Polarization-fused Binocular Stereo Vision Enhancement Method

Binocular stereo vision has become a mainstream three-dimensional imaging solution in medical imaging analysis and industrial inspection because of its low cost， passive sensing and real-time processing capability. However， conventional stereo matching algorithms exhibit intrinsic limitations when confronting highly reflective and weakly textured objects such as polished ceramics， metals or water-like surfaces. Specular highlights saturate the sensor and obscure surface micro-structure， while the lack of natural texture causes a reduction in reliable feature points， resulting in mismatches， depth discontinuities and holes in reconstructed point clouds. Existing methods can be broadly categorized into active feature augmentation， which relies on projected patterns， added particles or structured light， and passive algorithmic optimization， which refines feature detection or incorporates learning-based priors. Active schemes increase system complexity， cost and sensitivity to environmental illumination， whereas passive schemes remain vulnerable to strong glare and texture scarcity， and often depend on prior knowledge of material properties. Polarization imaging， which is sensitive to surface reflection mechanisms and material characteristics， offers a promising avenue， yet current polarization-stereo fusion strategies still suffer from reflection-model inaccuracies， normal ambiguity， dependence on refractive index， and error accumulation during normal integration. These limitations motivate the present work， which aims to develop a polarization-augmented binocular stereo framework that alleviates specular interference and texture deficiency without introducing additional active devices.The proposed method constructs a polarization-fused binocular stereo vision pipeline that combines physically based reflection decomposition with multimodal feature enhancement and conventional triangulation. A stereo pair of polarization-sensitive cameras acquires synchronized images at four polarization angles （0°， 45°， 90° and 135°）， and the system is calibrated using a polarization chessboard and Zhang's method to obtain intrinsic and extrinsic parameters and to ensure accurate epipolar rectification. At the image-formation level， the observed intensity at each polarization angle is modeled as a superposition of specular and diffuse components. Multi-angle intensities are assembled into an observation matrix and approximated as low rank via singular value decomposition， which factorizes the data into coefficient and reflection matrices. An unknown transformation matrix is then optimized by minimizing the Pearson correlation between specular and diffuse estimates， yielding a separation in which diffuse components retain texture while specular energy is strongly attenuated. In parallel， Stokes-parameter analysis of the four polarization channels provides the degree and angle of polarization， from which azimuth and zenith angles of surface normals are inferred under an assumed refractive index. Intensity-based SURF descriptors are extracted on the glare-suppressed diffuse image， polarization-normal descriptors are computed from the normal map， and both standardized descriptors are concatenated into a joint feature representation that guides stereo matching， disparity computation and depth estimation.Extensive experiments on highly reflective ceramic and metallic targets demonstrate that the proposed polarization-enhanced stereo framework markedly improves both feature richness and three-dimensional reconstruction quality. Using a dual MER2-502-79U3M polarization camera system with precise baseline control and synchronized acquisition of four-angle polarization sequences， the authors first examine feature detection on ceramic samples with strong specular highlights. SURF-based analysis shows that， before enhancement， only a few dozen stable feature points are available in each region， whereas after specular-diffuse separation and polarization-based feature augmentation， the number of keypoints increases by roughly an order of magnitude and feature coverage improves correspondingly. Gray-level histograms confirm that the proportion of pixels in high-intensity highlight ranges drops sharply after processing， indicating effective suppression of saturated regions and recovery of latent texture. Depth profiles along selected cross-sections reveal that original stereo reconstructions exhibit pronounced oscillations and artificial depressions in glare-dominated areas， while the enhanced method yields smooth， continuous depth curves that closely approximate the true object geometry. Quantitative distance measurements between reconstructed and physically measured point pairs show absolute errors below 5 mm and relative errors below about 0.5. On metallic heater targets， the method fills the vast majority of point cloud holes， elevates surface completeness to almost full coverage， reduces highlight-dominated pixel proportions， and increases local feature density by several tens of percent in difficult regions.Comparative evaluation against three representative baselines further highlights the advantages and limitations of the proposed approach. Traditional binocular stereo vision， although computationally efficient， exhibits the lowest structural similarity because it relies solely on intensity texture and is easily disrupted by glare and weak-patterned surfaces. Pure polarization-based three-dimensional imaging improves detail recovery by exploiting polarization-derived normals， yet its accuracy is constrained by unknown or spatially varying refractive indices and by error accumulation during normal integration. A hybrid method that fuses polarization with deep neural networks achieves higher reconstruction quality by resolving angular ambiguities， but requires substantial computational resources and long inference times， which challenge real-time deployment. In contrast， the proposed specular–diffuse separation and polarization-augmented stereo matching strategy achieves the highest structural similarity index while maintaining moderate runtime， thereby realizing a favorable balance between precision and efficiency without resorting to additional active projection hardware. Overall， the work shows that incorporating polarization cues at the level of reflection modeling and feature description can effectively overcome the dual challenges of high specular reflection and texture scarcity， significantly reducing point cloud hole rates and enhancing geometric fidelity. The method therefore offers a promising solution for high-precision， passive three-dimensional measurement of highly reflective， weakly textured objects in industrial inspection and related application scenarios.