Disparity Recovery in Stereo Matching Based on a Conditional Deep Convolutional Generative Adversarial Network

Deep learning technology is widely used everywhere, such as in the medical field, autonomous assistance systems, or recognition schemes. Depth-variant video for higher-level machine vision can provide depth perception and dynamic B-roll for accurate autonomous range. However, the stereo navigation technology of simultaneous localization and mapping (SLAM) for robots due to the poor quality of the depth map produced by the sensor affects the misjudgment of the surrounding area or obstacle distance, which can cause serious impacts, such as collisions on robots. Therefore, in this paper, we will design a feasible method based on a conditional deep convolutional generative adversarial network (CDCGAN) to repair the error disparity. The experimental results demonstrated that the proposed method can maintain a good quality of the corrected disparity maps by computing the structural similarity index (SSIM) and peak signal-to-noise ratio values (PSNR).