🧠 Image Caption Generator

This project demonstrates how Transfer Learning can be applied to automatically generate captions for images using a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN) (specifically, an LSTM).
The model combines InceptionV3 (pre-trained on ImageNet) for visual feature extraction and an LSTM network for text sequence generation.

---

🎯 Objective

To build a deep learning model that can look at an image and describe it in natural language showcasing understanding of both computer vision and natural language processing concepts through transfer learning.

---

🧩 Concept: What Is Transfer Learning?

Transfer learning means reusing a model’s pre-learned knowledge from one task to solve another, similar task.
Instead of training a model from scratch (which requires a large dataset and computational power), we take a pre-trained CNN that already understands general image patterns (edges, colors, shapes) and adapt it for image captioning.

---

⚙️ How It Works

1. Feature Extraction (Vision Part):

   • Use InceptionV3, a pre-trained CNN on ImageNet, to extract a 2048-length feature vector for each image.
   • These features represent high-level visual understanding of the image.

2. Caption Generation (Language Part):

   • Use an LSTM (Long Short-Term Memory) network to generate text sequences word-by-word.
   • The LSTM learns how to describe the extracted features in natural language.

3. Integration:

   • The CNN encodes the image → The LSTM decodes that information into words.
   • The model is trained end-to-end to minimize caption prediction error.

---

🧪 Dataset

For this demonstration, a small custom dataset of sample images (soccer player, dog, and car) was used to keep runtime light and GPU usage minimal.
Each image has a few manually defined captions for supervised training.

Precisely the Flickr8k dataset was used. (Dataset gets auto downloaded via code)

---

🚀 Model Architecture

• Feature Extractor: InceptionV3 (Transfer Learning)
• Language Model: LSTM with Embedding
• Loss: Categorical Crossentropy

---

🧰 Tech Stack

• Python
• TensorFlow / Keras
• NumPy, Pillow, Matplotlib

---

🖼️ Example Output

---

🧾 Results & Insights

• The model successfully learned to describe images in short, meaningful phrases.
• Demonstrates semantic understanding (colors, actions, and objects).
• Even with a small dataset, transfer learning helped achieve good generalization.