous regions were first detected,and the NLF of the image is estimated by exploited the statistics of multiple homogeneous patches.Next,the IID is queried according to the polynomial coefficients of the NLF to ascertain the category of the image.Then the corresponding pre-trained specified denoiser was adopted to perform blind image denoising task.The proposed denoising methods,namely,IQANet and NLFNet,are compared with state-of-the-art ones in terms of both denoising effect and execution efficiency.The dnoising effect is evaluated from subjective visual effect and objective image quality indicators,i.e.,PSNR,structural similarity(SSIM)and feature similarity(FSIM).And the execution efficiency is measured by the average computational time for restoring one image.The experimental results on three published image databases show that,compared with the state-of-the-art Gaussian-impulse mixed noise removal algorithms,IQANet achieves comparable denoising effect while having a great improvement in the execution efficiency,Compared with classical mixed Poisson-Gaussian noise removal algorithms,NLFNet achieves the state-of-the-art performance on denoising effect and higher efficiency across different noise levels.From the above,the proposed IQANet and NLFNet achieve better comprehensive performance on denoising effect and execution efficiency,which makes them more practical.