🧠 AI Projects: NLP & Computer Vision

This repository is a curated collection of AI projects spanning Computer Vision, Natural Language Processing (NLP), and Multi-Modal AI. Each project is implemented in a self-contained Jupyter Notebook (or script) with explanations, code, and results.

🔗 Useful Links:

• 📂 GitHub Repo
• 🤖 Hugging Face Transformers
• 🖼️ Keras Applications
• 📊 Fashion-MNIST Dataset
• 📸 Flickr30k Dataset

---

📊 Project Summary (Skills Matrix)

Project	Domain	Key Skills / Techniques	Tools & Frameworks
Face Mask Detection with VGG16	CV	Transfer Learning, Binary Classification, Data Augmentation	TensorFlow, Keras, VGG16
Facial Emotion Recognition with VGG16	CV	Multi-class Classification, Emotion Recognition, Transfer Learning	Keras, VGG16
Fashion MNIST Classification with CNNs	CV	CNN Architectures, Model Comparison, Feature Visualization	TensorFlow, Keras
Green Screening with OpenCV	CV	Chroma Keying, Real-time Video Processing	OpenCV, NumPy
Image Deblurring with VGG16 + DCGAN	CV	GANs, Perceptual Loss, Image Restoration	DCGAN, VGG16, TensorFlow
Image Captioning with Flickr30k	CV + NLP	Encoder-Decoder, Seq2Seq, BLEU Evaluation	VGG16, LSTM, Keras
Tweets Sentiment Analysis (3 Neural Nets)	NLP	Sentiment Analysis, DNN/CNN/RNN Comparison, Embeddings	TensorFlow, Keras, GloVe
GenZ Tweets Data Pipeline	NLP	Text Preprocessing, Regex, Lemmatization, Emoji Normalization	NLTK, SpaCy, Python
Next Word Prediction with Bi-LSTM	NLP	Language Modeling, Sequence Prediction, Perplexity	TensorFlow, Keras, Bi-LSTM
Prompt-to-Synopsis Generator	NLP	Fine-Tuning Transformers, Creative Text Generation	HuggingFace, GPT-2
AI Long-Form Story Generator	NLP	Long-Context Modeling, Story Generation	HuggingFace, Transformers
AI Imagining Stories from Images	Multi-Modal	Image-to-Text, Vision+Language, Storytelling	HuggingFace, Transformers

---

📂 Projects

🖼️ Computer Vision

1. Face Mask Detection with VGG16

• Situation: During COVID-19, monitoring mask compliance became critical in public spaces.

• Task: Build a system to automatically detect masks from images of people.

• Action:

  • Fine-tuned VGG16 (transfer learning) pretrained on ImageNet.
  • Applied data augmentation (rotation, flipping, zoom) for robustness.
  • Built a binary classifier on labeled mask/no-mask dataset.

• Result: Achieved 97% accuracy on validation data, demonstrating production feasibility for surveillance and healthcare use cases.

• Tags: TensorFlow · Keras · Transfer Learning · CNN · Image Classification · Model Deployment

  

---

2. Facial Emotion Recognition with VGG16

• Situation: Emotion recognition is important in mental health monitoring, human-computer interaction, and customer analytics.

• Task: Develop a model to classify facial images into multiple emotions.

• Action:

  • Preprocessed FER-2013 dataset with grayscale normalization & augmentation.
  • Fine-tuned VGG16 with added dense layers for 7-class classification.
  • Used categorical cross-entropy loss and early stopping.

• Result: Reached 72% accuracy, surpassing traditional ML baselines (e.g., SVMs ~45%).

• Tags: Keras · VGG16 · Image Classification · Emotion Recognition · Transfer Learning · FER-2013

  

---

3. Fashion MNIST Classification with CNNs

• Situation: Fashion MNIST is a standard benchmark for testing deep learning models on real-world classification tasks.

• Task: Classify clothing images into 10 categories.

• Action:

  • Built multiple CNN architectures (2–4 conv layers, pooling, dropout).
  • Compared models using accuracy and loss curves.
  • Visualized feature maps for interpretability.

• Result: Achieved 92% accuracy on test set with deeper CNN.

• Tags: TensorFlow · CNN · Fashion-MNIST · Model Comparison · Deep Learning

  

---

4. Green Screening with OpenCV

• Situation: Film, broadcasting, and AR applications rely on chroma key (green screen).

• Task: Replace green backgrounds in images and videos with arbitrary scenes.

• Action:

  • Used OpenCV to detect and mask green pixel ranges.
  • Replaced with background images/videos dynamically.
  • Implemented for both static images and live video streams.

