Cricket Shot Classification ML Project

🏆 Achieved 97.14% Accuracy with Ensemble Learning

This project implements advanced machine learning models to classify cricket shots into four categories with exceptional accuracy using ensemble learning techniques.

🎯 Project Status: COMPLETED ✅

Final Results:

• Best Individual Model: ResNet50V2 (95.24% accuracy)
• Best Ensemble Method: Weighted Averaging (97.14% accuracy)
• Target Achieved: ✅ 95%+ accuracy goal exceeded

---

📊 Quick Results Overview

Model	Accuracy	Method
Custom CNN	90.90%	Individual
ResNet50V2	95.24%	Transfer Learning
Advanced CNN	91.64%	Enhanced Architecture
Ensemble (Weighted)	97.14%	3-Model Combination

Class-wise Performance (Best Ensemble):

• Drive: 98% precision, 95% recall
• Legglance-Flick: 96% precision, 96% recall
• Pullshot: 97% precision, 99% recall
• Sweep: 97% precision, 99% recall

---

📁 Project Structure

shotpredictionml/
├── 🎯 main.py                                   # Main entry point (EASY TO USE!)
├── 🏆 ensemble_predictor.py                     # Core ensemble prediction engine
├── 🔮 predict.py                                # Original prediction script
├── 📊 quick_ensemble.py                         # Ensemble evaluation
├── 📋 README.md                                 # This file
├── 📦 requirements.txt                          # Dependencies
├── 📊 CRICKET_SHOT_CLASSIFICATION_PROJECT.md    # Complete documentation
├── 📁 saved_models/                             # Trained models
│   ├── ensemble_model_1.keras                   # Custom CNN (90.90%)
│   ├── ensemble_model_2.keras                   # ResNet50V2 (95.24%)
│   ├── best_advanced_model.h5                   # Advanced CNN (91.64%)
│   └── advanced_cricket_shot_classifier.h5      # Backup model
├── 📁 old_models/                               # Previous model versions
├── 📁 scripts/                                  # Additional scripts
├── 🏏 drive/                                    # Drive shot images (1,257)
├── 🏏 legglance-flick/                          # Legglance-flick images (1,117)
├── 🏏 pullshot/                                 # Pullshot images (1,257)
└── 🏏 sweep/                                    # Sweep shot images (1,117)

---

🚀 Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Use the Main Predictor (EASIEST!)

# Single image prediction
python main.py --image drive/drives1.png

# Interactive mode
python main.py --interactive

# Batch prediction (first 5 images)
python main.py --batch drive/

3. Alternative Methods

# Original prediction script
python predict.py --image path/to/image.jpg

# Ensemble evaluation
python quick_ensemble.py

---

🏗️ Model Evolution

Journey from 25% to 97.14% Accuracy:

1. Original Classifier (25% accuracy) ❌

   • Failed due to double preprocessing and over-aggressive augmentation

2. Working Classifier (88% accuracy) ✅

   • Fixed preprocessing issues with single pipeline

3. Advanced Classifier (91.64% accuracy) ✅

   • Added Batch Normalization and deeper architecture

4. Ensemble Classifier (97.14% accuracy) 🏆

   • Combined 3 models using weighted averaging

---

🎯 Ensemble Methods Used

1. Hard Voting (Majority Vote)

• Accuracy: 95.34%
• Method: Each model votes, majority wins

2. Soft Voting (Probability Averaging)

• Accuracy: 96.72%
• Method: Average predicted probabilities

3. Weighted Averaging (Best Method)

• Accuracy: 97.14% 🏆
• Method: Weighted average based on individual performance
• Weights: [0.3, 0.4, 0.3] for [Custom CNN, ResNet50V2, Advanced CNN]

---

🔧 Technical Implementation

Model Architectures:

• Custom CNN: Sequential CNN with 3 conv blocks
• ResNet50V2: Transfer learning with ImageNet pre-training
• Advanced CNN: Enhanced CNN with Batch Normalization

Training Configuration:

• Image Size: 224x224 pixels
• Batch Size: 32
• Learning Rate: 0.001 (Adam optimizer)
• Data Augmentation: Rotation, shifts, zoom, brightness
• Callbacks: Early stopping, learning rate reduction, model checkpointing

Data Preprocessing:

• Resize: All images to 224x224
• Color Conversion: BGR to RGB
• Normalization: Pixel values scaled to [0,1]

---

📈 Performance Analysis

Individual Model Performance:

