Application Of Deep Learning In Ghost Imaging

Ghost imaging(GI),also known as correlation imaging,is different from the traditional imaging techniques with a nonlocality property.Taking advantages of the properties of quantum entanglement or the fluctuation of light beam to recover the object at the end without the object,GI has become a hot topic in the field of quantum optics in recent years.We study the application of deep learning in GI for handwritten digit recognition and edge detection.The main work and results are as the following:(1)Based on the theory of computational ghost imaging and deep learning,we propose an unknown handwritten digit recognition method based on deep neural network.The characteristics of Discrete Cosine transform and Fourier transform in the low-frequency information are used to design the special ilumination patterns,and the bucket detector without spatial resolution in GI system is utilized to collect each detection result.Finally,the detection results are as the feature of unknown handwritten digit to input to the designed and trained deep neural network for recognition.The simulation and experiment results show that the method has a good recognition performance,and can classify the handwritten digital images without knowing the handwritten digital at first.The method has high recognition rate,short recognition time,non-locallity characteristics,and the designed deep neural network is simple,so it will have great application prospects.(2)Edge detection has always been the key problem in image processing.In this thesis,the deep learning and ghost imaging technology are applied to edge detection,we propose a new method of edge detection of unknown image based on holistically nested neural network(HED).Based on ghost imaging and HED network,we input the vague imaging from numerical GI system and its corresponding edge information into the network for training.After training,a clear edge information can be achieved when the data from experimental GI system.The results show that in the case of low compression ratio,that is,the measurement times is much less than that of the size of the unkown object image,the proposed method can achieve a clear edge information.Compared with the other edge detection methods based on GI,the proposed method has a good performance even if there is no edge information obtained from other edge detection methods based on GI.