Research On Deep Learning Denoise And Super Resolution Algorithm Of Terahertz Coaxial Holographic Reconstruction Image

Terahertz(THz)waves,with their low photon energy and high penetration ability to most non-metallic non-polar substances,have shown wide application prospects in aerospace,non-destructive testing and human body safety testing.Continuous THz coaxial holographic imaging system has the advantages of simple structure and convenient operation,so it has broad application prospects and popularization value.However,due to the low energy of the signal source,the low response rate and resolution of the area array detector,and the noise interference,the quality of the resulting image is difficult to improve.This thesis takes the continuous 2.52 THz coaxial holographic reconstruction image as the research object,analyzes the main features of the image,and introduces the method of deep learning into the processing of this type of image,from the perspective of image denoising and super resolution.Related research.In order to effectively remove the noise of the 2.52 THz coaxial holographic reproduced image,this paper is based on the deep neural network of the Dn CNN model.Under the condition that the training samples of the real THz coaxial holographic reproduced image are lacking,the simulated noise training set is used to train the deep neural network.For training.By optimizing the influence of hyperparameters such as the learning rate,loss function,and network depth of the network,the best network model is obtained.Using the trained network to denoise the 2.52 THz coaxial holographic reproduced image,the results are compared and analyzed with classic image denoising algorithms such as NLM and BM3 D,and both the subjective and objective indicators are improved.Research on super-resolution of 2.52 THz coaxial holographic reproduced image Based on the Dense Net model,a deep learning super-resolution network is constructed.The deep network is trained by using the standard training set after downsampling and the network-related hyperparameters are optimized to obtain Best for super-resolution network models.After applying the trained model to the 2.52 THz coaxial holographic reconstruction image,the results have been effectively improved in subjective and objective indicators in comparison with classic super-resolution algorithms such as sparse dictionary and anchored neighborhood regression.Finally,combining the above research on image denoising and super-resolution of 2.52 THz coaxial holographic reproduced image,combining the features of residual learning and sub-pixel convolution,a deep neural network with super-resolution collaborative denoising is designed.The network is trained using a simulation training set that is down-sampled and added with simulation noise,and the relevant network hyperparameters are optimized to obtain the best super-resolution collaborative denoising network.Applying it to the 2.52 THz coaxial holographic reproduction image,while achieving better super-resolution and denoising effects,it further improves the efficiency of the algorithm compared to by-steps.