Journal of Advances in Mathematics and Computer Science

Stereo Disparity estimation is widely used in computer vision applications that require reliable spatial information, but raw disparity maps often contain artefacts caused by mismatches, occlusions, illumination variation, low-texture regions, and noise. This study presents a guided-filter-based refinement approach for reducing artefacts in stereo Disparity estimation while preserving computational efficiency. The proposed pipeline uses rectified stereo image pairs from the Middlebury Stereo Vision Dataset and applies Semi-Global Matching to generate an initial raw disparity map. A left-to-right consistency check is then used to identify unreliable disparity values. Detected outliers are classified as mismatches or occlusions, followed by invalid-pixel filling using neighbourhood-based strategies. Finally, guided filtering is applied to refine the filled disparity map while preserving object boundaries. The method was evaluated under normal, bright and Dark illumination scenarios using mean absolute error, root mean square error, bad-pixel measures, and coverage percentage against available ground-truth disparity maps. In the normal lighting scenario, guided filtering with smoothing values of 0.002 and 0.003 reduced error measures compared with the raw Semi-Global Matching output. The best refined result reduced MAE from 4.1199 to 2.8308 and RMSE from 12.100 to 8.7966. In the bright lighting scenario, the method improved the raw disparity output by reducing MAE from 8.8559 to 7.2154 and RMSE from 17.739 to 14.382. The findings indicate that outlier detection, invalid-pixel filling, and guided filtering can improve disparity-map quality under the evaluated conditions. Further validation on larger datasets and real-time hardware platforms is needed to confirm wider applicability.