app.py

# -*- encoding: utf-8 -*-
import json
import time
from pathlib import Path

import cv2
import gradio as gr
from rapidocr_onnxruntime import RapidOCR

from utils import visualize

font_dict = {
    'ch': 'FZYTK.TTF',
    'japan': 'japan.ttc',
    'korean': 'korean.ttf',
    'en': 'FZYTK.TTF',
}


def inference(img_path, box_thresh=0.5, unclip_ratio=1.6, text_score=0.5,
              text_det=None, text_rec=None):
    out_log_list = []

    det_model_path = str(Path('models') / 'text_det' / text_det)
    rec_model_path = str(Path('models') / 'text_rec' / text_rec)
    if 'v2' in rec_model_path:
        rec_image_shape = [3, 32, 320]
    else:
        rec_image_shape = [3, 48, 320]

    out_log_list.append('Init Model')
    s = time.time()
    rapid_ocr = RapidOCR(det_model_path=det_model_path,
                         rec_model_path=rec_model_path,
                         rec_img_shape=rec_image_shape)
    elapse = time.time() - s

    if 'ch' in rec_model_path or 'en' in rec_model_path:
        lan_name = 'ch'
    elif 'japan' in rec_model_path:
        lan_name = 'japan'
    elif 'korean' in rec_model_path:
        lan_name = 'korean'
    else:
        lan_name = 'ch'

    out_log_list.append(f'Init Model cost: {elapse:.5f}')
    out_log_list.extend([f'det_model: {det_model_path}',
                         f'rec_model: {rec_model_path}',
                         f'rec_image_shape: {rec_image_shape}'])

    img = cv2.imread(img_path)
    ocr_result, infer_elapse = rapid_ocr(img, box_thresh=box_thresh,
                                         unclip_ratio=unclip_ratio,
                                         text_score=text_score)
    det_cost, cls_cost, rec_cost = infer_elapse
    out_log_list.extend([f'det cost: {det_cost:.5f}',
                         f'cls cost: {cls_cost:.5f}',
                         f'rec cost: {rec_cost:.5f}'])
    out_log = '\n'.join([str(v) for v in out_log_list])
    if not ocr_result:
        return img_path, '未识别到有效文本', out_log

    dt_boxes, rec_res, scores = list(zip(*ocr_result))
    font_path = Path('fonts') / font_dict.get(lan_name)
    img_save_path = visualize(img_path, dt_boxes, rec_res, scores,
                              font_path=str(font_path))
    out_dict = {str(i): {'rec_txt': rec, 'score': score} 
                for i, (rec, score) in enumerate(zip(rec_res, scores))}
    return img_save_path, out_dict, out_log


if __name__ == '__main__':
    examples = [['images/1.jpg'],
                ['images/ch_en_num.jpg'],
                ['images/air_ticket.jpg'],
                ['images/car_plate.jpeg'],
                ['images/idcard.jpg'],
                ['images/train_ticket.jpeg'],
                ['images/japan_2.jpg'],
                ['images/korean_1.jpg']]

    with gr.Blocks(title='RapidOCR') as demo:
        gr.Markdown("""
            <h1><center><a href="https://github.com/RapidAI/RapidOCR" target="_blank">Rapid⚡OCR</a></center></h1>

            ### Docs: [Docs](https://rapidocr.rtfd.io/)
            ### 运行环境：
            Python: 3.8 | onnxruntime: 1.14.1 | rapidocr_onnxruntime: 1.2.5""")
        gr.Markdown(
            '''**[超参数调节](https://github.com/RapidAI/RapidOCR/tree/main/python#configyaml%E4%B8%AD%E5%B8%B8%E7%94%A8%E5%8F%82%E6%95%B0%E4%BB%8B%E7%BB%8D)**
            - **box_thresh**: 检测到的框是文本的概率，值越大，框中是文本的概率就越大。存在漏检时，调低该值。取值范围：[0, 1.0]
            - **unclip_ratio**: 控制文本检测框的大小，值越大，检测框整体越大。在出现框截断文字的情况，调大该值。取值范围：[1.5, 2.0]
            - **text_score**: 文本识别结果是正确的置信度，值越大，显示出的识别结果更准确。存在漏检时，调低该值。取值范围：[0, 1.0]
            ''')
        with gr.Row():
            box_thresh = gr.Slider(minimum=0, maximum=1.0, value=0.5,
                                   label='box_thresh', step=0.1,
                                   interactive=True,
                                   info='[0, 1.0]')
            unclip_ratio = gr.Slider(minimum=1.5, maximum=2.0, value=1.6,
                                     label='unclip_ratio', step=0.1,
                                     interactive=True,
                                     info='[1.5, 2.0]')
            text_score = gr.Slider(minimum=0, maximum=1.0, value=0.5,
                                   label='text_score', step=0.1,
                                   interactive=True,
                                   info='[0, 1.0]')

        gr.Markdown('**[模型选择](https://github.com/RapidAI/RapidOCR/blob/main/docs/models.md)** (模型转换→[PaddleOCRModelConverter](https://github.com/RapidAI/PaddleOCRModelConverter))')
        with gr.Row():
            text_det = gr.Dropdown(['ch_PP-OCRv3_det_infer.onnx',
                                    'ch_PP-OCRv2_det_infer.onnx',
                                    'ch_ppocr_server_v2.0_det_infer.onnx'],
                                   label='选择文本检测模型',
                                   value='ch_PP-OCRv3_det_infer.onnx',
                                   interactive=True)
            rec_model_list = [v.name for v in Path('models/text_rec').iterdir()]
            text_rec = gr.Dropdown(rec_model_list,
                                   label='选择文本识别模型(包括中英文和多语言，欢迎提交更多模型)',
                                   value='ch_PP-OCRv3_rec_infer.onnx',
                                   interactive=True)

        with gr.Row():
            input_img = gr.Image(type='filepath', label='Input')
            out_img = gr.Image(type='filepath', label='Output')
        out_json = gr.JSON(label='Rec Res')
        out_log = gr.outputs.Textbox(type='text', label='Run Log')
        button = gr.Button('Submit')
        button.click(fn=inference,
                     inputs=[input_img, box_thresh, unclip_ratio, text_score,
                             text_det, text_rec],
                     outputs=[out_img, out_json, out_log])
        gr.Examples(examples=examples,
                    inputs=[input_img, box_thresh, unclip_ratio, text_score,
                            text_det, text_rec],
                    outputs=[out_img, out_json, out_log], fn=inference)
    demo.launch(debug=True, enable_queue=True)