Enhancing Glass Segmentation Accuracy via Boundary-Context Guided Physics-Aware Modeling

Transparent objects such as glass pose significant challenges and potential safety risks to vision systems due to their transparency and complex light interactions, which often lead to blurred boundaries and contextual ambiguity. Existing methods typically suffer from weak edge responses and insufficient cross-scale context integration. To address these issues, we propose the Boundary-Context Guided Glass Detection Network (BCGNet), a novel encoder-decoder architecture that combines global context modeling with boundary-aware refinement for accurate glass segmentation. Specifically, we design a dual-path Boundary Feature Refinement (BFR) module, where the spatial branch captures position-sensitive boundary features, and the channel branch, inspired by physical light propagation, leverages optical flow-guided feature warping to model geometry-aware deformations along transparent edges, thereby mitigating semantic confusion. Additionally, we develop a Multi-scale Context Aggregator that integrates a Contextual Cross-Attention Module (CCAM) and a Depthwise Separable Feed-Forward Network (D-FFN) through hierarchical feature interaction. CCAM captures long-range dependencies via dual-stream attention, while D-FFN reconstructs boundary-sensitive local representations using depthwise separable convolutions. Extensive experiments on four public datasets demonstrate that BCGNet consistently outperforms existing methods, validating its robustness and effectiveness in complex real-world scenarios.