• Result: Delivered real-time background replacement with smooth transitions.

• Tags: OpenCV · Computer Vision · Chroma Keying · Real-Time Video Processing

  

---

5. Image Deblurring with VGG16 + DCGAN

• Situation: Blurred images affect critical fields like surveillance and medical imaging.

• Task: Restore sharpness in blurred images.

• Action:

  • Built a DCGAN generator-discriminator architecture.
  • Used VGG16 perceptual loss to guide training.
  • Trained on custom mixed-blur dataset (motion blur, Gaussian blur).

• Result: Restored sharper images with SSIM score improvement of +18% over baseline interpolation.

• Tags: GAN · DCGAN · VGG16 · Image Restoration · Perceptual Loss

  

---

6. Image Captioning with Flickr30k

• Situation: Image captioning aids accessibility for the visually impaired and powers multimedia search.

• Task: Generate natural language descriptions of images.

• Action:

  • Extracted features with VGG16 encoder.
  • Trained LSTM decoder with teacher forcing on Flickr30k captions.
  • Evaluated with BLEU scores.

• Result: Generated fluent captions like “A boy playing with a dog in the grass” with BLEU-4 ≈ 0.41.

• Tags: VGG16 · LSTM · Seq2Seq · Encoder-Decoder · Image Captioning · Flickr30k

  

---

📝 Natural Language Processing (NLP)

7. Tweets Sentiment Analysis with 3 Neural Networks

• Situation: Businesses and political campaigns monitor sentiment on Twitter for decision-making.

• Task: Classify tweets into positive, negative, or neutral sentiment.

• Action:

  • Built three deep neural network architectures (DNN, CNN, RNN).
  • Preprocessed with regex, stopword removal, embeddings (GloVe).
  • Compared architectures on accuracy/F1.

• Result: Best-performing CNN achieved 88% accuracy on test data.

• Tags: NLP · Sentiment Analysis · DNN · CNN · RNN · Embeddings

  

---

8. GenZ Tweets Data Pipeline for Sentiment Analysis

• Situation: Raw social media data is noisy with slang, emojis, and hashtags.

• Task: Design a reusable pipeline for tweet preprocessing.

• Action:

  • Implemented regex cleaning, tokenization, lemmatization.
  • Normalized emojis, URLs, and @mentions.
  • Built pipeline both in Jupyter and as a standalone Python script.

• Result: Produced clean, structured text improving sentiment model accuracy by ~10%.

• Tags: NLP · Data Pipeline · Regex · NLTK · SpaCy · Preprocessing

  

---

9. Next Word Prediction with Bi-Directional LSTM

• Situation: Next-word prediction powers mobile keyboards and search engines.

• Task: Build a language model to predict the next word.

• Action:

  • Preprocessed text corpus into n-grams.
  • Trained a Bi-LSTM sequence model with embeddings.
  • Evaluated using perplexity and prediction accuracy.

• Result: Generated accurate predictions with perplexity reduced to ~35, suitable for autocomplete.

• Tags: Bi-LSTM · Language Modeling · Sequence Prediction · Text Generation

  

---

10. Prompt-to-Synopsis Generator (Fine-Tuning)

• Situation: Entertainment & content industries need tools to expand short prompts into story outlines.

• Task: Fine-tune transformer to generate synopses from short prompts.

• Action:

  • Fine-tuned GPT-2 using HuggingFace Transformers.
  • Applied causal LM loss, LR scheduling, and early stopping.
  • Evaluated coherence & diversity of outputs.

• Result: Produced multi-sentence coherent synopses with logical flow from prompts.

• Tags: Transformers · GPT-2 · Fine-Tuning · HuggingFace · Text Generation

  

---

11. AI Long-Form Story Generator with Varied Context

• Situation: Longer context improves story coherence but increases complexity.

• Task: Build a model for generating long-form stories.

• Action:

  • Used transformer causal language models.
  • Tested varying context sizes and prompt strategies.

• Result: Generated coherent multi-paragraph stories; longer context improved narrative consistency.

• Tags: Transformers · Causal LM · Text Generation · HuggingFace · Long-Context Modeling

---

🔮 Multi-Modal AI

12. AI Imagining Stories from Images

• Situation: Bridging computer vision with NLP enables storytelling from images.

• Task: Generate stories based on input images.

• Action:

  • Extracted image features via pretrained vision encoders.
  • Used HuggingFace causal language models to generate narratives.

• Result: Produced imaginative, contextually relevant stories from fantasy to real-world scenes.

• Tags: Multi-Modal AI · Vision + Language · Transformers · HuggingFace · Image-to-Text

---

📜 License

This project is licensed under the MIT License.