FAST (Chen et al., 2021) starts from the observation that the post-processing of segmentation detectors, not the network, had become the bottleneck. Two changes follow:
- Minimalist kernel representation. The network predicts a single-channel map of shrunken text kernels. Text instances are recovered by dilating those kernels, and the dilation is written as a GPU-parallel operation (a max-pool-style pass) instead of CPU contour code, so post-processing costs almost nothing.
- A searched backbone. Instead of reusing an ImageNet classification network, the authors search an architecture (TextNet) for the text detection task itself and find it produces stronger features at the same cost.
The abstract reports 81.6 % F-measure at 152 FPS for FAST-T on Total-Text, and more than 600 FPS with TensorRT. The Hugging Face Transformers library later added a FAST implementation, which is why the model shows up in “text detection” search results above older work.
Where it fits
FAST is the right reference point when the question is “how fast can arbitrary-shape detection be on a GPU”. For CPU and browser deployment, the lighter DB-family models used by PaddleOCR remain more common because their ONNX exports are small and well supported, which is why the detector on this site runs those.
Citation
@article{chen2021fast,
title = {FAST: Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation},
author = {Chen, Zhe and Wang, Jiahao and Wang, Wenhai and Chen, Guo and Xie, Enze and Luo, Ping and Lu, Tong},
journal = {arXiv preprint arXiv:2111.02394},
year = {2021}
}