sberbank-ai
Create app.py
c53b2c3
raw
history blame
2.72 kB
import os
os.system("pip install git+https://github.com/ai-forever/ReadingPipeline.git")
import cv2
import json
import gradio as gr
from huggingface_hub import hf_hub_download
from ocrpipeline.predictor import PipelinePredictor
from ocrpipeline.linefinder import get_structured_text
def get_config_and_download_weights(repo_id, device='cpu'):
# download weights and configs
pipeline_config_path = hf_hub_download(repo_id, "pipeline_config.json")
ocr_model_path = hf_hub_download(repo_id, "ocr/ocr_model.onnx")
kenlm_path = hf_hub_download(repo_id, "ocr/kenlm_corpus.arpa")
ocr_config_path = hf_hub_download(repo_id, "ocr/ocr_config.json")
segm_model_path = hf_hub_download(repo_id, "segm/segm_model.onnx")
segm_config_path = hf_hub_download(repo_id, "segm/segm_config.json")
# change paths to downloaded weights and configs in main pipeline_config
with open(pipeline_config_path, 'r') as f:
pipeline_config = json.load(f)
pipeline_config['main_process']['SegmPrediction']['model_path'] = segm_model_path
pipeline_config['main_process']['SegmPrediction']['config_path'] = segm_config_path
pipeline_config['main_process']['SegmPrediction']['num_threads'] = 2
pipeline_config['main_process']['SegmPrediction']['device'] = device
pipeline_config['main_process']['SegmPrediction']['runtime'] = "OpenVino"
pipeline_config['main_process']['OCRPrediction']['model_path'] = ocr_model_path
pipeline_config['main_process']['OCRPrediction']['lm_path'] = kenlm_path
pipeline_config['main_process']['OCRPrediction']['config_path'] = ocr_config_path
pipeline_config['main_process']['OCRPrediction']['num_threads'] = 2
pipeline_config['main_process']['OCRPrediction']['device'] = device
pipeline_config['main_process']['OCRPrediction']['runtime'] = "OpenVino"
# save pipeline_config
with open(pipeline_config_path, 'w') as f:
json.dump(pipeline_config, f)
return pipeline_config_path
def predict(image_path):
image = cv2.imread(image_path)
rotated_image, pred_data = PREDICTOR(image)
structured_text = get_structured_text(pred_data, ['shrinked_text'])
result_text = [' '.join(line_text) for page_text in structured_text
for line_text in page_text]
return '\n'.join(result_text)
PIPELINE_CONFIG_PATH = get_config_and_download_weights("sberbank-ai/ReadingPipeline-notebooks")
PREDICTOR = PipelinePredictor(pipeline_config_path=PIPELINE_CONFIG_PATH)
gr.Interface(
predict,
inputs=gr.Image(label="Upload an image of handwritten school notebook", type="filepath"),
outputs=gr.Textbox(label="Text on the image"),
title="School notebook recognition",
).launch()