Research On Off-axis Digital Holographic Microscopic Imaging Technology Based On Deep Learning

Digital holography has the characteristics of large field of view,non-contact,and accurate dynamic measurement of object 3D morphology,which is widely used in interferometry,3D imaging and quantitative phase imaging,and has great potential in the fields of materials science,industrial and biomedical measurements.In recent years,great progress has been made in the field of digital holographic microscopy research using deep learning methods to achieve fast and accurate 3D measurements using generative adversarial network convolution,convolutional neural networks,and other methods.At present,the application of convolutional neural networks in the field of digital holography still has some problems to be solved: on the one hand,some convolutional neural networks increase the model volume while improving the accuracy,and satisfying the accuracy cannot take into account the computational complexity of the network at the same time.On the other hand,the accuracy of some convolutional neural networks needs to be improved in the network learning process due to their own networks.To solve the above problems,this thesis designs a lightweight convolutional neural network and algorithms to achieve digital holographic unwrapping,data enhancement,phase reconstruction and other work.1.A deep learning framework incorporating lightweight Mobilenet V3 network is proposed to implement phase unwrapping.Using simulated wrapped phases and real phases as training sets,random simulated noise is added to simulate close to real wrapped phases and real phase distributions,and the mapping relationship between wrapped phases and real phases is learned through the network.The UMnet network based on the Unet framework incorporating Mobilenet V3 module is architected,which is more accurate and faster than the traditional phase unwrapping algorithm,instead of the traditional method to realize the phase unfolding work.2.A deep learning framework incorporating Efficient Net V2 networks is proposed to achieve digital hologram data enhancement.A frequency domain fusion method is used to generate holograms as training set labels,and network training is performed through neural networks to obtain holograms quickly.The Unet network framework based on the fusion of the Efficient Net V2 module UEFFnet network was developed,and the network model was obtained after network training to achieve fast hologram generation.A small field-of-view training model was tested on large field-of-view images,allowing the generation of large field-of-view holograms,providing the possibility that more training sets are needed for network training,and good network accuracy was obtained using different samples for network training.3.A deep learning framework incorporating Covn Next network is proposed to achieve digital holographic end-to-end phase reconstruction.PCA principal component analysis is used to achieve phase distortion compensation,and TIE-FFT method is applied to obtain the real phase information of the object after phase unwrapping by PCA phase distortion compensation.A new network structure is adopted to realize digital holographic phase reconstruction,and the UCovnnet network is constructed by integrating the Covn Next network module on the basis of the Unet network framework.UCovnnet adopts a new visual attention network module with different parameters and module adjustments to achieve higher computational accuracy for improving the quality of phase reconstruction,compared with UMnet and UEFFnet network,UCovnnet phase reconstruction accuracy is higher.