Research On Super Resolution Of Metallic Sample Droplet Image

Metastable materials usually present the characteristics of rich structure and singular performance,and the research on them has become a hot topic in the world.The metal sample droplet is a metastable material obtained through a containerless,microgravity and deep supercooling process.In order to obtain the morphological change of the sample droplet during the free fall process,a high-speed camera was utilized to photograph the falling process.Due to the limitations of the internal memory and data transmission speed of high-speed cameras,the image resolution is very low,which cannot fully reflects the complete form of the material sample.It is an urgent problem for researchers to take reasonable scientific methods to reconstruct images into high-resolution images.SR(Super-resolution reconstruction,SR)is an important part of image processing,and it has become a hot issue in scientific research today.When the current super-resolution algorithms are applied to the droplet image,it often reflects the phenomenon of blurred edges and insufficient texture detail information.This paper adopts the method of multi-scale feature fusion,the single-frame image super-resolution reconstruction method based on frequency domain feature learning,and the multi-frame image super-resolution reconstruction method based on WGAN(Wasserstein Generative Adversarial Nets,WGAN).Three super-resolution reconstruction algorithms suitable for droplet images are implemented.The main work of this paper is as follows:(1)A single-frame image super-resolution reconstruction algorithm based on hierarchical feature fusion network is proposed,which effectively expands the receptive field and more fully extracts the image context information.The algorithm effectively expands the receptive field and extracts image context information more comprehensively.First,a multi-core multi-scale residual block is proposed,and five different sizes of features are extracted for low-resolution images.Then,according to the decomposition of convolution,a single-core multi-scale residual block is constructed,which reduces the network size,reduces the amount of network training parameters,and improves the network reconstruction performance.This algorithm solves the problem of the huge amount of parameters caused by the large-scale convolution kernel to the network in the multi-core multi-scale residual block.Experimental results show that under three different magnification factors of ×2,×3 and×4,the PSNR(Peak Signal to Noise Ratio)and SSIM(Structural Similarity,SSIM)indicators of the reconstructed metal sample droplet images are improved.In terms of visual effects,it can not only reconstruct the clear edge of the droplet,but also accurately extract the contour and calculate the diameter and area with small errors.(2)From the perspective of frequency domain features,a single-frame image super-resolution reconstruction algorithm based on the frequency domain channel attention mechanism is proposed,which is applied to the reconstruction of the molten drop image of deep subcooled metal samples.First,the frequency-domain channel attention block is utilized to fully extract multiple frequency-domain component information in low-resolution images,and an adaptive learning channel is performed on it.Then the Non-Local module is utilized to effectively capture the feature dependence between the long distances of the network,and to enhance the correlation between the shallow features and the deep features.Finally,the network is deepened by long and short hop connection to extract deeper features.The experimental data show that the algorithm has achieved a certain improvement in PSNR and SSIM performance.At the same time,the visual effect can clearly reconstruct the image edge contour,effectively reduce noise blur and other phenomena,which verifies the effectiveness and applicability of the algorithm.The algorithm achieves a better effect in edge reconstruction.(3)From the perspective of sequence image reconstruction,a multi-frame image super-resolution reconstruction algorithm based on densely connected adversarial network is proposed,which is applied to metal sample droplet image reconstruction.The algorithm combines the advanced spatial transformation module STN(Spatial Transformer Networks,STN)in the SRGAN(Super-Resolution Generative Adversarial Network,SRGAN)network,and then integrates the associated information of multiple consecutive frames,and conducts stable training through the WGAN network.This algorithm improves the subjective and objective evaluation of reconstructed images,while focuses on the reconstruction of the sequence deformation process.Experiments show that the proposed algorithm can clearly reconstruct the deformation process of the metal sample droplet.The reconstruction results improve the PSNR and SSIM indicator.Compared with the single-frame algorithm SRGAN,the PSNR of the model reconstruction is improved by about 7.77 d B,and the SSIM is improved by about 0.0399.Compared with the current advanced multi-frame algorithm VESPCN,the PSNR of the model reconstruction in this paper is improved by 3.2d B,and the SSIM is improved by about 0.0004.