An open API service indexing awesome lists of open source software.

https://github.com/narius2030/imcp-support-blinders

This project focuses on image captioning by creating two primary models: DarkNetLM and DarkNetVG2. Both models leverage the CSP DarkNet53 architecture as the backbone of YOLOv8 for feature extraction from images. Combining with Transformers or LSTM to generating captions.
https://github.com/narius2030/imcp-support-blinders

computer-vision data-lake image-captioning large-language-model mobile-app

Last synced: 8 days ago
JSON representation

This project focuses on image captioning by creating two primary models: DarkNetLM and DarkNetVG2. Both models leverage the CSP DarkNet53 architecture as the backbone of YOLOv8 for feature extraction from images. Combining with Transformers or LSTM to generating captions.

Awesome Lists containing this project

README

          

# 🚦 Image Caption Support Blinder

## 📑 Table of Contents

- [🚦 Image Caption Support Blinder](#-image-caption-support-blinder)
- [📑 Table of Contents](#-table-of-contents)
- [📝 Overview](#-overview)
- [📁 Folder Structure](#-folder-structure)
- [🔗 Pipeline](#-pipeline)
- [🤖 Models](#-models)
- [🚀 Usage](#-usage)
- [1. Environment Setup](#1-environment-setup)
- [2. Run the Pipeline Steps](#2-run-the-pipeline-steps)
- [📊 Results](#-results)
- [🤝 Contributing](#-contributing)
- [🪪 License](#-license)
- [🙏 Acknowledgments](#-acknowledgments)
- [💁‍♂️ Author](#️-author)

## 📝 Overview

This project provides a full pipeline for building a Vietnamese traffic image captioning dataset and system, supporting both computer vision research and accessibility for the visually impaired. The pipeline includes data crawling, cleaning, caption generation, and image augmentation.

## 📁 Folder Structure

```
IMCP-Support-Blinders/
├── README.md
├── Data/
│ ├── dashboard.dio
│ ├── image.png
│ ├── README_en.md
│ ├── README_vn.md
│ ├── README.md
│ ├── 1.crawl_data/
│ │ ├── env_crawl_data.yaml
│ │ ├── output/
│ │ │ ├── metadata.json
│ │ │ ├── traffic_images_dataset_v1.csv
│ │ │ ├── traffic_images_dataset_v2.csv
│ │ │ └── traffic_images_dataset_v3.csv
│ │ └── python/
│ │ └── traffic_raw.py
│ ├── 2.data_preprocessing/
│ │ ├── jupyter/
│ │ │ └── data_preprocessing.ipynb
│ │ └── output/
│ ├── 3.labels_short_captions/
│ │ ├── output/
│ │ │ └── csv_with_captions/
│ │ └── python/
│ └── 4.Image_data_augument/
│ ├── output/
│ └── python/
├── Model/
│ ├── DarkNetLM/
│ └── DarkNetVG2/
```

- **Data/**: Main pipeline and dataset scripts, including crawling, preprocessing, captioning, and augmentation.

👉 For more details, see the [Germini-Captioning-Dataset-2025 repo](https://github.com/TrieuPhi/Germini-Captioning-Dataset-2025)
- **Model/**: Model code and experiments (details to be updated).

## 🔗 Pipeline

1. **Crawl Data**
Collect traffic images from Google Images via SerpApi, save metadata and images locally.

2. **Data Preprocessing**
Clean data: remove broken URLs, handle nulls, standardize fields.

3. **Labels Short Captions**
Use Gemini 2.0 Flash API to generate concise (10-15 words) captions for each image.

4. **Image Augmentation**
Augment data using modern image transformation techniques with [Albumentations](https://albumentations.ai/).

## 🤖 Models

**To be updated.**

## 🚀 Usage

### 1. Environment Setup

```bash
cd Data/1.crawl_data/
conda env create -f env_crawl_data.yaml
conda activate crawl_data
pip install -r ../3.labels_short_captions/python/requirements.txt
pip install -r ../4.Image_data_augument/python/requirements.txt
```

### 2. Run the Pipeline Steps

**Step 1: Crawl data**

```bash
python 1.crawl_data/python/traffic_raw.py
```

**Step 2: Data cleaning**

- Use the notebook or script in `2.data_preprocessing/jupyter/`

**Step 3: Caption generation**

```bash
python 3.labels_short_captions/python/label_short_captions.py
```

**Step 4: Augmentation**

```bash
python 4.Image_data_augument/python/data_augument.py
```

## 📊 Results

- Augmented images and captions are stored in `Data/4.Image_data_augument/output/` and `augmented/`.
- Cleaned CSVs and intermediate results are in the corresponding `output/` folders.

## 🤝 Contributing

Contributions are welcome! Please submit a pull request or open an issue for any suggestions or improvements.

## 🪪 License

This project is licensed under the MIT License. See the LICENSE file for more details.

## 🙏 Acknowledgments

- BERT, GPT-2, and Albumentations authors.

## 💁‍♂️ Author

- [Narius2030](https://github.com/Narius2030) 🦉
- [HTN-DT-Beo](https://github.com/HTN-DT-Beo) 🐻
- [TrieuPhi](https://github.com/TrieuPhi) 🚀