Research On Removal Of The Twin Imagein Holographic Image Reconstruction Methods Based On Deep Learning

Holographic technology is a technique that records and reconstructs a true three-dimensional image of an object using the principles of interference and diffraction.During the recording process,since the imaging sensor can only record the amplitude of the hologram wavefront,the phase information is lost,resulting in the reconstructed image being seriously affected by the twin image.This paper focuses on research on removal of the twin image in holographic image reconstruction methods based on deep learning.The specific research contents are as follows:Firstly,in order to remove the twin image in the reconstruction of in-line digital holography,multi-scale residual dense network is proposed.The network will input images with twin image artifacts to perform different sampling operations with different sampling rates to obtain different scales.At different scales,to extract different levels of features,performing residual dense blocks learning and corresponding up-sampling operations,and finally merging different scales of information to output reconstructed images.The experimental results show that the method proposed effectively removes the twin image in holographic reconstruction.Secondly,in order to reduce the reconstruction time of network,a residual memory network is proposed.Firstly,the first-order Haar wavelet transform is performed on the input image,and the mapping relationship between the different sub-bands of input twin image and the original image is learned in the wavelet domain to reduce the mapping complexity.At the same time,the network introduces a memory mechanism that can transmit the information of the network front-end convolution layer to the network back-end convolution layer.By comparing with other methods,it shows that the method proposed can not only improve the quality of reconstruction,but also shorten the reconstruction time.Finally,in order to improve the reconstruction performance of network,a multi-scale memory network is proposed.The network decomposes the input twin image into three scales by two down-sampling operations,while increasing the network receptive field.The first scale and the second scale at the front end of the network introduce branches to supplement the information lost during the network delivery process.Information of different scales of the image is extracted by multiple memory blocks.Finally,the information of different scales of the image is merged to reconstruct a high quality image.The experimental results show that the performance of the network proposed is better than the comparison methods.