Image compressed sensing algorithm based on multi-scale wavelet transform and deep learning

Abstract

The invention discloses an image compressed sensing algorithm based on multi-scale wavelet transform and deep learning, which comprises an image acquisition stage, to be specific, a convolutional layer is used for sampling, and a sampling vector shown in the specification is obtained; an initial reconstruction stage, to be specific, every 1*1*B<2> in the initial reconstruction vector is rearrangedas a B*B image block by adopting Reshape operation; a depth reconstruction stage, to be specific, four residual blocks are adopted to deeply reconstruct the image, an initial reconstructed image block vector in the residual blocks is used as input, and a depth reconstructed image with the size of B*B is output, after depth reconstructed image blocks are obtained, the image blocks are rearranged,and finally a reconstructed image is obtained. In the sampling stage, the convolutional neural network is used for sampling, so that the sampling efficiency is improved; at the reconstruction end, theconvolutional neural network is used for initial reconstruction, then the residual network is used for deep reconstruction, and multiple networks are used for reconstruction, so that the reconstruction performance is remarkably improved; by using the residual network, the network depth is increased, and meanwhile, an efficient training effect can still be kept, so that a better reconstruction effect is obtained.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image compression sensing algorithm based on multi-scale wavelet transform and deep learning. Background technique [0002] Compressed Sensing (CS) is an emerging sampling method that simultaneously samples and compresses via linear projection to reduce coding complexity. It transforms sparse or compressible signals by linear projection captured as a compressed signal M<<N, where is the measurement rate, and its mathematical model is: [0003] y=Φx (1) [0004] in is a sampling matrix. Gaussian random matrices are widely used due to their theoretical interpretability, but have significant computational and storage costs. Over the past few decades, many researchers have attempted to alleviate the computational complexity by exploiting prior knowledge about the signal in its However, how to use prior information beyond sparsity has become a bottleneck r...

AI Technical Summary

Problems solved by technology

Although good reconstruction results have been obtained, there is still room for improvement in reconstruction performance, and further research on multi-scale sampling and multi-neural network cascaded reconstruction is still needed

Images