Deep Learning Improving The Imaging Quality Of Single Pixel Microscope

The traditional optical microscope is a combination of lenses,which is used as measuring instrument with the ability to enlarge images.CCD component mounted on the microscope system outputs images through collecting optical signals and converting to electrical signals.However,CCD components generally have higher costs and are not applicable when the optical source is invisible.Single-pixel imaging technology breaks through this limitation and widens the spectrum of traditional microscopes.Simultaneously,the compressive sampling like approach works in terms of reducing costs.Fourier single-pixel imaging technology can provide higher quality images,except for the low sampling frequency cases.In order to improve the quality of images,the post-processing methods have been widely researched,especially deep learning method,which has the advantage of achieving end-to-end learning without any intermediate process.In this paper,the deep learning methods are used to obtain high quality images through the single-pixel imaging for traditional microscopes.First,based on Fourier single-pixel imaging technology,a complete experiment system of traditional optical microscope was established and a series of original images were obtained through experiments.Second,the principles of improving image quality using deep learning method were introduced.There were three deep learning models based on convolutional neural network,three-layer,ten-layer and nineteen-layer network structures.The corresponding datasets were used to train and test the network model,and the improved model was tested.Verification shows that deeper networks have better optimization results.The same image is input to different deep learning models to quantitatively analyze the impact of single-pixel imaging results of different models.Finally,by using the image captured by the single-pixel experiment as the input of the deep learning model,the output image can be obtained in real time.By adjusting the structure,parameters and convergence speed in the model,it is more suitable for training experimental images.In addition,the training set is further processed.Due to the existence of the aperture diaphragm in the experiment,a circular area is defined to the training set and Gaussian noise signals are added to analyze its impact on the results.After that,the data set is directly cut into a circle for training,and the experimental image of the circle illustrates the effect of deep learning on the single-pixel microscopic imaging results.