DeepRFWT: Approach to mitigate defocus blur effects in DIC-based strain analysis

Abstract

Digital Image Correlation (DIC) technology is confronted with considerable challenges due to defocus blur in deformation measurements. Out-of-plane displacement-induced defocus degradation severely compromises DIC accuracy, particularly in microscopic applications requiring sub-pixel precision. To address this issue, DeepRFWT (Deep Residual Fourier-Wavelet Transform), a deep learning-based deblurring algorithm was specifically designed for speckle image restoration. The algorithm integrates three innovative components: 1) the Multi-Scale Feature Enhancement Module (MSFE) for spatial context preservation, 2) the Multi-Transform Domain Encoder-Decoder (MTDED) for dual-channel frequency-spatial domain processing, and 3) the Frequency Domain Spatial Transformer (FDST) for high-frequency information recovery. Comprehensive validations demonstrate superior performance over state-of-the-art methods, achieving 26.70 dB Peak Signal-to-Noise Ratio (PSNR)/0.829 Structural Similarity Index Measure (SSIM) on the Speckle Blur Dataset (SBD), and 26.21 dB PSNR/0.819 SSIM on the Dual-Pixel Defocus Deblurring Dataset (DPDD) with 11.49 M parameters. Micro-displacement experiments confirm exceptional robustness under varying defocus conditions (0.5–1.5 mm), yielding reconstruction errors in the 10⁻⁵ to 10⁻⁴ mm range. Engineering validation via polyurethane 90A tensile tests show DIC strain measurement relative errors below 1.65 %, verifying practical efficacy.