Traditional optical property measurement techniques typically rely on complex detection systems and are easily affected by environmental disturbances. To solve these problems, this paper proposes a low?cost and disturbance?resistant single?image inversion method for two?layer dental optical properties, achieved by integrating physical priors with a deep learning fusion framework. By modeling the optical transmission process within the tooth structure, the optical properties of teeth inversion network (OPTNet) is proposed, which preserves multi?layer features through residual concatenation and incorporates a spatial attention mechanism to focus on critical diffuse reflectance regions, thereby enabling precise prediction from a single image. Natural teeth are collected and corresponding restorative samples are fabricated, an optical acquisition system is then assembled to capture the surface light distributions of these samples, thereby constructing the clinical reflectance dataset. On clean clinical samples, the proposed method achieves an average relative error (MRE) below 0.2% and a coefficient of determination (R2) above 0.99. To evaluate robustness, tests are also conducted on samples with measurement noises. Results show that, OPTNet reduces the MRE of the reduced scattering coefficient and absorption coefficient by 55.08% and 72.38% compared to a fully connected network (FCN), respectively, and by 43.01% and 46.77% compared to a traditional convolutional neural network (CNN), respectively. The proposed physics-prior-driven deep learning approach for biophysical measurement achieves high-precision inversion of dental optical properties without requiring complex optical instrumentation, offering a novel solution for tooth color matching and customized restorative treatments.