vlad-m-dev/mobilenetv2_doc_photo_quant

MobileNetV2 — ONNX, Quantized

🔥 Lightweight mobile model for image classification into two categories:

• document (scans, receipts, papers, invoices)
• photo (regular phone photos: scenes, people, nature, etc.)

🟢 Overview

• Designed for mobile devices (phones and tablets, Android/iOS), perfect for real-time on-device inference!
• Architecture: MobileNetV2
• Format: ONNX (both float32 and quantized int8 versions included)
• Trained on balanced, real-world open-source datasets for both documents and photos.
• Ideal for tasks like:
  • Document detection in gallery/camera rolls
  • Screenshot, receipt, photo, and PDF preview classification
  • Image sorting for privacy-first offline AI assistants

🏷️ Model Classes

• 0 — document
• 1 — photo

⚡️ Versions

• mobilenetv2_doc_photo.onnx — Standard float32 for maximum accuracy (best for ARM/CPU)
• mobilenetv2_doc_photo_quant.onnx — Quantized int8 for even faster inference and smaller file size (best for low-power or edge devices)

🚀 Why this model?

• Ultra-small size (~10-15MB), real-time inference (<100ms) on most phones
• Runs 100% offline (privacy, no cloud required)
• Easy integration with any framework, including React Native (onnxruntime-react-native), Android (ONNX Runtime), and iOS.

🗃️ Datasets

• Photos: alfredplpl/Japanese-photos
• Documents: 3sara/colpali_italian_documents

🤖 Author

@vlad-m-dev Built for edge-ai/phone/tablet offline image classification: document vs photo Telegram: https://t.me/dwight_schrute_engineer

🛠️ Usage Example

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("mobilenetv2_doc_photo_quant.onnx")
img = np.random.randn(1, 3, 224, 224).astype(np.float32)  # Replace with your image preprocessing!
output = session.run(None, {"input": img})
pred_class = np.argmax(output[0])
print(pred_class)  # 0 = document, 1 = photo```