✊✋✌️ Rock-Paper-Scissors Image Classification Using CNN

📌 Project Overview

This project classifies hand gestures into Rock, Paper, or Scissors categories using Convolutional Neural Networks (CNNs).

To better understand performance differences, we first implemented traditional Machine Learning (ML) and Artificial Neural Network (ANN) approaches. These methods were limited in generalization capability for unseen image data, as they rely heavily on manually extracted features.

This served as a baseline comparison to highlight the superior performance of CNNs, which can automatically learn spatial hierarchies from raw images. CNNs outperformed ML and ANN models significantly in terms of accuracy and robustness.

---

---

🗂️ Dataset Overview

We used an image dataset containing labeled hand gesture images:

• rock/ – images of a closed fist
• paper/ – images of an open hand
• scissors/ – images of two extended fingers

⚠️ The dataset is not included in this repository due to size limits.
Please use your own or download a public Rock-Paper-Scissors image dataset and structure it as shown above.
📩 If you would like access to the dataset used in this project, feel free to contact me at uqashazahid@gmail.com.

---

Dataset/
├── train/
│ ├── rock/
│ ├── paper/
│ └── scissors/
├── validation/
│ ├── rock/
│ ├── paper/
│ └── scissors/
└── test/
├── rock/
├── paper/
└── scissors/

---

---

📁 Files Overview

📂 CNN/

The CNN/ folder contains all files related to the Convolutional Neural Network implementation:

• cnn.ipynb – This Jupyter Notebook includes all code for:

  • Training the CNN model
  • Performing real-time predictions
  • Evaluating the model
  • Visualizing performance

• Saved Keras model files – Best-performing models are saved after training.

• Image Augmentation – Various techniques (rotation, zoom, flip, etc.) were used to:

  • Expand dataset diversity
  • Improve accuracy
  • Reduce overfitting

📂 Machine_Learning/

The Machine_Learning/ folder contains:

• ml.ipynb – A complete ML pipeline including:
  • Image preprocessing & path extraction
  • Saving labels and paths to Excel
  • Feature extraction
  • Standardization
  • Model training & evaluation

---

🔬 Traditional ML Models

We trained several ML models using extracted image features:

models = {
    "Logistic Regression": LogisticRegression(multi_class='multinomial', solver='lbfgs', max_iter=1000),
    "SVM (RBF Kernel)": SVC(kernel='rbf'),
    "Random Forest": RandomForestClassifier(),
    "k-NN": KNeighborsClassifier(n_neighbors=5),
    "Decision Tree": DecisionTreeClassifier(),
}

📈 Model Evaluation Results

✅ Best Machine Learning Model: SVM (RBF Kernel)

      precision    recall  f1-score   support
   0     0.9802    0.9914    0.9858       349
   1     0.9970    0.9880    0.9925       333
   2     0.9884    0.9855    0.9870       346

🤖 Artificial Neural Network (ANN) Results

129/129 ━━━━━━━━━━━━━━━━━━━━ 1s 7ms/step accuracy: 1.0000 loss: 9.6035e-04

📌 Although ANN achieved perfect training accuracy, this indicates possible overfitting, highlighting why CNNs are more suitable for image-based tasks.

🧠 Why CNN?

CNNs were ultimately preferred because:

• 🚫 No need for manual feature extraction
• 📈 Better performance on unseen image data
• 🧭 Automatically learn spatial hierarchies from raw pixels
• 💪 More scalable and robust for real-world applications

---

📊 CNN Final Results

After applying data augmentation and tuning:

• ✅ Training Accuracy: 99.98%
• ✅ Validation Accuracy: 100%

---

🛠️ Requirements

⚠️ Note: Please use Python 3.9 to 3.12.
TensorFlow is not supported in Python versions above 3.12.

• Python 3.9–3.12
• TensorFlow / Keras
• NumPy
• Pandas
• Matplotlib
• Seaborn
• OpenCV (optional)