周先春,史振婷,王子威,李婷,张影.基于纹理先验的扩张残差注意力相似性去噪网络[J].电子测量与仪器学报,2024,38(5):75-89 |
基于纹理先验的扩张残差注意力相似性去噪网络 |
Expanded residual attention similarity denoising network based on texture prior |
|
DOI: |
中文关键词: 图像去噪 卷积神经网络 纹理信息 注意力相似性模块 扩张残差模块 |
英文关键词:image denoising convolutional neural network texture information attention similarity module dilated residual module |
基金项目:国家自然科学基金(11202106,61302188)项目资助 |
|
Author | Institution |
Zhou Xianchun | School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China |
Shi Zhenting | School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China |
Wang Ziwei | School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China |
Li Ting | School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China |
Zhang Ying | School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China |
|
摘要点击次数: 2 |
全文下载次数: 275 |
中文摘要: |
目前,大多数基于卷积神经网络的图像去噪模型不能充分利用图像数据的冗余性,这限制了模型的表达能力。而且,为了有效去噪,往往将边缘信息用作先验知识,而纹理信息通常被忽略。针对这些问题,提出一种新的图像去噪网络,该网络首先使用注意力相似性模块提取图像的全局相似性特征,通过平均池化来平滑和抑制注意力相似性模块中的噪声,以进一步提高网络性能;其次使用扩张残差模块来提取图像的局部和全局特征;最后使用全局残差学习增强网络从浅层到深层的去噪效果。此外,还设计一种纹理提取网络从噪声图像中提取局部二值模式以获取纹理信息,利用纹理信息作为先验知识,可在去噪过程中保留演化图像中的细节。实验结果表明,与一些先进的去噪网络相比,新提出的去噪网络在图像视觉上有很大改善、效率更高且峰值信噪比提高了2 dB左右,结构相似性提高了3%左右,更有利于实际应用。 |
英文摘要: |
Currently, most image denoising models based on convolutional neural networks cannot fully utilize the redundancy of image data, which limits the expressive power of the models. Moreover, edge information is often used as a priori knowledge for effective denoising, while texture information is usually ignored. To address these issues, a new image denoising network is proposed, which firstly uses the attentional similarity module to extract global similarity features of the image, and smooths and suppresses the noise in the attentional similarity module through average pooling to further improve the network performance; secondly, the dilated residual module is used to extract both local and global features of the image; finally, a global residual learning is utilized to enhance the denoising performance from shallow to deep layers. In addition, a texture extraction network is designed to extract local binary patterns from noisy images to obtain texture information, which can be utilized as a priori knowledge to preserve the details in the evolved images during the denoising process. The experimental results show that compared with some advanced denoising networks, the newly proposed denoising network has a great improvement in image vision, higher efficiency and peak signal-to-noise ratio by about 2 dB, and structural similarity by about 3%, which is more conducive to practical applications. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|