A New Fuzzy Smoothing Term Model For Stereo Matching.

This article focuses on developing a fuzzy smoothing term model for binocular stereo matching to improve disparity estimation in low-quality and vague images. The proposed approach constructs fuzzy relationships at three levels: between pixels and superpixels, among superpixels via a fuzzy network grounded in graph theory (supported by five proven theorems), and between pixels themselves, enabling the model to handle uncertainty inherent in indistinguishable image regions. Experimental results, including ablation and comparative studies against state-of-the-art algorithms such as PSM-Net, AA-Net, and RAFT-Stereo, demonstrate that this fuzzy model significantly enhances disparity accuracy, especially in scenes with moderate to severe vagueness, outperforming deterministic models in boundary preservation and overall robustness. The study also discusses computational complexity, noting that while the algorithm has quadratic complexity relative to pixel count, it remains computationally feasible for low-resolution images. This work has potential applications in autonomous vehicle vision under challenging weather conditions like dense fog and heavy rain.