基于深度学习SDA的压缩感知图像重构方法

Image reconstruction in compressed sensing based on deep learning SDA approach

为减少传统压缩感知对图像信号测量与重构的时间,提高重构精度,提出一种深度学习SDA的压缩感知框架,采用堆叠去噪自动编码器(SDA)作为无监督特征学习器,支持信号的线性和非线性测量,捕获特定信号的不同元素之间的统计依赖性;利用前馈深度神经网络代替传统重构算法,从训练数据中学习信号的结构化表示,实现图像信号重构。实验结果表明,与压缩感知SPL算法、D-AMP算法和TV算法比较,该方法峰值信噪比(PSNR)更高,且重构时间仅仅为0.002s,远低于其它3种算法。

To decrease the measurement and reconstruction time and improve the reconstruction accuracy of traditional compressive sensing,a compressed sensing framework based on deep learning SDA was proposed.Stacked denoising autoencoder(SDA)was used as an unsupervised feature learner,in which linear and non-linear measurements of the signal were supported,to capture the statistical dependencies between different elements of a given signal.The feed-forward neural network was used to replace the traditional reconstruction algorithm to learn the structured representation of the signal from the training data and to reconstruct the image signal.Experimental results show that the PSNR of the proposed method is higher than that of compressed sensing SPL algorithm,D-AMP algorithm and TV algorithm.And the reconstruction time is only 0.002 seconds,which is much lower than the other three algorithms.