Computational Holographic Imaging Through Scattering Media By Deep Learing

When light passes through non-uniform media such as smoke,fog,biological tissue and ground glass and so on,it will be scattered,resulting in distortion of the wavefront of the beam and the formation of speckle patterns on the image plane,which hinders the development of optical imaging in remote sensing observation,public safety,biomedical and other fields.Therefore,scattering imaging is one of the research hotspots in the field of optics.As a new imaging technology,computational holographic imaging can load the phase information of the image through a spatial light modulator(SLM)to modulate the wavefront of the beam,making it suitable for a wider range of display types.But light scattering is still regarded as an obstacle to computing optical imaging.Obviously,researching and exploring how to recover the target image from the speckle image of the hologram is of great significance to promote the development of computational optical imaging.The propagation process of light waves in random scattering media is researched,modeled and simulated,and also a Generative Adversarial Networks(GAN)based on the encoding-decoding structure is proposed and used for computational holographic imaging of targets in a scattering environment.The main research work is as follows:(1)The propagation characteristics of light waves in random scattering media are studied,the properties of random scattering media are statistically analyzed,and the transmission matrix is simulated.Meanwhile,based on the single-plane GS algorithm,the pure phase computational hologram was recovered.And combined with the Fresnel diffraction theory and the angular spectrum theory,a computational holographic scattering imaging simulation model was constructed,and the imaging of the phase hologram through a random scattering medium was simulated.(2)Based on the prior knowledge of physics and the characteristics of optical devices,we designed and built a computational holographic scattering imaging system,and realized automatic system acquisition of a large amount of image data through programming control.(3)We researched and analyzed deep networks with generative effects,proposed an end-to-end generative confrontation network structure,and explained the principles of network design.At the same time,the loss function of the original generation confrontation network is improved,and the L1 norm,L2 norm and structural similarity function are added as content loss functions to the original confrontation loss function,which not only alleviates the network overfitting,but also improves the imaging quality.(4)We have explored the reconstruction ability of the improved generative adversarial network under ground glass and A4 paper scattering conditions,MINIST handwritten numbers and Fashion-MINIST fashion item data sets,and three loss functions.Based on experiments,it was found that the MINIST handwritten digit data set through ground glass has the best speckle image restoration effect when the structural similarity function is used as the content loss function.Furthermore,images completely different from the training set are selected to test the generalization ability of the network.Experiments showed that the network we proposed not only has good generalization ability,but also learned the functional relationship between the reconstructed image speckle image and the target intensity image.