is-image-document-ai 0.0.2

A simple AI package for classifying images as documents or general images.

is_image_document_ai

A lightweight pipeline for classifying whether an image is a document (e.g., scanned text, forms, PDFs) or a generic image (photographs, art, etc.). This repository provides two model options:

1. MobileNetV2 (pretrained on ImageNet, then fine-tuned)
2. TinyCNN (a small custom CNN)

Both achieve ~99% accuracy on our dataset, with MobileNetV2 (recommended & default) slightly higher (~99.8%) and TinyCNN slightly faster (~99.2% accuracy).

Quickstart:

Install with:

pip install is-image-document-ai==0.0.2

Classify a folder of images of documents or generic images...

from is_image_document_ai import infer_folder

result = infer_folder("testfolder")
print(result)

Expected result:

paper.jpg => document
image.png => image
['document', 'image']

Table of Contents

• Overview
• Data
• Installation
• Inference
  • Single Image
  • Folder of Images
  • Test-Set Evaluation (DataLoader)
• Training
  • MobileNetV2 Model
  • TinyCNN Model
• Usage Examples
• Results
• Contributing
• License

---

Overview

• Purpose: Quickly determine if an image is a “document” vs. a “regular image.” Useful for filtering scanned PDFs, forms, or text-based images from photo-based datasets.
• Models:
  • MobileNetV2 for high accuracy (99.8%).
  • TinyCNN for a more lightweight approach (~99.2% accuracy, fewer parameters).

---

Data

• We downloaded 10,000 document images from HuggingFaceM4/DocumentVQA and 10,000 generic images from jackyhate/text-to-image-2M, resulting in:

  images/
   ├─ document/
   │   ├─ 0.jpg
   │   ├─ 1.jpg
   │   ...
   └─ image/
       ├─ 0.jpg
       ├─ 1.jpg
       ...
  

• We perform a 3-way random split: 80% training, 10% validation, 10% test (i.e., train_ratio=0.8, val_ratio=0.1, test_ratio=0.1).
• Each image is resized to 224×224 before feeding into the models.

Installation

Using the PyPi package:

pip install is-image-document-ai==0.0.1

Using normal python environment:

1. Clone this repository:

   git clone https://github.com/logophoman/is_image_document_ai.git
   cd is_image_document_ai
   

2. Install requirements:

   pip install -r requirements.txt
   

   This installs torch, torchvision, requests, and datasets.

3. (Optional) Download data if you haven’t:

   python download_data.py
   

   By default, it streams from the Hugging Face datasets to get 10k document images and 10k image images, saved into images/.

   Note: You will need ~3.8GB of free storage for the image data. Training requires ~4GB of VRAM for mobilenet and ~800MB VRAM for TinyCNN

---

Inference

Using the PyPi package:

Input Folder:

from is_image_document_ai import load_tinycnn, infer_folder

model = load_tinycnn()
result = infer_folder("testfolder", model)
print(result)

Expected result:

paper.jpg => document
image.png => image
['document', 'image']

Input File:

from is_image_document_ai import load_mobilenet, infer_single_image

model = load_mobilenet()
result = infer_single_image("paper.jpg", model)
print(result)

Expected result:

paper.jpg => document
document

Or for the lazy ones (defaults to mobilenet):

from is_image_document_ai import infer_folder

result = infer_folder("testfolder")
print(result)

Manual using the scripts yourself:

We provide multiple ways to run inference. You can:

1. Use the test.py (for MobileNetV2) or tinytest.py (for TinyCNN) scripts for quick checks.
2. Use the inference.py script, which supports command-line arguments to load either model and do single, folder, or testset evaluation.

Single Image

python inference.py --model-type mobilenet \
                    --weights mobilenet_document_vs_image.pth \
                    --mode single \
                    --input path/to/your_image.jpg

Folder of Images

python inference.py --model-type tinycnn \
                    --weights tinycnn_document_vs_image.pth \
                    --mode folder \
                    --input test_images/

Test-Set Evaluation (DataLoader)

python inference.py --model-type mobilenet \
                    --weights mobilenet_document_vs_image.pth \
                    --mode testset \
                    --root-dir images \
                    --batch-size 32

---

Training

MobileNetV2 Model

• File: train.py
• Command:

  python train.py
  

• This script:
  1. Loads data from images/ folder using data_loader.py.
  2. Initializes a MobileNetV2 (pretrained on ImageNet).
  3. Fine-tunes the final layer for 2 classes: document vs. image.
  4. Trains until ~99% accuracy or until epochs finish.
  5. Saves the weights to mobilenet_document_vs_image.pth.
  6. Evaluates on the test set (the 10% split).

TinyCNN Model

• File: tinytrain.py
• Command:

  python tinytrain.py
  

• This script:
  1. Loads the same images/ dataset with an 80/10/10 split.
  2. Initializes our custom TinyCNN architecture (model.py).
  3. Trains for up to 5 epochs (or early stops at ~99% val accuracy).
  4. Saves weights to tinycnn_document_vs_image.pth.
  5. Prints final test accuracy.

---

Results

After training on 20k images (10k document + 10k generic), both models achieve high accuracy:

• MobileNetV2: ~99.8% accuracy on the test set.
• TinyCNN: ~99.2% accuracy on the test set.

The TinyCNN has fewer parameters and uses less VRAM for the model itself, but PyTorch’s baseline overhead may still be a few hundred MB. MobileNetV2 is slightly more accurate and can converge faster due to ImageNet pretraining.

---

Contributing

Contributions are welcome! To get started:

1. Fork the repo and clone your fork.
2. Create a new branch for your feature or bug fix.
3. Make changes, add tests, and ensure everything passes.
4. Open a Pull Request describing your changes.

Possible improvements:

• Quantization or Half-Precision (FP16) to reduce model size / VRAM usage.
• Additional data augmentations to handle more varied document layouts.
• Implementation of other lightweight architectures (e.g. ShuffleNet, EfficientNet-Lite).

---

License

This project is licensed under the MIT License. You are free to use, modify, and distribute this code as permitted by the license.

---

Enjoy classifying your images! If you have any questions or issues, feel free to open an issue or pull request.