Model	Architecture	Accuracy	Precision	Recall	F1-Score
Custom CNN	Sequential CNN	90.90%	0.91	0.91	0.91
ResNet50V2	Transfer Learning	95.24%	0.95	0.95	0.95
Advanced CNN	Enhanced CNN	91.64%	0.92	0.92	0.92

Ensemble Performance:

Ensemble Method	Accuracy	Drive	Legglance-Flick	Pullshot	Sweep
Hard Voting	95.34%	0.95	0.94	0.96	0.96
Soft Averaging	96.72%	0.97	0.96	0.97	0.97
Weighted Averaging	97.14%	0.98	0.96	0.97	0.97

---

🎮 Usage Examples

Basic Prediction:

from predict import CricketShotPredictor

# Load the best ensemble model
predictor = CricketShotPredictor('best_advanced_model.h5')
predictor.load_model()

# Predict single image
shot_type, confidence, probabilities = predictor.predict_single_image('image.jpg')
print(f"Predicted: {shot_type} with {confidence:.2%} confidence")

Batch Prediction:

# Predict multiple images
results = predictor.predict_batch('folder_with_images/', max_images=50)
for result in results:
    print(f"{result['image']}: {result['prediction']} ({result['confidence']:.2%})")

Interactive Mode:

# Interactive prediction session
predictor.interactive_predictor()

---

🔍 Key Features

✅ What Works:

• Ensemble Learning: Combines multiple models for better accuracy
• Transfer Learning: Uses pre-trained ResNet50V2 for feature extraction
• Data Augmentation: Balanced augmentation prevents overfitting
• Batch Normalization: Stabilizes training and improves convergence
• Early Stopping: Prevents overfitting with automatic stopping

🎯 Problem Solutions:

• Fixed Double Preprocessing: Single clean data pipeline
• Reduced Augmentation: Conservative augmentation strategy
• Enhanced Architecture: Deeper networks with regularization
• Model Diversity: Different architectures capture different features

---

📚 Technical Concepts

Convolutional Neural Networks (CNNs)

• Extract spatial features from cricket shot images
• Translation invariant feature detection
• Hierarchical feature learning

Transfer Learning

• Use pre-trained models (ResNet50V2) on new tasks
• Faster training and better performance
• Leverage ImageNet features for cricket shots

Ensemble Learning

• Combine multiple models for improved accuracy
• Reduce variance and improve generalization
• Weighted averaging based on individual performance

---

🚀 Future Improvements

Data Quality:

• Collect more diverse angles and lighting conditions
• Add new shot types (cut, hook, defensive, etc.)
• Include different players and batting styles

Model Enhancements:

• Attention mechanisms for key body parts
• Multi-scale feature extraction
• Temporal information from video frames

Advanced Techniques:

• Curriculum learning
• Focal loss for hard examples
• Advanced ensemble methods (stacking, boosting)

---

🏆 Achievements

Performance Milestones:

1. ✅ Baseline Working: 88.00% accuracy
2. ✅ Advanced Model: 91.64% accuracy
3. ✅ Target Achieved: 97.14% accuracy
4. ✅ Ensemble Success: Weighted averaging method

Technical Achievements:

• Systematic problem diagnosis and solution
• Architecture optimization with Batch Normalization
• Successful transfer learning implementation
• Effective ensemble learning techniques
• Comprehensive error analysis

---

📞 Support & Documentation

• Complete Documentation: CRICKET_SHOT_CLASSIFICATION_PROJECT.md
• Model Evolution: old_models/README.md
• Error Analysis: misclassification_analyzer.py
• Prediction Script: predict.py

Troubleshooting:

1. Model Loading: Ensure .keras or .h5 files are in correct location
2. Dependencies: Install all requirements with pip install -r requirements.txt
3. Memory Issues: Reduce batch size for large datasets
4. Data Format: Ensure images are in supported formats (PNG, JPEG)

---

🎯 Conclusion

This project successfully demonstrates the power of ensemble learning and transfer learning in computer vision tasks. Starting from a failed baseline model stuck at 25% accuracy, we systematically diagnosed issues, implemented solutions, and achieved an impressive 97.14% accuracy using a weighted ensemble of three diverse models.

Project Status: ✅ COMPLETED - Target accuracy achieved
Final Accuracy: 97.14% 🏆
Ensemble Method: Weighted Averaging
Models Used: Custom CNN, ResNet50V2, Advanced CNN