Surface diffuse reflectance \(\:{R}_{d}\) is affected by the sample geometric contour, which subsequently affects the optical parameter inversion and parameter distribution diagram. To reduce the contour impacts, a geometry-aware correction method guided by imaging geometry and physical response characteristics is proposed. A spatial frequency domain imaging (SFDI) system was built to obtain the SFDI image of the sample. Surface height reconstructed from phase information is used to derive the imaging angle, and the Lambertian model is applied to correct \(\:{R}_{d}\) , followed by compensation for sample-to-camera height variations. This correction compensates for geometry-induced measurement deviations. Finally, the absorption coefficient \(\:{\mu\:}_{a}\) and the reduced scattering coefficient \(\:{\mu\:}_{s}^{{\prime\:}}\) were inferred using the established neural network model. The model shows the accuracy of \(\:{\mu\:}_{a}\) and \(\:{\mu\:}_{s}^{{\prime\:}}\) for the solid phantom is 96.80% and 98.09%, respectively, and the inversion time for a single image is 0.83 s. The contour correction method was validated using the solid phantom. The results showed that the relative error of \(\:{\mu\:}_{s}^{{\prime\:}}\) after correction is within 1%, the standard deviation is reduced by 78.38%, and the correction is more effective within the angle range from 0 \({}^\circ\:\) to 53 \({}^\circ\:\) . The proposed method is further applied to hidden apple bruise detection across different fruit regions, including apple body, stem and calyx areas using deep learning algorithms. Compared with before correction, the mAP value increases by 2.2%. The results confirm that the proposed method effectively reduces contour influence on optical parameter inversion, providing a practical approach for non-invasive quality assessment using imaging techniques.
