Boosting Stereo Image Noise Removal by Learning Uncertainty and Enriched Features

Bingcai Wei, Hui Liu, Chuang Qain

Abstract: Stereo image denoising is crucial to improve perceptual quality and autonomous driving perception. Existing methods often fall short in accurately estimating the uncertainty inherent in noisy data, leading to suboptimal denoising outcomes. To address this, we introduce a novel framework named the Cross-view Uncertainty-aware Gradient-assisted Fusion Network (CUGFNet), which pioneers the integration of uncertainty estimation, gradient-assisted feature extraction, and frequency interaction-based feature fusion for superior stereo image de-noising. CUGFNet’s primary innovation lies in its Cross-view Uncertainty-aware Probabilistic Feature Extractor (CUPFE), which provides a probabilistic interpretation that enables the model to make informed decisions during the denoising process. The Gradient-assisted Attentive Feature Aggregation Module (GAFAM) enhances the network’s ability to detect details and structure in stereo images by using spatial gradients to highlight important features. Extensive experimental validation across a variety of stereo image datasets substantiates the effectiveness of CUGFNet. This model delivers not only higher PSNR and SSIM scores but also demonstrates superior robustness against diverse noise levels.