基于深度学习的多尺度分块压缩感知算法

Multi-scale Block Compressed Sensing Based on Deep Learning

压缩感知算法可以突破传统采样定理的限制,在采样的同时完成对数据的压缩.利用深度学习方法解决压缩感知算法存在的缺陷,在图像处理领域十分受欢迎.现有的深度学习框架下的压缩感知算法,多采用全连接层进行采样,对于自然图像来说,所需计算量巨大,不利于数据的存储.分块压缩感知算法,多采用单一尺度分块,如何选取合适的分块尺寸成为难题.本文提出了基于深度学习的多尺度分块压缩感知算法,利用卷积层代替全连接层,实现对原始图像的多尺度分块,同时添加了卷积自编码器对重构图像进一步优化.实验结果表明,本文算法对于图像特征的提取及重构,都表现出了明显的优势,取得了良好的效果.

Compressed sensing(CS),a technique for simultaneous data sampling and compression,can break through the limitation of traditional sampling theorem.It has become a popular research direction in the field of image processing that using deep learning method to solve the defects of CS algorithm.The existing algorithms under the deep framework mostly use afully-connected layer data sampling.For the natural images,the amount of computation is vary large,which is not conducive to the data storage.Block Compressed Sensing(BCS)methods mostly use single-scale block segmentation,and the selection of appropriate block size becomes a problem.In this paper,a deep learning method based on multi-scale block compressed sensing is proposed.The convolutional layer is used to replace the full-connected layerto realize multi-scale block segmentation of the original image,and the convolutional self-encoder is used to further optimize the reconstructed image.The experimental results show that the algorithm presented in this paper has obvious advantages in image festure extraction and reconstruction and achieved good results.