538
M. Ma et al.
transform the complex CS reconstruction problem into a classic denoising problem. The denoising performance directly affects the quality of CS reconstruction
[5]. At present, there is no denoiser that is universally applicable to all types
of images. The mainstream denoisers can be roughly divided into three categories: non-local self-similarity (NSS) prior-based, convolutional neural network
(CNN)-based and combine NSS-based and CNN-based methods. The NSS-based
methods are favorable on images with regular and repetitive patterns. Since the
priors are based on human observation and contain some parameters that have
to be tuned by users, it is difficult to find the irregular priors implied in the
images [6,7]. The CNN-based methods perform better on irregular structures.
However, the patch-based denoising is a local processing and the existing neural
network methods do not consider the NSS prior [7,8]. The combine NSS-based
and CNN-based methods show advantages for both regular and irregular priors.
Since the block matching number of the trained network structure is fixed, the
performance of the combine-based method is weaker than the NSS-based method
for strong regular feature images. In addition, the network part only simulates
the filtering process, so the performance of the combine-based method is weaker
than the CNN-based method for strong irregular feature images [7].
In this paper, aiming at the issue that the denoiser is not universal, we propose a CS image reconstruction strategy that can adaptively select the denoiser
for different types of images. We note that regular texture and irregular smoothing are two features that universally exist in natural images [9]. Based on the CS
measurement image, we design a method that first uses fast Fourier transform
(FFT) to extract image features and then applies an adaptive denoiser-selection
reconstruction algorithm to different types of images. The main contributions of
this paper are summarized as follows:
• We propose a CS image reconstruction framework which can select denoisers
adaptively based on the feature extraction from CS measurement.
• We propose a feature-aware adaptively denoiser-selection method which is
based on the features extracted by FFT.
The rest of the paper is organized as follows. Section 2 introduces the compressive sensing theory. In Sect. 3, we introduce the D-AMP framework and the
typical denoisers. Section 4 introduces the proposed feature-aware adaptively
denoiser-selection strategy. In Sect 5, we display the simulation results of our
research to demonstrate the effectiveness of the compressed image reconstruction method which can select denoiser adaptively by feature-aware. Conclusion
is given in the last section.
M. Ma et al.
transform the complex CS reconstruction problem into a classic denoising problem. The denoising performance directly affects the quality of CS reconstruction
[5]. At present, there is no denoiser that is universally applicable to all types
of images. The mainstream denoisers can be roughly divided into three categories: non-local self-similarity (NSS) prior-based, convolutional neural network
(CNN)-based and combine NSS-based and CNN-based methods. The NSS-based
methods are favorable on images with regular and repetitive patterns. Since the
priors are based on human observation and contain some parameters that have
to be tuned by users, it is difficult to find the irregular priors implied in the
images [6,7]. The CNN-based methods perform better on irregular structures.
However, the patch-based denoising is a local processing and the existing neural
network methods do not consider the NSS prior [7,8]. The combine NSS-based
and CNN-based methods show advantages for both regular and irregular priors.
Since the block matching number of the trained network structure is fixed, the
performance of the combine-based method is weaker than the NSS-based method
for strong regular feature images. In addition, the network part only simulates
the filtering process, so the performance of the combine-based method is weaker
than the CNN-based method for strong irregular feature images [7].
In this paper, aiming at the issue that the denoiser is not universal, we propose a CS image reconstruction strategy that can adaptively select the denoiser
for different types of images. We note that regular texture and irregular smoothing are two features that universally exist in natural images [9]. Based on the CS
measurement image, we design a method that first uses fast Fourier transform
(FFT) to extract image features and then applies an adaptive denoiser-selection
reconstruction algorithm to different types of images. The main contributions of
this paper are summarized as follows:
• We propose a CS image reconstruction framework which can select denoisers
adaptively based on the feature extraction from CS measurement.
• We propose a feature-aware adaptively denoiser-selection method which is
based on the features extracted by FFT.
The rest of the paper is organized as follows. Section 2 introduces the compressive sensing theory. In Sect. 3, we introduce the D-AMP framework and the
typical denoisers. Section 4 introduces the proposed feature-aware adaptively
denoiser-selection strategy. In Sect 5, we display the simulation results of our
research to demonstrate the effectiveness of the compressed image reconstruction method which can select denoiser adaptively by feature-aware. Conclusion
is given in the last section.