EasyOCR in ONNX
EasyOCR's official models by Jaided AI (Apache-2.0), exported to ONNX unchanged so they run in the browser with ONNX Runtime Web. Used by OCR Arena.
craft.onnx: the CRAFT text detector. Inputimage(1, 3, H, W): RGB, (x - mean) / std with ImageNet mean and std times 255. Outputscores(1, H/2, W/2, 2): text score and link score.rec_<script>.onnx: the recognizer for one script (latin, english, arabic, devanagari, cyrillic, zh_sim, japanese, korean). Inputimage(1, 1, 64, W): greyscale, x / 255 * 2 - 1. Outputlogits(1, T, classes), decoded with greedy CTC.rec_<script>.json: its character list (index 0 is the CTC blank) and the EasyOCR languages it serves.
Exported with torch.onnx.export from easyocr 1.7.2. All credit for the models to Jaided AI.