| Optical imaging technology is a very promising direction in the field of optics,which has been widely used in medical diagnosis,remote sensing,virtual reality and other scenes.With the development of biomedical imaging,especially the medical endoscopy,ultra-fine endoscopic imaging with high resolution and low cost has become a research hotspot.Because of its small core diameter,rich internal transmission modes and large information throughput,the multimode fiber is considered to be an ideal image medium for ultra-fine endoscopic imaging.Due to the special optical characteristics of multimode fibers,the original image will form a highly randomized speckle pattern after transmission through multimode fibers.How to restore the original image from the speckle pattern has become the key problem of multimode fiber imaging.Using big data analysis,deep learning can automatically learn the abstract mapping relationship between speckle patterns and original images but without the prior knowledge of multimode fibers.It has become a new research direction of multimode fiber imaging technology.In this paper,we will research multimode fiber imaging technology along the main line of deep learning.Firstly,we use the self-built multimode fiber imaging experimental system to obtain enough speckle patterns corresponding to original images one by one.For the problem of large demand of data samples in deep learning,Pix2Pix for image translation in generation adversarial networks is proposed to recover speckle patterns.The idea of adversarial generation is used for reference to improve the feature extraction ability of the network,reduce the data set and obtain better recovery effect.Compared with U-Net,Pix2Pix can achieve better recovery effect than U-Net with the same training set size,and Pix2Pix can also achieve recovery effect compared with U-Net with smaller training set.Especially when the data set is small and the images are complex,the advantages of Pix2Pix are more obvious. |
PDF Full Text Request