Scene text detection is the problem of finding text in photographs of the world rather than in scanned documents. The field moved fast between 2016 and 2021, then settled into toolkits. This is the sequence of ideas, with each paper’s own headline number, so the models table makes sense.
Before 2016: engineered features
Stroke Width Transform (2010) and MSER-based pipelines found candidate character regions with hand-designed rules and filtered them with classifiers. They worked on clean signage and fell apart on incidental text. ICDAR 2015’s incidental benchmark, with blurred and rotated words from a wearable camera, made the limitation obvious.
2016: text as a sequence
CTPN (Tian et al., ECCV 2016) adapted the Faster R-CNN recipe to text: fixed-width vertical anchors slide along a line, a bidirectional LSTM passes context between them, and neighbouring proposals are linked into lines. It reported 0.88 F on ICDAR 2013 and 0.61 on ICDAR 2015. Its weakness was orientation: the anchors only move up and down.
2017: one stage, rotated boxes
EAST (Zhou et al., CVPR 2017) removed proposals altogether. A fully convolutional network predicts, at every pixel, a text score and the geometry of the surrounding word as a rotated box or quadrilateral; non-maximum suppression merges them. The abstract reports 0.782 F at 13.2 FPS on ICDAR 2015. Rotation was solved; curves and very long lines were not.
Contemporaries worth knowing: TextBoxes and TextBoxes++ (SSD-style anchors with text-shaped aspect ratios), RRPN (rotated region proposals), and SegLink (small oriented segments linked together, a rotated cousin of CTPN).
2019: segmentation and characters
Two ideas arrived at once.
CRAFT (Baek et al., CVPR 2019) predicts a heat map of character centres and a second map of the affinity between characters, then groups them into words of any shape. It trained on synthetic character labels and pseudo-labels for real images, and became the detector in EasyOCR.
PSENet and PAN (Wang et al., 2019) predicted shrunken text kernels and grew them back with progressive scale expansion or pixel aggregation, showing that segmentation plus a good post-processing rule handles arbitrary shapes.
2020: differentiable binarization
DBNet (Liao et al., AAAI 2020) made the threshold that turns a probability map into text regions a learned output, so post-processing collapsed to one contour pass and a polygon expansion. With a ResNet-18 backbone it reported 0.828 F at 62 FPS on MSRA-TD500. The combination of accuracy, speed and simplicity is why PaddleOCR, MMOCR, docTR and OpenCV all ship DB-family detectors, and why the detector on this site runs one. DB++ (TPAMI 2022) added adaptive scale fusion.
2021: speed as the objective
FAST (Chen et al., 2021) cut post-processing to a GPU-parallel dilation of a single-channel kernel map and searched a backbone for text specifically, reporting 0.816 F at 152 FPS for its smallest model on Total-Text.
2022 onward: transformers, spotting, and models that read
DETR-style detectors (TESTR, DeepSolo and their successors) predict text instances as sets of points and read them in the same model, closing the gap between detection and end-to-end spotting. Separately, vision-language models learned to read documents and screenshots directly, outputting text or Markdown without an explicit box step. For scene text, dedicated detectors remain far cheaper and give you coordinates; for documents, the benchmarks page covers how the two families are compared.
What to use in 2026
| Situation | Reach for |
|---|---|
| Boxes and text from photos, on CPU or in a browser | PP-OCR detectors (DB-family): PaddleOCR, or the detector on this site |
| Curved text, research comparisons, custom training | DBNet or DBNet++ in MMOCR; FAST when speed dominates |
| Documents and forms | docTR’s DB models, then a layout model |
| Reading whole pages into structured text | A document VLM, evaluated on OmniDocBench or olmOCR-bench |
| Reproducing a 2016 paper | CTPN: the page links to the authors’ trained model |
Reading the numbers
Every number above is the value stated in the paper’s abstract; results quoted from tables are marked as unverified on the model pages until someone checks the PDF. Datasets differ across the timeline (ICDAR 2013 → ICDAR 2015 → MSRA-TD500 → Total-Text), so the sequence shows what each paper cared about, not a single leaderboard. For the protocols behind the F-measures, see benchmarks.