曹世鹏,
崔美玥,
练秋生,
1.燕山大学信息科学与工程学院 秦皇岛 066004
2.河北省信息传输与信号处理重点实验室 秦皇岛 066004
基金项目:国家自然科学基金(61471313)，河北省自然科学基金(F2019203318)
详细信息
作者简介:陈书贞：女，1968年生，副教授，研究方向为图像处理、压缩感知、深度学习、相位恢复
曹世鹏：男，1993年生，硕士生，研究方向为深度学习、动态场景去模糊
崔美玥：女，1996年生，硕士生，研究方向为深度学习、人脸图像去模糊、超分辨率
练秋生：男，1969年生，教授，博士生导师，研究方向为稀疏表示、深度学习、压缩感知及相位恢复
通讯作者:练秋生 lianqs@ysu.edu.cn
中图分类号:TN911.73计量
文章访问数:543
HTML全文浏览量:283
PDF下载量:143
被引次数:0
出版历程
收稿日期:2019-11-27
修回日期:2020-10-29
网络出版日期:2020-11-25
刊出日期:2021-01-15
Image Blind Deblurring Algorithm Based on Deep Multi-level Wavelet Transform
Shuzhen CHEN,Shipeng CAO,
Meiyue CUI,
Qiusheng LIAN,
1. Institute of Information Science and Technology, Yanshan University, Qinhuangdao 066004, China
2. Hebei Key Laboratory of Information Transmission and Signal Processing, Qinhuangdao 066004, China
Funds:The National Natural Science Foundation of China (61471313), The Natural Science Foundation of Hebei Province (F2019203318)
摘要
摘要:近年来卷积神经网络广泛应用于单幅图像去模糊问题，卷积神经网络的感受野大小、网络深度等会影响图像去模糊算法性能。为了增大感受野以提高图像去模糊算法的性能，该文提出一种基于深度多级小波变换的图像盲去模糊算法。将小波变换嵌入编-解码结构中，在增大感受野的同时加强图像特征的稀疏性。为在小波域重构高质量图像，该文利用多尺度扩张稠密块提取图像的多尺度信息，同时引入特征融合块以自适应地融合编-解码之间的特征。此外，由于小波域和空间域对图像信息的表示存在差异，为融合这些不同的特征表示，该文利用空间域重建模块在空间域进一步提高重构图像的质量。实验结果表明该文方法在结构相似度(SSIM)和峰值信噪比(PSNR)上具有更好的性能，而且在真实模糊图像上具有更好的视觉效果。
关键词:图像去模糊/
深度学习/
小波变换/
多尺度/
特征融合
Abstract:In recent years, convolutional neural networks are widely used in single image deblurring problems. The receptive field size and network depth of convolutional neural networks can affect the performance of image deblurring algorithms. In order to improve the performance of image deblurring algorithm by increasing the receptive field, an image blind deblurring algorithm based on deep multi-level wavelet transform is proposed. Embedding the wavelet transform into the encoder-decoder architecture enhances the sparsity of the image features while increasing the receptive field. In order to reconstruct high-quality images in the wavelet domain, the paper leverges to multi-scale dilated dense block to extract multi-scale information of images, and introduces feature fusion blocks to fuse adaptively features between encoder and decoder. In addition, due to the difference in representation of image information between the wavelet domain and the spatial domain, in order to fuse these different feature representations, the spatial domain reconstruction module is used to improve further the quality of the reconstructed image in the spatial domain. The experimental results show that the proposed method has better performance on Structural SIMilarity index (SSIM) and Peak Signal-to-Noise Ratio, and has better visual effects on real blurred images.
Key words:Image deblurring/
Deep learning/
Wavelet transform/
Multi-scale/
Feature fusion
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=5db949e6-b84b-4180-aa5a-3839b4984f28
