Research On Super Resolution Reconstruction Algorithm Of Deep Undercooling Melt Image Based On Deep Learning

In space environment,liquid alloy is in thermodynamic metastable state,which is very helpful to study the material structure and thermophysical properties of deep undercooling melt.Limited by the cost and technology,researchers built electrostatic levitation combined with drop pipe device to simulate the space environment,and used high-speed camera to capture the falling image of deep undercooling melt to study its melting and solidification process.However,due to the exposure time and other hardware limitations of the image acquisition equipment,the image resolution of the deep undercooling melt is low,which is not conducive to further study its thermophysical properties and solidification interface.Software design methods such as super-resolution reconstruction can more accurately reconstruct the image contour information and effectively improve the image resolution.The traditional super-resolution reconstruction algorithms often use machine learning methods such as interpolation,reconstruction and dictionary learning,which can improve the image resolution to a certain extent.However,due to the limitation of the algorithm itself,the reconstructed image often appears edge blur,noise and details disappear.In recent years,with the development of the computing power of hardware devices and the in-depth study of related theories,deep learning method has been successfully applied in the field of super-resolution reconstruction,which surpasses the traditional algorithms in the evaluation methods of objective index and subjective feeling.At present,most of the super-resolution reconstruction algorithms based on deep learning use y-channel or y-channel,Cb-channel and Cr-channel learning directly,which is not enough to provide more prior information to solve the ill-posed problem.In this paper,a single frame image super-resolution reconstruction algorithm based on deep learning is proposed.The algorithm uses the frequency domain features learning method to better reconstruct the deep undercooling melt image.In the proposed method,the frequency domain feature extraction module first performs discrete cosine transform to realize the time-frequency conversion of the image,and then arranges and reorganizes the acquired frequency domain features,so as to refine the shallow features in the model learning and add constraints for the reconstruction model;the multichannel feature selection module uses the adaptive learning method to weight the features of each channel,so as to characterize the corresponding effect of different feature channels on the image composition and the relationship between the features of the same layer;The whole network uses long-skip and short-skip connection to deepen the network depth and extract deep features.In this paper,the above algorithm model is compared with the current popular interpolation algorithm,sparse coding algorithm,SRCNN and EDSR algorithm.Experimental results show that the algorithm proposed in this paper achieves good reconstruction results in both objective index and visual perception for general scene images and deep undercooling melt images.At the same time,it also achieves better results in Task-based evaluation for deep undercooling melt images than the above comparison algorithm.