Why return_word_box didn't works?

[mkrzywda]

问题描述 / Problem Description

I have a Python script with RapidOCR set to return_word_box: True, but it doesn't work as it still generates bounding boxes per line instead of per word. This is critical for me because RapidOCR is a key component of my project, and if it doesn't work correctly with word-level bounding boxes, I'll have to look for another solution.

运行环境 / Runtime Environment

WSL + OpenVINO

复现代码 / Reproduction Code

from rapidocr import RapidOCR
import time

# Initialize RapidOCR with parameters
config_path = "default_rapidocr.yaml"
engine = RapidOCR(config_path=config_path)

print(engine.config)
img_path = "test.png"

# Measure start time
start_time = time.time()

# Run OCR
result = engine(img_path, use_det=True, use_cls=False, use_rec=True)

# Measure end time and calculate duration
end_time = time.time()
duration_ms = (end_time - start_time) * 1000  # Convert to milliseconds

print("OCR Result:", result)
print(f"Processing time: {duration_ms:.2f} ms")

# Visualize result
result.vis('vis_result-openvino-rapid-ocr.jpg')

default_config.yaml

Global:
    lang_det: "en_mobile" # ch_server
    lang_rec: "en_mobile"
    text_score: 0.5

    use_det: true
    use_cls: false
    use_rec: true

    min_height: 30
    width_height_ratio: 8
    max_side_len: 2000
    min_side_len: 30

    return_word_box: true

    with_onnx: false
    with_openvino: true
    with_paddle: false
    with_torch: false

    font_path: null

EngineConfig:
    onnxruntime:
        intra_op_num_threads: -1
        inter_op_num_threads: -1
        use_cuda: true
        use_dml: false

    openvino:
        inference_num_threads: -1

    paddlepaddle:
        cpu_math_library_num_threads: -1
        use_cuda: false
        gpu_id: 0
        gpu_mem: 500

    torch:
        use_cuda: false
        gpu_id: 0

Det:
    model_path: null
    model_dir: null

    limit_side_len: 736
    limit_type: min
    std: [ 0.5, 0.5, 0.5 ]
    mean: [ 0.5, 0.5, 0.5 ]

    thresh: 0.3
    box_thresh: 0.5
    max_candidates: 1000
    unclip_ratio: 1.6
    use_dilation: true
    score_mode: fast

Cls:
    model_path: null
    model_dir: null

    cls_image_shape: [3, 48, 192]
    cls_batch_num: 6
    cls_thresh: 0.9
    label_list: ['0', '180']

Rec:
    model_dir: null
    dict_url: null
    rec_keys_path: null
    rec_img_shape: [3, 48, 320]
    rec_batch_num: 6

可能解决方案 / Possible solutions

• Handling Bouding Boxes per word instead of whole lline.

[wukonggeo]

@mkrzywda
You can manually pass parameters, such as.

from rapidocr_onnxruntime import RapidOCR

img_file = "test.png"
engine = RapidOCR()
result, elapse = engine(img_file, return_word_box=True)

[mkrzywda]

@wukonggeo

I modified the code to:

result = engine(img_path, use_det=True, use_cls=False, use_rec=True, return_word_box=True)

However, even though I set return_word_box=True in the config, the screenshot shows bounding boxes around individual letters instead of words. How can I fix this to get bounding boxes per word?

Also, why should I use rapidocr_onnxruntime when it doesn't have the .vis method?

[SWHL]

The return_word_box parameter means to return the coordinates of a single character. In Chinese, it is a single character, while in English it is a single letter.

If you want to get the word bbox of English, you could use the v1.4.0 of rapidocr_onnxruntime (See link for details)