计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (4): 199-204.DOI: 10.3778/j.issn.1002-8331.1709-0155

• 图形图像处理 • 上一篇    下一篇

基于改进栈式稀疏去噪自编码器的图像去噪

马红强,马时平,许悦雷,吕  超,辛  鹏,朱明明   

  1. 空军工程大学 航空工程学院,西安 710038
  • 出版日期:2018-02-15 发布日期:2018-03-07

Image denoising based on improved stacked sparse denoising auto-encoder

MA Hongqiang, MA Shiping, XU Yuelei, LV Chao, XIN Peng, ZHU Mingming   

  1. Aeronautics Engineering College, Air Force Engineering University, Xi’an 710038, China
  • Online:2018-02-15 Published:2018-03-07

摘要: 为了提高栈式稀疏去噪自编码器(SSDA)的图像去噪性能,解决计算复杂度高,参数不易调节,训练收敛速度慢等问题,提出了一种栈式边缘化稀疏去噪自编码器(SMSDA)的图像去噪方法。首先,由于边缘化去噪自编码器(MDA)具有收敛速度快这一特性,对SDA网络损失函数作边缘化处理,形成边缘化稀疏去噪自编码器(MSDA),使其同时满足边缘性和稀疏性。其次,将多个MSDA堆叠构成深度神经网SMSDA,为避免模型参数局部最优,采用非监督逐层训练法分别训练每一层网络,再用BP算法对整个网络微调,从而获得最优权重。最后,用SMSDA对给定图像去噪。仿真结果表明,较SSDA而言,所提算法在降低计算复杂度、提高收敛速度的同时,拥有较高峰值信噪比(PSNR),且保留了更多原始图像的细节信息,具有更好的降噪性能。

关键词: 图像去噪, 深度学习, 纹理细节, 降噪自编码器, 稀疏自编码器

Abstract: In order to improve the image denoising performance of Stacked Sparse Denoising Auto-encoder(SSDA)and solve the problems of high computational complexity, difficult parameters adjustment and slow training convergence speed, the image denoising algorithm based on Stacked Marginalized Sparse Denoising Auto-encoder(SMSDA) is proposed. Marginalized Sparse Denoising Auto-encoder(MSDA) is formed by marginalizing the loss function of SDA network because of the fast convergence speed of the Marginalized Denoising Auto-encoder(MDA), which has the characteristics of SAE and MDA. Then, multiple MSDA is stacked to form deep neural network SMSDA. The unsupervised greedy layer-wise training algorithm is used to train each layer of network for avoiding the local optimization of the model parameters. The BP(Back Propagation) algorithm is used to fine tune the whole network and can obtain the optimal weight. Last, SMSDA is used to denoise a given image. Simulation results show that the proposed algorithm has higher Peak Signal to Noise Ratio(PSNR) while reducing the computational complexity and improving the convergence rate, retains more details of the original image and has better denoising performance than SSDA.

Key words: image denoising, deep learning, texture detail, Denoising Auto-Encoder(DAE), Sparse Denoising Auto-encoder(SDA)