Research On Removal Of The Twin Image In In-Line Digital Holographic Methods Based On Deep Learning

In-line digital holography technology uses an optical sensor(CCD)to record the interference spectrum,and then uses a computer to simulate the optical diffraction process to achieve object reproduction.Because the in-line holographic object light is in the line as the reference light,the twin image and the original image appear superimposed during the reproduction process.This article mainly focuses on the in-line digital holographic twin image removal method based on deep learning.The specific research contents are as follows:Firstly,to adaptively rescale each channel-wise feature by modeling the interdependencies across feature channels.A convolutional channel attention network is proposed.Adding a channel attention mechanism behind the convolutional layer of the network can cause specific attention to the channel features output by the convolutional layer,focusing more on useful features to enhance the network’s discriminative learning ability.The experimental results show that the network with attention mechanism has a better effect of removing the twin images.Secondly,in order to further improve the performance of the network and restore the literary and artistic details of the image,a multi-scale residual network based on different loss functions is proposed.The input image is multi-scale decomposition with different sampling rates,and then each branch enters the residual After feature extraction in the block,the respective upsampling operations are performed separately,and finally,the multi-scale features are fused.At the same time,Total Variation(TV)loss and perception loss are added to the network training to optimize the network.The experiment proves that the network structure optimized by various loss functions has reached a certain expected effect.Finally,in order to better remove the twin image in the in-line holographic reconstruction,an hourglass memory network is proposed.Before the image is down-sampled,in order to prevent the loss of the original-scale information,the upper half of the branch will be divided to save the original-scale image features to retain the original-scale information memory function,and then information fusion with the up-sampled high-resolution image In addition,each time the feature information is extracted,residual block learning is used to retain the original level of information without changing the data size.Experiments show that the method can effectively remove the twin image and is better than the comparison method.