Study On Learning-based Image Super Resolution Algorithms And Its Application In Bio-microscopy

Efficient acquisition of high resolution bio-electron microscopic images is an important part of brain science research.The acquisition time-consuming and acquisition accuracy are two difficult problems in brain micro-reconstruction engineering.Due to the limitation of hardware,it is necessary to solve this problem from the perspective of algorithm.In this paper,we focus on the speed and accuracy of reconstruction,study the learning-based super-resolution algorithm,and apply it to the field of electron microscopy image reconstruction.From the perspective of reconstruction speed,a fast super-resolution algorithm based on Bayesian hierarchical tree local regression is proposed in this paper.Based on the Bayesian learning principle,the hierarchical tree is trained in the image feature space to obtain the optimal tree node parameters and the maximum probability of the regression matrix of leaf nodes.By means of image space compression(amplitude,phase and frequency),search compression and regression compression,the parameters of the model are reduced and the computational complexity is reduced,thus the reconstruction speed of the model is improved and the requirement for hardware is reduced.Finally,the experiment verifies the advantage of the algorithm in the speed of reconstruction.The algorithm can be used in the early stage of electron microscope image acquisition.From the perspective of reconstruction accuracy,a full-convolution end-to-end neural network based on multi-scale wavelet decomposition is proposed.First,the main structure of the network is constructed,which consists of two modules,encoding and decoding.Then the indirect prediction of multi-scale wavelet coefficients is used to replace the direct prediction of gray images.The training constraint modes include gray domain and wavelet domain.Then,the output coefficients are grouped according to the wavelet scale,which makes the model training more stable and the hyper-parameter adjustment more flexible.Experiments show that compared with other classical algorithms,this algorithm is more prominent and true in detail texture restoration.The algorithm can be applied to the post-analysis stage of electron microscope image.Using super-resolution technology to reconstruct electron microscopic images can shorten the period of electron microscopic acquisition,meet the needs of industrial production and scientific research of high-throughput acquisition,and speed up the process of micro-reconstruction of nerve circuits.