Research Of Photon-limited Imaging Based On Deep Learning

Imaging under ultra weak-lighting conditions is about how to obtain a high quality image when the light intensity is weak until the light signal is in the particle characteristics.Photon counting imaging can be used for detection in the case.To obtain a high quality image,we usually need to combine ultra weak-lighting detection with a series of post-processing algorithms.Benefiting from the development of detectors,especially the invention of singlephoton array detector(SPC),photon-limited imaging has developed rapidly.In recent years,deep learning has shown its advantages in various fields.Its strong learning ability from the training datasets has attracted widespread attention.People has tried to combine it with many problems to get better results than traditional methods in various fields.In the field of ultra weak-lighting imaging,no one has tried to combine them.And in this paper,we made such attempt.The main content includes the following aspects:(1)Photon-limited imaging super-resolution based on deep learning: We acquire high quality images by combining detection using SPC with deep learning.The photon detection accuracy and sensitivity of SPC can reach the single photon level.And we use deep learning to solve the under-sampled problem and Possion noise inherent in SPC detection.To demonstrate the feasibility of our scheme,we conducted real experiment.We detected objects with SPC under ultra weak-lighting conditions.Then we use the training dataset to train a deep learning network.Finally,we use the test samples to measure the performance.With deep learning,we only need to detect each object once instead of detecting an object multiple times translationally.The experimental results demonstrate that the network can realize 4× super-resolution and denoising at the same time.On the other hand,our scheme can still work when the photon detected is only about 0.2 per pixel.(2)Photon-limited scattering imaging based on deep learning: We focus on the complex imaging situations where both ultra weak-lighting and scattering occur simultaneously.We propose a scheme that combines deep learning with it.And then we validate it through experiments.This scheme is not affected by the low resolution characteristics of SPC detection.The low resolution can be seen as downsampling of data,and it is still sufficient to implement scattering imaging.At the same time,the scheme is not limited by the scattering model.The experimental results show that our scheme can still work when the photon detected is only about12 per pixel.This paper focuses on the application of deep learning in photon-limited detection and photon-limited scattering imaging.From the design of the algorithm to the experimental verification,we prove the feasibility of the deep learning scheme.The advantages of deep learning make up the operation difficulty of the traditional photon-limited detection scheme and the low generalization ability of the traditional algorithm.It presents a new solution to ultra weak-lighting imaging.