基于深度学习与压缩感知理论的通用图像重构算法

General image reconstruction algorithm based ondeep learning and compressed sensing theory

数字图像作为信息的高效载体,在信息传输中发挥着重要的作用。随着图像数据不断增大,需要压缩感知技术解决数据存储和传输过程中成本浪费与耗时问题。传统压缩感知运算复杂重构时间长,重构质量较差,在低采样率下将无法恢复。提出一种基于深度学习压缩感知理论的图像重构算法,同时适用于灰度图与彩色图像。压缩重构网络使用双线性插值对图像的宽高压缩,损失的信息由全连接层学习。网络中多次使用全连接层进行构建,使其具有更多的网络参数学习图像特征。对于彩色图像,通过卷积神经网络将3通道压缩为1通道,重构网络使用双线性插值将压缩图像放大,使用卷积神经网络和全连接层重构得到高质量图像。实验表明,在不同采样率下,提出的CCSNet网络的PSNR和SSIM值均为最优,重构性能优于基于深度学习的ReconNet、DR^(2)-Net和MSRNet网络。算法同时适用于灰度图像与RGB格式彩色图像,在保证运行时间尽量短的情况下,提高重构质量和缩短重构时间有较大优势。

As an efficient carrier of information,digital images play an increasingly important role in information transmission.With the increasing of image data,compressed sensing technology is needed to solve the cost waste and time in the process of data storage and transmission.Traditional compressed sensing operation takes a long time to reconstruct and has poor reconstruction quality,so it cannot be recovered at low sampling rate.In this paper,an image reconstruction algorithm based on deep learning compressed sensing theory is proposed,which is suitable for both grayscale and color images.The compression reconstruction network uses bilinear interpolation to compress the width and height of the image,and the lost information is learned by the fully connected layer.The full connection layer is used many times to construct the network,so that it has more network parameter learning image features.For color images,the 3 channels are compressed into 1 channel by convolutional neural network.Finally,the reconstruction network uses bilinear interpolation to enlarge the compressed image,and the convolutional neural network and the fully connected layer are used to reconstruct the high-quality image.Experiments show that under different sampling rates,the proposed CCSNet network has the optimal PSNR and SSIM values,and the reconstruction performance is better than the deep learn-based ReconNet,DR^(2)-Net and MSRNet networks.The algorithm is suitable for both grayscale image and RGB format color image,and has great advantages in improving reconstruction quality and shortening reconstruction time while keeping the running time as short as possible.