Real-time semantic segmentation network via bidirectional feature alignment

Real-time semantic segmentation network via bidirectional feature alignment

연구 분야: Strategies

학회: Journal of Real-Time Image Processing

We propose a real-time semantic segmentation network based on bidirectional feature alignment, termed BFASNet. Existing methods, such as BiSeNet, learn semantic and detailed information through context and spatial paths, respectively. However, they often fail to adequately address the feature alignment issue between these paths, leading to feature misalignment during progressive downsampling and context information fusion, which adversely affects segmentation performance. To address this, we introduce a novel network that achieves feature alignment between paths by learning flow fields. Our approach incorporates two key modules: the Gated Bidirectional Feature Alignment Module (GBFAM) and the Gated Context Feature Alignment Module (GCFAM). GBFAM employs a learnable interpolation strategy and a gating mechanism to bidirectionally align high- and low-resolution features, reducing noise and semantic discrepancies during fusion. GCFAM adaptively selects personalized contextual information for each pixel, enhancing the alignment of contextual features. Extensive experiments on the Cityscapes and ADE20K datasets demonstrate that BFASNet achieves mean Intersection over Union (mIoU) scores of 80.47% and 45.47%, respectively, validating its effectiveness. The network also achieves a frame rate of 33 FPS on the Cityscapes validation set, meeting the demands of real-time segmentation.