jadechoghari / textnet-tiny

TextNet-T/S/B: Efficient Text Detection Models

Overview

TextNet is a lightweight and efficient architecture designed specifically for text detection, offering superior performance compared to traditional models like MobileNetV3. With variants TextNet-T , TextNet-S , and TextNet-B (6.8M, 8.0M, and 8.9M parameters respectively), it achieves an excellent balance between accuracy and inference speed.

Performance

TextNet achieves state-of-the-art results in text detection, outperforming hand-crafted models in both accuracy and speed. Its architecture is highly efficient, making it ideal for GPU-based applications.

How to use

Transformers

pip install transformers

import torch
import requests
from PIL import Image
from transformers import AutoImageProcessor, AutoBackbone

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

processor = AutoImageProcessor.from_pretrained("jadechoghari/textnet-tiny")
model = AutoBackbone.from_pretrained("jadechoghari/textnet-base")

inputs = processor(image, return_tensors="pt")
with torch.no_grad():
  outputs = model(**inputs)

Training

We first compare TextNet with representative hand-crafted backbones, such as ResNets and VGG16. For a fair comparison, all models are first pre-trained on IC17-MLT [52] and then finetuned on Total-Text. The proposed TextNet models achieve a better trade-off between accuracy and inference speed than previous hand-crafted models by a significant margin. In addition, notably, our TextNet-T, -S, and -B only have 6.8M, 8.0M, and 8.9M parameters respectively, which are more parameter-efficient than ResNets and VGG16. These results demonstrate that TextNet models are effective for text detection on the GPU device.

Applications

Perfect for real-world text detection tasks, including:

• Natural scene text recognition
• Multi-lingual and multi-oriented text detection
• Document text region analysis

Contribution

This model was contributed by Raghavan , jadechoghari and nielsr .