Frequency enhanced multi-scale fusion network for high temperature speckle image restoration

Abstract High temperature optical imaging is highly susceptible to turbulence induced thermal haze, which degrades speckle contrast, blurs texture details, and distorts structural information. These effects significantly hinder the accuracy of digital image correlation (DIC) in high temperature measurement scenarios. Existing deep learning based dehazing methods, mainly developed for natural scenes and struggle to address the frequency domain attenuation and structural distortions caused by thermal disturbances, limiting their ability to restore fine details in thermally degraded speckle images. To address this challenge, we propose a multi-scale frequency enhanced network that performs haze removal and reconstruction through frequency compensation and cross-scale feature fusion. Specifically, frequency enhanced residual blocks compensate for the frequency domain energy loss induced by thermal haze, while a multi-scale feature aggregation module mitigates structural degradation commonly caused by traditional sampling based fusion, thereby improving the quality of cross-scale information integration. Experimental results demonstrate that the proposed model achieves superior performance, reaching 31.05 dB PSNR and 0.9649 SSIM, and reducing the maximum relative error in high temperature DIC displacement measurement from 8.9% to 4.3%. This method can restore clear speckle images under 800 °C conditions, meeting the requirements for high precision measurements, and provides an effective solution for thermal disturbance suppression in high temperature optical measurements.