EGSNet: An Efficient Glass Segmentation Network Based on Multi-Level Heterogeneous Architecture and Boundary Awareness

Existing glass segmentation networks have high computational complexity and large memory occupation, leading to high hardware requirements and time overheads for model inference, which is not conducive to efficiency-seeking real-time tasks such as autonomous driving. The inefficiency of the models is mainly due to employing homogeneous modules to process features of different layers. These modules require computationally intensive convolutions and weight calculation branches with numerous parameters to accommodate the differences in information across layers. We propose an efficient glass segmentation network (EGSNet) based on multi-level heterogeneous architecture and boundary awareness to balance the model performance and efficiency. EGSNet divides the feature layers from different stages into low-level understanding, semantic-level understanding, and global understanding with boundary guidance. Based on the information differences among the different layers, we further propose the multi-angle collaborative enhancement (MCE) module, which extracts the detailed information from shallow features, and the large-scale contextual feature extraction (LCFE) module to understand semantic logic through deep features. The models are trained and evaluated on the glass segmentation datasets HSO (Home-Scene-Oriented) and Trans10k-stuff, respectively, and EGSNet achieves the best efficiency and performance compared to advanced methods. In the HSO test set results, the IoU, F_β, MAE (Mean Absolute Error), and BER (Balance Error Rate) of EGSNet are 0.804, 0.847, 0.084, and 0.085, and the GFLOPs (Giga Floating Point Operations Per Second) are only 27.15. Experimental results show that EGSNet significantly improves the efficiency of the glass segmentation task with better performance.