Research On Super-resolution Reconstruction Method Of Droplet Image Based On Deep Learning

Super resolution(SR)is an important part of image processing and a hot issue in the field of image processing.The object of this paper is low resolution(LR)material droplet image collected on the electrostatic suspension and vertical vacuum drop tube dropping experimental equipment.In view of the existing SR method,the edge blur and texture details are not enough in the droplet image.In this paper,the convolutional neural network is used to solve the problem of fuzzy edges and insufficient texture details,CNN is the main method to study the super-resolution reconstruction method of high resolution(HR)droplet image based on low resolution image data.1)Based on different features fusion methods of CNN,two kinds of image reconstruction algorithms are proposed: DRSub Net and Dense Sub Net.The two algorithms use CNN to extract features on low resolution droplet images.Compared with the operation on interpolation enlarged images,the complexity of calculation is reduced.The feature fusion is carried out by adding the matrix of feature graphs of residual network and the connection principle of feature graph channel of dense network,and combining the sub-pixel convolution sampling module to construct SR algorithm model.DRSub Net makes full use of the complementarity of high-level and low-level features,adds the semantic of low-level features in SR model and adds more spatial information in the high-level features,and increases the model volume and the range of sense field.Dense Sub Net further optimizes the propagation of information flow,and can use less parameters than DRSub Net,and realizes feature reusability by feature connection.The above two algorithms improve the learning ability of the image reconstruction model,and optimize the gradient propagation effect in the training process,making the network model more easy to train.2)Based on CNN's attention mechanism and cyclic structure,two kinds of image reconstruction algorithms are proposed: RSASR and UASR.RSASR extends the multi-scale sense field by using less convolution layer through branch cycle CNN structure.It can reduce the weight parameter significantly when the ideal reconstruction precision is obtained.The RSASR can obtain more abundant spatial semantic information in LR image by using self attention mechanism,and be good at capturing long-distance features,which is conducive to improve the reconstruction effect of droplet image.UASR uses semantic segmentation to identify high frequency regions such as contour and texture in input image,normalizes the processing to get attention weight matrix,weights it into DRSub Net reconstruction image,focuses on reconstruction of high frequency contour and texture area,and suppresses noise in reconstruction results.The experimental results show that,compared with the existing methods,the proposed image super-resolution reconstruction method has better reconstruction effect for low resolution droplet image in terms of peak signal-to-noise ratio,structural similarity and the reconstruction accuracy of the specific diameter and area of the droplet image in this paper.