xuebinqin/U-2-Net

Employs a nested U-structure architecture with residual U-blocks at multiple scales to capture both fine details and global context for salient object detection. Provides pre-trained models optimized for real-time inference, enabling integration into mobile applications (iOS/Android via CoreML and Swift) and web services for background removal, portrait matting, and semantic segmentation tasks. Supports PyTorch framework with straightforward inference APIs for custom model training and deployment across edge devices.

9,672 stars. No commits in the last 6 months.