Cross-Attention Joint Denoising and Semantic Segmentation for Robust Image Analysis
2169-3536
Abstract
Semantic segmentation assigns class labels at the pixel level, forming a key part of various computer vision applications, including autonomous driving, medical imaging, and robotics. However, segmentation accuracy often drops significantly under noise corruption in real-world images. Traditional image restoration methods primarily focus on perceptual quality, which does not always result in improved segmentation performance. To bridge this gap, we propose a novel cross-attention joint learning framework that explicitly combines features from denoising and segmentation networks through a context-aware module. Our approach also introduces a mid-frequency loss to guide the denoising process toward preserving structural details essential for semantic boundaries. Extensive experiments on the VOC2012 dataset under various noise types, including Gaussian, salt-and-pepper, JPEG compression, and motion blur, demonstrate that our method outperforms existing pipelines in segmentation accuracy and robustness. These results confirm the practical effectiveness of our cross-attention joint optimization strategy, establishing a new benchmark for robust semantic segmentation in degraded visual environments.