Recently, the Feyn team officially launched a new SOTA (state-of-the-art) background removal model FeyNoBg, and simultaneously released the corresponding NoBg training and inference library in the open-source community, aiming to provide developers with a more lightweight and efficient image segmentation solution.

image.png

Excellent background removal algorithms not only need to accurately identify the foreground but also handle edge details (such as hair, semi-transparent veils, etc.) perfectly. To achieve this goal, the development team conducted deep optimization based on the open-source BiRefNet architecture. By expanding the modules in the key feature extraction stage from 18 to 24, the model's parameter count increased slightly from 222M to 263M, while retaining the original pre-trained weights, endowing the model with stronger learning and information retention capabilities.

image.png

In terms of training strategy, the team broke through the limitations of a single dataset, collecting and cleaning 10 datasets, including complex scenes, high-resolution images, camouflaged objects, and anime, totaling 26,100 images for multi-source mixed training. After careful data ratio adjustment and unified binary mask conversion, the model performed outstandingly in multiple authoritative benchmark tests. Data shows that FeyNoBg achieved the best S-measure score in four out of eight mainstream benchmarks, and remained closely behind the leader in other tests, with performance almost equal.

image.png

To facilitate developers' application, the team also open-sourced the underlying Python library NoBg. Practical tests show that, in batch processing mode, this library achieves higher throughput, lower latency, and smaller memory usage compared to the original BiRefNet. Currently, users can obtain the model via FeyNoBg on Hugging Face, or explore GitHub in-depth.