基于有损压缩编码的降噪自编码器

Denoising Autoencoders Based on Lossy Compress Coding

图像预处理算法的优劣程度直接关系到图像后置处理的效果,如图像分割、目标检测、边缘提取等。为了获取高质量的数字图像,对图像进行降噪处理成了必不可少的前置步骤。图像降噪旨在尽可能地保持原始信息完整性(即主要特征)的同时,又能够去除信号中无用的信息。为此,提出了一种基于有损压缩编码的卷积自编码器(AutoEnconders,AE)去噪模型;并根据最大编码率下降原则(the principle of Maximal Coding Rate Reduction,MCR 2)设计了新的损失函数代替主流深度学习算法中常用的均方误差(Mean Squared Error,MSE)损失,以提高模型的鲁棒性和适应性。模型首先通过编码器处理带噪图像,得到隐变量,然后使用解码器进行解码,消除噪声并得到重构图像。接下来,保持编码器不变,将重构图像输入编码器,使编码器继续学习并得到重构隐变量。最后,通过计算隐变量与重构隐变量的距离来间接衡量重构图像与原始图像的误差,并将其作为收敛代价进行模型训练。在thumbnails128×128和CBSD68数据集上对所提模型进行了大量实验验证。实验结果表明,该自编码器框架(AE-MCR 2)在不同类型的噪声(高斯噪声、伯努利噪声和泊松噪声)下均表现出良好的性能,并具有一定的可解释性。

The performance of image preprocessing algorithms is directly related to the effect of image post-processing,such as image segmentation,target detection,edge extraction,etc.In order to obtain high-quality digital images,image noise reduction has become an essential pre-step.Image noise reduction aims to maintain the integrity of the original information(i.e.,the main features)as much as possible,while being able to remove the useless information in the signal.To this end,this paper proposes a lossy compression coding based convolutional auto-encoders(AutoEnconders,AE)denoising model.According to the principle of maximal coding rate reduction(MCR 2),a new loss function is designed to replace the mean squared error(MSE)loss commonly used in mainstream deep learning algorithms to improve the robustness and adaptability of the model.The model first processes the noisy image through an encoder to obtain the hidden variables,and then decodes it using a decoder to remove the noise and obtain the reconstructed image.Next,keeping the encoder unchanged,the reconstructed image is fed into the encoder so that the encoder continues to learn and obtains the reconstructed hidden variables.Finally,the error between the reconstructed image and the original image is indirectly measured by calculating the distance between the hidden variable and the reconstructed hidden variable,which is used as the convergence cost for model training.The proposed model is validated extensively on thumbnails128x128 and CBSD68 datasets,and the experimental results show that the self-encoder framework(AE-MCR 2)exhibits good performance under different types of noise(Gaussian,Bernoulli,and Poisson)and has some interpretability.