Disparity refinement based on least square support vector machine for stereo matching

Disparity refinement is a crucial step in obtaining accurate disparity for stereo matching method. Outlier disparity value still exists in some areas (such as feeble texture and discontinuous regions), even the advanced stereo matching algorithm based on deep learning. To address this issue, a novel disparity refinement method based on the least square support vector machine (LSSVM) is proposed. In this method, the least square support vector machine model is first applied to every horizontal line of the obtained initial disparity map to model the disparity values, corresponding image color values, and coordinates of pixels. According to corresponding feature, the predicted disparity value of each pixel is calculated by this regression model. Subsequently, the outliers are detected and removed based on the residual between the real and predicted disparity value for obtaining more accurate initial disparity map. Then, along a horizontal line of the disparity map, the LSSVM with different parameters is applied to train the valid disparity values and its feature for obtaining the trained regression model. Finally, the invalid disparity values are redefined by the trained regression model. Experimental results demonstrate that the proposed method shows a better performance compared with current some disparity refinement methods. When the proposed algorithm is implemented on the disparity map of the deep learning method, the error rate has decreased and the maximum decline rate is 4.3 and 3.0 in nonocc and all regions, respectively.