| Single pixel imaging(SPI)utilizes a single pixel detector without spatial resolution to achieve two-dimensional imaging,which has advantages such as high sensitivity and low cost.It has good application prospects in spectral imaging,three-dimensional imaging,ultra fast imaging,terahertz imaging,and other fields.Fourier Single Pixel Imaging(FSI)is a type of single pixel imaging that uses Fourier transform orthogonal substrates to modulate the target image.The Fourier coefficients of the target image are calculated using the modulated light intensity values,and then the target image is obtained through inverse Fourier transform.Fourier single pixel imaging can obtain high-quality imaging results,but there are problems with multiple sampling times and long imaging time.In order to reduce the number of samples,it is common to only sample the low-frequency Fourier coefficients of the Fourier spectrum and ignore the high-frequency Fourier coefficients.Due to the lack of high-frequency components,reconstructed images may lose imaging textures and details.In recent years,deep learning has made great progress in image reconstruction of single pixel imaging,which not only avoids the excessive computation time caused by traditional algorithms due to iterative operations,but also achieves higher reconstruction quality.In this paper,Fourier single pixel imaging based on deep compressive sensing reconstruction network is studied.The main work and achievements are as follows:(1)A Fourier single pixel imaging model based on deep compressive sensing reconstruction network is proposed.The whole model consists of three sub-networks:sampling network,inverse Fourier transform network and deep convolutional reconstruction network.The sampling network samples and preliminarily reconstructs the Fourier spectrum of the image,while the inverse Fourier transform network maps the Fourier spectrum of the image to the real value image.The deep convolutional reconstruction network further improves the quality of reconstructed images.(2)In order to study the reconstruction effect of sampling networks,three network models and training methods are designed.The first method is to use the low-frequency part of the Fourier spectrum as the input for the reconstruction network.The experimental results show that this method can effectively reconstruct simple images such as MNIST handwritten digits;The second method uses the Fourier spectrum of the image as input for compression reconstruction,and compares the experimental results of five sampling networks,including FC-Net,TFC-Net,WFC-Net,DFC-Net,and DWFC-Net.The experimental results show that FC-Net has the best reconstruction effect;The third method is to separate the real and imaginary parts of the Fourier spectrum of the image as inputs and enter a sampling network with shared weights,and compare the experimental results of five sampling networks.The experimental results show that this method can significantly reduce the phenomenon of blockiness in reconstructed images and improve the quality of image reconstruction.(3)Two types of deep reconstruction networks were designed.The first deep reconstruction network takes the Fourier spectrum of the image as input,first undergoes FC-Net compression sampling,and then undergoes IF-Net and deep convolutional reconstruction networks for reconstruction.Four deep convolutional reconstruction networks are compared,and the experimental results show that DC4-Net has the greatest improvement in the quality of reconstructed images.The second type of deep reconstruction network separates the real and imaginary parts of the low-frequency Fourier spectrum as inputs,passes through IFT-Net first,and then uses U-Net deep reconstruction network to reconstruct the image.The experimental results show that at lower sampling rates,the reconstructed image quality of this network is higher than that of IFT-Net under the same conditions.(4)A Fourier single pixel imaging system was built for experimental verification of the network.A series of digital grayscale matrices constructed using a series of two-dimensional Fourier base patterns were used as measurement matrices and loaded onto the DMD to measure the target object.The image was then reconstructed using IFT-Net+U-Net. |
