| Single-pixel imaging(Single-pixel imaging,SPI),as a cutting-edge computing imaging technology,has developed rapidly with the improvement of computer computing power.Area array detectors in the non-visible light band are often expensive or difficult to obtain.The single-pixel imaging method can greatly reduce the imaging detection cost in these bands.In addition,the single-pixel imaging method can increase the acquisition speed through compression sampling.However,in the current singlepixel imaging method,the image quality after compression sampling is still difficult to meet the imaging needs of the actual application in the optical system,so we can only make a trade-off between time cost and imaging quality.We seek a new single-pixel imaging method that can obtain high-quality images while reducing time costs.After a lot of experiments and in-depth research,we have proposed a single-pixel imaging method based on deep learning to meet the above requirements.We used the Generative Adversarial Network(GAN)in the deep neural network,which optimizes the measurement matrix and reconstruction algorithm to obtain better image quality at the same compression rate.Compared with other single-pixel imaging methods for artificial encoding selection during compression imaging,the deep neural network can obtain a more objective illumination coding pattern by learning a large number of images,so that it has a better image at the same compression rate than other methods quality.Through simulation experiments,we will quantitatively analyze the single-pixel imaging method based on deep learning and the compressed sensing method,Hadamard single-pixel imaging method and Fourier single-pixel imaging method.At a very low compression ratio of 12.5%,The peak signal-to-noise ratio of the method is 20% higher than the best results in other methods.If GPU accelerated,we can achieve 50 times the reconstruction speed of the traditional method.In order to study that our method is the result of practical application,we built a single-pixel imaging system to implement the application of the single-pixel imaging method.By building a single-pixel imaging system,we compared the imaging results of different single-pixel imaging methods at different compression rates in practical applications and explained the advantages of single-pixel imaging based on deep learning methods in terms of imaging quality.Finally,we summarized our work in this paper,analyzed some of the problems that have yet to be solved in single-pixel imaging based on the experience of this subject,and conceived a future solution.We also made prospects for the potential areas of development of single-pixel imaging methods. |
