Meiki.txt.recognition.v0

--- license: lgpl-3.0 tags: - onnx - text-recognition - textline_recognition - computer-vision - japanese - ocr - OCR library_name: tensorrt language: - en - ja pipeline_tag: image-to-text --- --- **Official GitHub Repository:** [meikiocr](https://github.com/rtr46/meikiocr) This model is a core component of the `meikiocr` pipeline. For the full implementation, command-line script, and documentation, please see the official GitHub repository. --- # meiki.text.recognition.v0 pareto-optimal text recognition model. trained on japanese video games. meiki.text.recognition achieves state-of-the-art text recognition accuracy as well as latency by redefining "text recognition" as "character detection". the model is a fine-tune of https://github.com/Peterande/D-FINE object detecor combined with a mobilenetv4 CNN backbone. to my knowledge there is no existing, open weight text recognition model with a better accuracy/latency tradeoff for japanese text recognition. ## intended use and constraints - it is specifically trained on japanese video games, therefore performance may vary outside of this use case - input needs to be resized and padded to be 960x32px - outputs detection results as characters + bbox + confidence. check the inference.py script for a suggested post processing algorithms - can detect up to 48 characters - only works on horizontal text ## benchmarks | cpu | gpu | |:---:|:---:| |  |  | ## how to use please refer to this demo inference script: https://huggingface.co/rtr46/meiki.txt.recognition.v0/blob/main/inference.py ## examples  ``` その一つの実情が、第一層の一画、空気には黴臭さと変に饐えた甘 ``` ```json { "text":"その一つの実情が、第一層の一画、空気には黴臭さと変に饐えた甘", "chars":[ {"char":"そ","bbox":[2,0,33,32]}, {"char":"の","bbox":[33,0,65,32]}, {"char":"一","bbox":[65,0,96,32]}, {"char":"つ","bbox":[97,0,128,32]}, {"char":"の","bbox":[129,0,160,32]}, {"char":"実","bbox":[161,0,193,32]}, {"char":"情","bbox":[192,0,224,32]}, {"char":"が","bbox":[225,0,256,32]}, {"char":"、","bbox":[258,0,288,32]}, {"char":"第",...