🧠 Image Recognition with Convolutional Neural Networks: Rock-Paper-Scissors 🪨✋✌️

This project uses Deep Learning techniques to classify hand gestures in the classic game of Rock-Paper-Scissors. With varying hand positions, orientations, and gesture styles, the challenge lies in creating robust models capable of accurate predictions.

---

📂 Dataset

The dataset is sourced from Kaggle: Rock-Paper-Scissors Images, containing thousands of labeled images. Data was split into:

• Training (80%)
• Validation (6%)
• Testing (14%)

---

🎯 Objective

To design, implement, and evaluate models capable of classifying images into one of the three categories:

• ✊ Rock
• ✋ Paper
• ✌️ Scissors

The study explores:

• The effectiveness of Transfer Learning (VGG16 and ResNet architectures).
• Building a custom CNN from scratch, both with and without Ridge regularization.
• Techniques like data augmentation, fine-tuning, and kernel visualization.

---

⚙️ Approach

🔄 Transfer Learning

1. VGG16 and ResNet were pretrained on ImageNet and fine-tuned for this dataset.
2. The architecture included frozen layers, dropout regularization, and task-specific dense layers.

🛠️ Custom Models

• A CNN inspired by Inception, analyzing images at two scales using convolutional layers with 7x7 and 3x3 filters.
• Ridge regularization was applied to combat overfitting and gradient vanishing issues.

---

🌟 Results

Performance Metrics

• Custom Model (with Ridge): Achieved 100% accuracy on the test set.
• VGG16 (Transfer Learning): High accuracy with notable misclassifications in "Paper" gestures.

Visual Insights

Heatmaps:

• Correct classifications focused on hand contours.
• Misclassifications were often due to incomplete hands or unconventional gestures.

Kernel Activations:

• The VGG16 kernels retained texture-detection biases from ImageNet, emphasizing the need for fine-tuning.
• Custom models demonstrated effective edge and shadow detection.