parseq

Why we included this project

PARSeq is a scene text recognition model from ECCV 2022 that reads words out of cropped images of signs, receipts, labels, and other real-world scenes. The core idea is a single Transformer that changes how it decodes text just by swapping its attention mask, so one model covers context-free recognition, context-aware decoding, and iterative refinement without a separate language model. That keeps the compute footprint smaller than comparable approaches. The repo ships pretrained weights, training and evaluation scripts, and a live demo, and the model has been adopted into both PaddleOCR and docTR, which makes it a safe implementation to borrow from or integrate. Researchers comparing STR architectures will also find it a clean, reproducible baseline against ABINet, TRBA, and ViTSTR.