To meet the specific demands of cultural heritage digitization， such as grotto murals and metallic handicrafts， optical three-dimensional reconstruction often faces high dynamic range challenges caused by the coexistence of complex surface texture and strong reflections. Under conventional sensing pipelines， these challenges lead to key issues including signal saturation， holes in low signal-to-noise ratio regions， and outliers in the reconstructed point cloud. Parallel Fourier single-pixel imaging is inherently suitable for complex texture and reflection conditions. However， in practical sensing chains it is still jointly constrained by the loss of physical information introduced by conventional image processing pipelines and by limited effective dynamic range. As a result， distortions in frequency domain measurements propagate and accumulate in the stages of light transport estimation and three-dimensional reconstruction， thereby limiting the completeness and stability of high dynamic range surface reconstruction.This work adopts parallel Fourier single-pixel-imaging-based three-dimensional reconstruction as the overall framework and treats the per-pixel light transport coefficient as the key intermediate quantity for geometric recovery. On this basis， a robust estimation and regularized correction method guided by frequency domain distribution characteristics is developed. First， a frequency domain observation model consistent with phase-shift demodulation is established on raw measurements， explicitly modeling noise propagation in the form of frequency domain estimation errors. To justify the necessity of using raw data in terms of physical consistency and controllable error behavior， typical sensing pipelines are compared， showing that outputs after conventional image processing alter the amplitude ratio relationships across phase-shifted measurements， which is detrimental to subsequent spectrum estimation and peak localization. Raw data are therefore used instead of conventionally processed images. Second， the noise impact and spectrum degradation mechanisms under low-photon conditions are addressed， and a decomposed frequency domain optimization strategy is proposed. Magnitude and phase are decoupled in the frequency domain； robust estimation is applied to suppress noise-induced abnormal perturbations； phase preserving adaptive denoising is performed to lower the noise floor and emphasize reliable frequency components； and regularized restoration is finally applied to refine the spectrum and improve the reliability of effective frequency domain structures， thereby obtaining a more faithful light transport response. Finally， the corrected light transport estimates are used for camera-projector geometric recovery to generate point clouds via triangulation， and a geometry-texture consistent color mapping is further constructed by combining the Bayer channel responses in raw data with the reconstructed light transport response.Experimental results demonstrate that the proposed frequency domain robust estimation and regularized correction method significantly improves the stability and completeness of three-dimensional reconstruction in challenging environments where low photon conditions and strong reflections coexist. Compared with reconstructions based on conventional data processing pipelines， the method effectively suppresses frequency domain noise perturbations， reduces spurious peak elevation and sidelobe spreading in the light transport coefficients， and markedly enhances both the energy concentration and positional stability of the primary peak associated with direct illumination. For highly reflective targets such as a metallic step specimen， the method substantially alleviates holes and outliers in specular regions and improves overall geometric accuracy by 54.3% relative to the conventional processing mode. For textured targets with clear material or color boundaries such as a colored plaster bear， the method preserves greater sharpness and continuity of texture boundaries and is less sensitive to imaging conditions. Overall， by improving the reliability of effective frequency domain structures， the proposed method significantly reduces the accumulation of frequency domain measurement distortions induced by sensing chain nonlinearity and photon noise in the final three-dimensional results， achieving unified improvements in point-cloud completeness and accuracy.The proposed raw based frequency domain robust estimation and regularized correction approach can serve as a physically grounded complement to existing high dynamic range strategies. It substantially enhances the reliability of light transport peaks， point-cloud completeness， and texture expression consistency in three-dimensional measurement of complex reflective objects， providing a practical technical pathway that balances low-level physical consistency with high dynamic range capability for noncontact precision measurement of high value targets such as cultural heritage preservation and provenance tracing.