Here’s a sample README.md for your project on GitHub:

---

Automated Image Captioning using Deep Learning Techniques

Overview

This project focuses on automated image captioning using deep learning techniques. We employed Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models to generate captions for images. The ResNet-50 model, pretrained on large image datasets, was used for feature extraction. This work was conducted as part of the Neural Networks and Deep Learning course under the guidance of Professor Braun Jerome.

Project Details

Dataset

• MS COCO 2017 Dataset: A large-scale dataset for object detection, segmentation, and captioning.
  • train2017: ~118,000 images with captions (used for training).
  • val2017: ~5,000 images with captions (used for validation).

Approach

1. Data Preparation:

   • Images: Resized, cropped, flipped, converted to tensors, and normalized.
   • Captions: Tokenized, converted to lowercase, added special tokens, and transformed into numerical indices.

2. Encoder-Decoder Architecture:

   • Encoder: Pre-trained ResNet-50 model used for feature extraction.
   • Decoders:
     • LSTM: Captures long-term dependencies in sequences.
     • GRU: Simpler architecture, potentially fewer parameters, and competitive performance.

3. Training and Evaluation:

   • Metrics: METEOR and BLEU scores used to evaluate caption quality.
   • Hyperparameters: Batch size of 256, embedding size of 256, hidden size of 512, 5 epochs, Adam optimizer, and cross-entropy loss.

Results

• LSTM Model: Achieved a slightly higher METEOR score (0.2360) indicating better alignment with human judgment in content and fluency.
• GRU Model: Achieved a slightly higher BLEU score (0.2868) reflecting better precision in n-gram overlap.

Future Work

• Explore advanced encoder architectures.
• Integrate attention mechanisms to improve caption quality.