| | --- |
| | license: apache-2.0 |
| | datasets: |
| | - AadityaJain/Fromula_text_classification |
| | language: |
| | - en |
| | base_model: |
| | - google/siglip2-base-patch16-224 |
| | pipeline_tag: image-classification |
| | library_name: transformers |
| | tags: |
| | - Formula-Text-Detection |
| | - SigLIP2 |
| | - Image-Classification |
| | --- |
| | |
| |  |
| |
|
| | # **Formula-Text-Detection** |
| |
|
| | > **Formula-Text-Detection** is a vision-language encoder model fine-tuned from **google/siglip2-base-patch16-224** for **binary image classification**. It is built using the **SiglipForImageClassification** architecture to distinguish between **mathematical formulas** and **natural text** in document or image regions. |
| |
|
| | > [!Note] |
| | > Note: This model works best with plain text or formulas using the same font style |
| |
|
| |
|
| | ```py |
| | Classification Report: |
| | precision recall f1-score support |
| | |
| | formula 0.9983 1.0000 0.9991 6375 |
| | text 1.0000 0.9980 0.9990 5457 |
| | |
| | accuracy 0.9991 11832 |
| | macro avg 0.9991 0.9990 0.9991 11832 |
| | weighted avg 0.9991 0.9991 0.9991 11832 |
| | ``` |
| |
|
| |  |
| |
|
| | --- |
| |
|
| | > [!note] |
| | *SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features* https://arxiv.org/pdf/2502.14786 |
| |
|
| | --- |
| |
|
| | ## **Label Space: 2 Classes** |
| |
|
| | The model classifies each input image into one of the following categories: |
| |
|
| | ``` |
| | Class 0: "formula" |
| | Class 1: "text" |
| | ``` |
| |
|
| | --- |
| |
|
| | ## **Install Dependencies** |
| |
|
| | ```bash |
| | pip install -q transformers torch pillow gradio |
| | ``` |
| |
|
| | --- |
| |
|
| | ## **Inference Code** |
| |
|
| | ```python |
| | import gradio as gr |
| | from transformers import AutoImageProcessor, SiglipForImageClassification |
| | from PIL import Image |
| | import torch |
| | |
| | # Load model and processor |
| | model_name = "prithivMLmods/Formula-Text-Detection" # Replace with your model path if different |
| | model = SiglipForImageClassification.from_pretrained(model_name) |
| | processor = AutoImageProcessor.from_pretrained(model_name) |
| | |
| | # Label mapping |
| | id2label = { |
| | "0": "formula", |
| | "1": "text" |
| | } |
| | |
| | def classify_formula_or_text(image): |
| | image = Image.fromarray(image).convert("RGB") |
| | inputs = processor(images=image, return_tensors="pt") |
| | |
| | with torch.no_grad(): |
| | outputs = model(**inputs) |
| | logits = outputs.logits |
| | probs = torch.nn.functional.softmax(logits, dim=1).squeeze().tolist() |
| | |
| | prediction = { |
| | id2label[str(i)]: round(probs[i], 3) for i in range(len(probs)) |
| | } |
| | |
| | return prediction |
| | |
| | # Gradio Interface |
| | iface = gr.Interface( |
| | fn=classify_formula_or_text, |
| | inputs=gr.Image(type="numpy"), |
| | outputs=gr.Label(num_top_classes=2, label="Formula or Text"), |
| | title="Formula-Text-Detection", |
| | description="Upload an image region to classify whether it contains a mathematical formula or natural text." |
| | ) |
| | |
| | if __name__ == "__main__": |
| | iface.launch() |
| | ``` |
| | ## **Demo Inference** |
| |
|
| | > [!Important] |
| | > Text |
| |
|
| |
|
| |  |
| |  |
| |  |
| |
|
| | > [!Important] |
| | > Formula |
| |
|
| |  |
| |  |
| |  |
| |
|
| | --- |
| |
|
| | ## **Intended Use** |
| |
|
| | **Formula-Text-Detection** can be used in: |
| |
|
| | - **OCR Preprocessing** – Improve document OCR accuracy by separating formulas from text. |
| | - **Scientific Document Analysis** – Automatically detect mathematical content. |
| | - **Educational Platforms** – Classify and annotate scanned materials. |
| | - **Layout Understanding** – Help AI systems interpret mixed-content documents. |
