![2.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/yUKVKSX2E18k0h3YwCx1h.png)

# **Imgscope-OCR-2B-0527**

> The **Imgscope-OCR-2B-0527** model is a fine-tuned version of *Qwen2-VL-2B-Instruct*, specifically optimized for *messy handwriting recognition*, *document OCR*, *realistic handwritten OCR*, and *math problem solving with LaTeX formatting*. This model is trained on custom datasets for document and handwriting OCR tasks and integrates a conversational approach with strong visual and textual understanding for multi-modal applications.

> [!warning]
Colab Demo : https://huggingface.co/prithivMLmods/Imgscope-OCR-2B-0527/blob/main/Imgscope%20OCR%202B%200527%20Demo/Imgscope-OCR-2B-0527.ipynb

---

### Key Enhancements

* **SoTA Understanding of Images of Various Resolution & Ratio**
  Imgscope-OCR-2B-0527 achieves state-of-the-art performance on visual understanding benchmarks such as MathVista, DocVQA, RealWorldQA, and MTVQA.

* **Enhanced Handwriting OCR**
  Specifically optimized for recognizing and interpreting **realistic and messy handwriting** with high accuracy. Ideal for digitizing handwritten documents and notes.

* **Document OCR Fine-Tuning**
  Fine-tuned with curated and realistic **document OCR datasets**, enabling accurate extraction of text from various structured and unstructured layouts.

* **Understanding Videos of 20+ Minutes**
  Capable of processing long videos for **video-based question answering**, **transcription**, and **content generation**.

* **Device Control Agent**
  Supports decision-making and control capabilities for integration with **mobile devices**, **robots**, and **automation systems** using visual-textual commands.

* **Multilingual OCR Support**
  In addition to English and Chinese, the model supports **OCR in multiple languages** including European languages, Japanese, Korean, Arabic, and Vietnamese.

---

### Demo Video Inference

https://github.com/user-attachments/assets/3ca9ef10-8a71-4cd1-8be1-951a9f6d5a00

```

The video starts with a group of people gathered around a table filled with snacks and drinks, indicating a casual social gathering. One person is seen holding a can of Pringles, suggesting that the snack is being enjoyed by the attendees.

As the scene progresses, the focus shifts to a man who is seen pouring a drink from a can into a glass. This action implies that the drink is being served or shared among the group.

The next scene shows a different setting where a man is walking down a hallway while holding a can of Pringles. This could indicate that he is on his way to join the group or has just arrived at the location.

The following scene takes place in a diner where two people are seated at a booth. The man is seen holding a can of Pringles, which suggests that they might be enjoying a meal together.

The video then transitions to a wedding ceremony where a man is feeding a woman a piece of cake using a can of Pringles. This unusual gesture adds a humorous element to the otherwise traditional event.

Next, the scene changes to a bedroom where a man is seen feeding a woman a piece of cake using a can of Pringles. This scene further emphasizes the playful nature of the video.

The video then shifts to an office setting where a man is seen working at a desk. The presence of a can of Pringles on the desk suggests that it might be part of his workspace or a snack during work hours.

Finally, the video ends with a scene of a funeral where a woman is seen crying over a casket. The presence of a can of Pringles on the casket adds an unexpected and humorous touch to the solemn occasion.

Throughout the video, the recurring theme of Pringles is evident, with various scenes featuring the snack as a central element. The video concludes with the text "GET STUCK IN," encouraging viewers to enjoy the snack and engage with the content.

```

### How to Use

```python
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

# Load the model
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/Imgscope-OCR-2B-0527",  # replace with updated model ID if available
    torch_dtype="auto",
    device_map="auto"
)

# Optional: Flash Attention for performance optimization
# model = Qwen2VLForConditionalGeneration.from_pretrained(
#     "prithivMLmods/Imgscope-OCR-2B-0527",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# Load processor
processor = AutoProcessor.from_pretrained("prithivMLmods/Imgscope-OCR-2B-0527")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Recognize the handwriting in this image."},
        ],
    }
]

# Prepare input
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")

# Generate output
generated_ids = model.generate(**inputs, max_new_tokens=128)
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(output_text)
```

---

### Demo Inference

![Screenshot 2025-05-27 at 03-40-34 Gradio.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/9KiRkOGPB8cLl6VHwh2UD.png)
![Screenshot 2025-05-27 at 03-40-56 (anonymous) - output_e0fbfa20-686e-4bce-b2e8-25991be5a5a0.pdf.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/VOHQIrT7hCs5afGMRROvD.png)

---

### Buffering Output (Streaming)

```python
buffer = ""
for new_text in streamer:
    buffer += new_text
    buffer = buffer.replace("<|im_end|>", "")
    yield buffer
```

---

### Key Features

1. **Realistic Messy Handwriting OCR**

   * Fine-tuned for **complex and hard-to-read handwritten inputs** using real-world handwriting datasets.

2. **Document OCR and Layout Understanding**

   * Accurately extracts text from structured documents, including scanned pages, forms, and academic papers.

3. **Image and Text Multi-modal Reasoning**

   * Combines **vision-language capabilities** for tasks like captioning, answering image-based queries, and understanding image+text prompts.

4. **Math Problem Solving and LaTeX Rendering**

   * Converts mathematical expressions and problem-solving steps into **LaTeX** format.

5. **Multi-turn Conversations**

   * Supports **dialogue-based reasoning**, retaining context for follow-up questions.

6. **Video + Image + Text-to-Text Generation**

   * Accepts inputs from videos, images, or combined media with text, and generates relevant output accordingly.

---

## **Intended Use**

**Imgscope-OCR-2B-0527** is intended for:

* Handwritten and printed document digitization
* OCR pipelines for educational institutions and businesses
* Academic and scientific content parsing, especially math-heavy documents
* Assistive tools for visually impaired users
* Robotic and mobile automation agents interpreting screen or camera data
* Multilingual OCR processing for document translation or archiving