Research On Photon Counting Imaging Via Compressive Sensing And Deep Learning

Visualizing an object at low illumination conditions has important application in various fields including quantum imaging,astronomy,biological fluorescence measurements and medical imaging.The challenge with light reduction to an extremely low level is photon starvation and Poisson noise starts to dominate.This consequently causes low-quality images reconstructed in their system.To suppress photon noise inherent in the optical detection process,they typically require a large number of photon detections.A commercially available detection typically collects greater than10~9 photons to provide the user with a high quality photograph.However,limitations on the optical flux and integration time preclude the collection of such a large number of photons.A key challenge in such scenarios is to make use of a small number of photon detections to accurately recover the desired scene information.In this paper,based on Compressive sensing(CS)and Deep learning method have been proposed to solve the problem of photon counting imaging.In this paper,the collected photon sequences are optimized to reconstruct a high-quality photon counting image using CS recovery algorithm.Compared with the original method,a photon counting image with higher signal-to-noise ratio can be obtained.Integrated imaging(II)as a popular three dimensional(3D)stereo imaging technology currently has greater advantages than other 3D imaging methods.For example,it includes full parallax,has continuous viewing angle information,simple imaging equipment,no visual fatigue compared to other 3D imaging methods.The proposed method further improves the visual quality of the3D reconstructed image by II.Experimental results show that the proposed method can effectively solve the photon counting imaging in dark environments.As the research on deep learning theory becomes more and more mature,many problems in the field of image processing can be effectively solved by deep learning methods,such as super-resolution images,demosaicing,denoising,and target recognition.This data-driven approach can fully learn from corrupted observations mapped into unobserved clean versions.This paper introduces a convolutional neural network for the photon counting image visualization method.We introduce up/down projection block(UDPB)and select memory block(SMB)to self-correcting reconstructions of the visual structure and visual feature.We repeatedly stack UDPB and SMB in decoding stages of a U-Net like architecture to form the generator of a generative adversarial network(GAN)for visualizing images dominated by Poisson noise.In addition,we have tested on different datasets and real pictures taken by accessible consumer-grade cameras.The experimental results show that the proposed method is superior to the state-of-the-art methods.The two method proposed in this paper are optimized at the element image,and do not need to provide extra precision equipment to solve the photon counting image visualization problem.