SFMHANet: Surface Fitting Constrained Multidimensional Hybrid Attention Network for Aero-Optics Thermal Radiation Effect Correction

Abstract

When an aircraft is flying at hypervelocity in the atmosphere, the airflow and the optical cowl rub against each other, and the airflow's kinetic energy in the boundary layer is transformed into thermal energy, which causes the cowl's surface temperature to rise nonuniformly and produces thermal radiation interference with the imaging system of the detector. In practical application scenarios, the aero-optical thermal radiation patterns in degraded images are not fixed, and types of aero-optics thermal radiation are more variable and complex. In order to handle multiple types of aero-optics thermal radiation effects effectively and to combine the advantages of image prior constraints and deep learning networks, we propose a surface fitting constrained multidimensional hybrid attention aero-optics thermal radiation correction network (SFMHANet) in this article. First, according to the characteristics of the aero-optics thermal radiation bias field belonging to low frequency, we initially estimate the aero-optics thermal radiation bias field using biharmonic spline interpolation surface fitting based on wavelet decomposition. Second, we design a multidimensional hybrid attention aero-optics thermal radiation correction network constrained by the supervision of aero-optics thermal radiation bias field for asymmetric information exchange. Finally, to achieve cross-dimensional information interaction of features, we propose a multidimensional hybrid attention module, a second-order pooling channel attention block, and a cross-convolution spatial attention block in the correction network. According to experiments on aero-optics thermal radiation correction of simulated and real degraded images, the SFMHANet can correct the aero-optics thermal radiation effects of multitype degraded images in comparison to other existing methods.