An Image Reconstruction Algorithm Based on Frequency Domain for Deep Subcooling of Melt Drops

Abstract

In a space environment, liquid alloys are in a thermodynamically metastable state, which facilitates research on the material structure and thermophysical properties of deep undercooling melt. Limited by the cost and technology of performing experiments in space, researchers developed electrostatic levitation that uses a drop pipe device to simulate the space environment. A high-speed camera was used to capture the falling image of the deep undercooling melt to study the melting and solidification process. Due to the exposure time and hardware limitations of the image acquisition equipment, the image resolution of the deep undercooling melt is low, which is not conducive to studying the thermophysical properties and solidification interface of the melt. Software design methods, such as super-resolution reconstruction, can more accurately reconstruct image contour information and effectively improve the image resolution. The most current deep learning-based super-resolution reconstruction algorithms directly perform Y-channel or Y, Cb, and Cr three-channel learning on the reconstructed image. This is insufficient in terms of providing more prior information to solve the super-resolution reconstruction. In this study, a single-frame image super resolution reconstruction network that is based on frequency-domain feature learning is proposed. It builds a time–frequency transformation layer at the front end of the neural network and uses the frequency to realize the neural network in the frequency domain. To evaluate the super-resolution reconstruction performance, the proposed algorithm is compared with the current mainstream interpolation, sparse coding, super resolution convolutional neural network, and enhanced single-image super-resolution deep residual algorithms. The proposed algorithm achieves good reconstruction effects on deep
undercooled melt images in terms of the objective evaluation and visual perception. At the same time, the peak signal-to-noise ratio and structural similarity index measure achieved results that exceed the aforementioned comparison algorithms.