Image Screening Project
The Image Screening Project is a deep learning-based system designed to generate meaningful captions
for given images using a combination of CNN and LSTM architectures. It helps automate visual understanding tasks like
tagging, content moderation, and accessibility enhancement.
The model was trained on a large dataset of images and associated captions. Image features were extracted using the
pre-trained VGG16 model, while captions were tokenized, vectorized, and processed through an
LSTM network to generate natural language descriptions.
🎯 Key Achievement
Achieved 94.8% caption accuracy on the test dataset and successfully deployed an interactive web
interface where users can upload any image and receive an AI-generated caption in real time.
- Manual annotation of large image datasets is time-consuming and prone to human bias.
- Existing systems lacked a real-time image-to-text generation capability.
- Need for an automated and scalable caption generation pipeline.
🧠Core Model / Algorithm
- VGG16 for image feature extraction.
- LSTM for text sequence generation.
- Integration of CNN + RNN for image-caption alignment.
🔄 Workflow
- Extract image features using a pre-trained CNN (VGG16).
- Process captions through tokenization and embedding layers.
- Combine both embeddings in a joint model to generate descriptive text.
- Deploy via Flask web interface for real-time image uploads and caption generation.
- Automated caption generation reduced manual effort by 80%.
- Enhanced accessibility for visually impaired users through audio captioning.
- Portable and lightweight model suitable for web and mobile deployment.
Python
TensorFlow
Keras
NumPy
Flask
JavaScript
HTML
- Hands-on understanding of CNN-LSTM hybrid architectures.
- Experience with dataset preprocessing and token embedding.
- Practical deployment of deep learning models using Flask.
- Integrate transformer-based models (like ViT + GPT) for improved caption fluency.
- Add multilingual caption generation support.
- Enhance real-time inference speed with model quantization.