Digital Holographic Reconstruction Based On Deep Learning

The reconstruction distance of Digital Holographic?DH?is related to the recording distance of hologram^[1].Due to the complexity of the aberrations,diffraction limits,resolution limitations of recording instruments?CCD,CMOS sensors,SLM space optical modulators?,and some human-made experimental loss errors make it difficult to reconstruct clear images,This requires a large and complex post-processing process for the image at a later stage to improve the resolution of the DH.Because ordinary optical systems generally can't achieve real-time calculation on-line,at the same time,the high-precision equipment required to eliminate the impact of image propagation noise and error caused by various image problems is extremely expensive,resulting in no public high-resolution data sets for the field of digital holographic imaging,so that its accurate high-resolution expected imaging target has been a major problem.Therefore,in order to optimize and improve the reconstruction system and enhance the spatial bandwidth of the imaging system,this paper first studies and analyzes the image processing method of the iterative DH super-resolution reconstruction based on interpolation and the non-iterative multi-frame DH resolution enhancement reconstruction algorithm to enhance the spatial resolution of the DH.Second,because iterative methods are computationally intensive,iterative-based methods are computationally time-consuming and expensive and are not suitable for real-time applications.In addition,iterative algorithms sometimes fall into local minimums and can result in poor performance.In order to overcome the problem of iterative algorithm,for fast or real-time processing,this paper studies the single image super-resolution reconstruction technology based on deep learning and proposes the residual local dense path network based on end-to-end deep learning-Holo DSR.The network can learn the upper sampling operation?i.e.high resolution from low resolution?from a single DH reconstruction image to make the image detail clearer.Finally,it also verifies that the network shows its unique advantages and research prospects in the reconstruction results of different subsampling factors,both in the evaluation of image reconstruction quality and in eliminating the influence of artifacts to improve the resolution of DH reconstruction image.DH numerical reconstruction based on the principle of light-based diffraction propagation follows certain rules,the combination of super-resolution algorithms based on deep machine learning and optical imaging allows neural networks to understand the laws of interference,diffraction,etc.,trying to solve the problem of lack of fine details,failure to remove defects and compression artifacts based on the upgrade methods of traditional algorithms.For those who perform these tasks manually,this is a very slow and arduous process.The benefits of obtaining higher quality images from a place that never existed or has been lost,which may be beneficial in many areas,even saving lives in medical applications.