syntheticbot/Qwen-VL-7B-ocr

Introduction

syntheticbot/Qwen-VL-7B-ocr is a fine-tuned model for Optical Character Recognition (OCR) tasks, derived from the base model Qwen/Qwen2.5-VL-7B-Instruct. This model is engineered for high accuracy in extracting text from images, including documents and scenes containing text.

Key Enhancements for OCR:

Enhanced Text Recognition Accuracy: Superior accuracy across diverse text fonts, styles, sizes, and orientations.
Robustness to Document Variations: Specifically trained to manage document complexities like varied layouts, noise, and distortions.
Structured Output Generation: Enables structured output formats (JSON, CSV) for recognized text and layout in document images such as invoices and tables.
Text Localization: Provides accurate localization of text regions and bounding boxes for text elements within images.
Improved Handling of Text in Visuals: Maintains proficiency in analyzing charts and graphics, with enhanced recognition of embedded text.

Model Architecture Updates:

Dynamic Resolution and Frame Rate Training for Video Understanding:
Streamlined and Efficient Vision Encoder

This repository provides the instruction-tuned and OCR-optimized 7B Qwen-VL-7B-ocr model. For comprehensive details about the foundational model architecture, please refer to the Qwen/Qwen2.5-VL-7B-Instruct repository, as well as the Blog and GitHub pages for Qwen2.5-VL.

Requirements

For optimal performance and access to OCR-specific features, it is recommended to build from source:

pip install git+https://github.com/huggingface/transformers accelerate

Quickstart

The following examples illustrate the use of syntheticbot/Qwen-VL-7B-ocr with 🤗 Transformers and qwen_vl_utils for OCR applications.

pip install git+https://github.com/huggingface/transformers accelerate

Install the toolkit for streamlined visual input processing:

pip install qwen-vl-utils[decord]==0.0.8

Using 🤗 Transformers for OCR

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "syntheticbot/Qwen-VL-7B-ocr",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained("syntheticbot/Qwen-VL-7B-ocr")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "path/to/your/document_image.jpg",
            },
            {"type": "text", "text": "Extract the text from this image."},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda" if torch.cuda.is_available() else "cpu")

generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Extracted Text:", output_text[0])

Example for Structured Output (JSON for Table Extraction)

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
import json

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "syntheticbot/Qwen-VL-7B-ocr",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("syntheticbot/Qwen-VL-7B-ocr")


messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "path/to/your/table_image.jpg",
            },
            {"type": "text", "text": "Extract the table from this image and output it as JSON."},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda" if torch.cuda.is_available() else "cpu")

generated_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Extracted Table (JSON):\n", output_text[0])

try:
    json_output = json.loads(output_text[0])
    print("\nParsed JSON Output:\n", json.dumps(json_output, indent=2))
except json.JSONDecodeError:
    print("\nCould not parse output as JSON. Output is plain text.")

Batch inference for OCR

messages1 = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "path/to/image1.jpg"},
            {"type": "text", "text": "Extract text from this image."},
        ],
    }
]
messages2 = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "path/to/image2.jpg"},
            {"type": "text", "text": "Read the text in this document."},
        ],
    }
]
messages = [messages1, messages2]

texts = [
    processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True)
    for msg in messages
]
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=texts,
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda" if torch.cuda.is_available() else "cpu")

generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_texts = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Extracted Texts (Batch):\n", output_texts)

🤖 ModelScope

For users in mainland China, ModelScope is recommended. Use snapshot_download for checkpoint management. Adapt model names to syntheticbot/Qwen-VL-7B-ocr in ModelScope implementations.

More Usage Tips for OCR

Input images support local files, URLs, and base64 encoding.

messages = [    {        "role": "user",        "content": [            {"type": "image", "image": "http://path/to/your/document_image.jpg"},            {"type": "text", "text": "Extract the text from this image URL."},        ],    }]

Image Resolution for OCR Accuracy

Higher resolution images typically improve OCR accuracy, especially for small text. Adjust resolution using min_pixels, max_pixels, resized_height, and resized_width parameters with the processor.

min_pixels = 512 * 28 * 28
max_pixels = 2048 * 28 * 28
processor = AutoProcessor.from_pretrained(
    "syntheticbot/Qwen-VL-7B-ocr",
    min_pixels=min_pixels, max_pixels=max_pixels
)

Control resizing dimensions directly:

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "file:///path/to/your/document_image.jpg",
                "resized_height": 600,
                "resized_width": 800,
            },
            {"type": "text", "text": "Extract the text."},
        ],
    }
]

Citation from base model

@misc{qwen2.5-VL,
    title = {Qwen2.5-VL},
    url = {https://qwenlm.github.io/blog/qwen2.5-vl/},
    author = {Qwen Team},
    month = {January},
    year = {2025}
}

@article{Qwen2VL,
  title={Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution},
  author={Wang, Peng and Bai, Shuai and Tan, Sinan and Wang, Shijie and Fan, Zhihao and Bai, Jinze and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Fan, Yang and Dang, Kai and Du, Mengfei and Ren, Xuancheng and Men, Rui and Liu, Dayiheng and Zhou, Chang and Zhou, Jingren and Lin, Junyang},
  journal={arXiv preprint arXiv:2409.12191},
  year={2024}
}

@article{Qwen-VL,
  title={Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond},
  author={Bai, Jinze and Bai, Shuai and Yang, Shusheng and Wang, Shijie and Tan, Sinan and Wang, Peng and Lin, Junyang and Zhou, Chang and Zhou, Jingren},
  journal={arXiv preprint arXiv:2308.12966},
  year={2023}
}