摘要
针对目前基于卷积神经网络的图像去噪算法中,存在的卷积层数少、模型简单、参数计算量大和收敛速度慢等问题,导致无法有效地保留高频纹理信息,易产生振铃效应,提出一种结合密连网络的图像去噪算法。首先利用低通滤波器将噪声图像分解成高频层和低频层,对包含高频层的图像信息进行傅里叶域上的预处理,将预处理后的初始图像小块作为网络的输入样本,对应的清晰图像小块作为标签训练该密连网络,得到从噪声图像到潜在清晰图像的映射函数,实现基于该网络的图像去噪。实验结果表明引入密连模块后,在保持网络深度的同时,可以保留更多被模糊的边缘细节,提高图像的视觉质量。
In view of the current image denoising algorithms based on convolutional neural networks,there are problems such as fewer convolutional layers,simple models,large parameter calculations,and slow convergence speed,which result in the inability to effectively retain high-frequency texture information and easily produce ringing effects,Propose an image denoising algorithm combined with dense network.First,a low-pass filter is used to decompose the noise image into a high-frequency layer and a low-frequency layer,and the image information containing the high-frequency layer is preprocessed in the Fourier domain,and the preprocessed initial image block is used as the input sample of the network,The corresponding clear image block is used as a label to train the densely connected network,and the mapping function from the noisy image to the potentially clear image is obtained,and the image denoising based on the network is realized.The experimental results show that after the introduction of the dense connection module,while maintaining the depth of the network,it can retain more blurred edge details and improve the visual quality of the image.
作者
王延年
李雄飞
Wang Yannian;Li Xiongfei(School of Electronic Information,Xi’an Polytechnic University,Xi'an 710048,China)
出处
《国外电子测量技术》
北大核心
2021年第3期23-27,共5页
Foreign Electronic Measurement Technology
基金
陕西省科技厅工业领域一般项目(2019GY-109)
西安市科技局科技计划(201805030YD8CG14(1))
西安工程大学柯桥纺织产业创新研究院项目(19KQYB02)资助。
关键词
图像去噪
深度密连网络
傅里叶域
视觉质量
image denoising
deep dense network
Fourier domain
visual quality
