Image Classification using Convolutional Neural Networks

🎯 Goal

To implement and compare various convolutional neural network (CNN) architectures for image classification tasks using the CIFAR-10 and MNIST datasets, analyzing their efficiency and accuracy.

🧵 Dataset

• CIFAR-10 Dataset: Link to CIFAR-10 Dataset
• MNIST Dataset: Link to MNIST Dataset

🧾 Description

This project involves the implementation of different CNN models to classify images from the CIFAR-10 and MNIST datasets accurately. The models range from simple architectures suitable for educational purposes to complex networks designed for high accuracy on larger, more complex datasets.

🧮 What I had done!

1. Dataset Preparation:
   • Loaded and preprocessed the CIFAR-10 and MNIST datasets (normalization, resizing, etc.).
2. Model Implementation:
   • Implemented several CNN architectures (LeNet-5, MobileNet, ResNet50, Simple CNN, VGG16).
3. Training and Validation:
   • Trained each model on the training datasets.
   • Validated the models on the validation datasets.
4. Performance Evaluation:
   • Evaluated model performance using accuracy, precision, recall, and F1-score.

🚀 Models Implemented

1. LeNet5_Model
   • Chosen for its simplicity and historical significance in CNN development, particularly for digit recognition tasks such as MNIST.
2. MobileNet_Model
   • Selected for its efficiency in mobile and embedded applications, utilizing depthwise separable convolutions, tested on both CIFAR-10 and MNIST.
3. ResNet50_Model
   • Implemented for its deep architecture and residual connections, solving the vanishing gradient problem, evaluated on CIFAR-10.
4. Simple_CNN_Model
   • Used as a basic model to understand CNN architecture and its implementation, tested on both CIFAR-10 and MNIST.
5. VGG16_Model
   • Included for its deep architecture with small 3x3 convolutions and proven performance on complex datasets like CIFAR-10.

📚 Libraries Needed

• TensorFlow/Keras
• PyTorch
• NumPy
• Pandas
• Matplotlib
• scikit-learn

📈 Performance of the Models based on the Accuracy Scores

• LeNet5_Model:
  • CIFAR-10: Accuracy - 0.5802000000%
  • MNIST: Accuracy - 0.9862000000%
• MobileNet_Model:
  • CIFAR-10: Accuracy - 0.5521000000%
  • MNIST: Accuracy - 0.9843000000%
• ResNet50_Model:
  • CIFAR-10: Accuracy - 0.3908000000%
  • MNIST: Accuracy - 0.9789000000%
• Simple_CNN_Model:
  • CIFAR-10: Accuracy - 0.6947000000%
  • MNIST: Accuracy - 0.9929000000%
• VGG16_Model:
  • CIFAR-10: Accuracy - 0.7196000000%
  • MNIST: Accuracy - 0.9919000000%

📢 Conclusion

From the results obtained, it is evident that different CNN architectures perform variably on the CIFAR-10 and MNIST datasets due to their unique design and complexity. Here are the key observations and conclusions derived from this project:

1. Simple_CNN_Model:

   • Achieved the highest accuracy on the MNIST dataset (99.29%), making it the most suitable for this specific task.
   • Also performed well on the CIFAR-10 dataset (69.47%), showcasing its robustness and versatility.

2. VGG16_Model:

   • Attained the highest accuracy on the CIFAR-10 dataset (71.96%), indicating its effectiveness for more complex image classification tasks.
   • Also showed strong performance on the MNIST dataset (99.19%).

3. LeNet5_Model:

   • Performed exceptionally well on the MNIST dataset (98.62%) due to its design tailored for digit recognition.
   • However, it was less effective on the CIFAR-10 dataset (58.02%), which is expected given its simpler architecture.

4. MobileNet_Model:

   • Showed decent performance on both CIFAR-10 (55.21%) and MNIST (98.43%), highlighting its efficiency and suitability for mobile applications despite slightly lower accuracy.

5. ResNet50_Model:

   • Surprisingly, had the lowest accuracy on the CIFAR-10 dataset (39.08%) among all models tested, possibly due to overfitting or insufficient training epochs.
   • Achieved a good accuracy on the MNIST dataset (97.89%), but it was outperformed by simpler models in this context.

In conclusion, the Simple_CNN_Model and VGG16_Model stood out as the best performers for MNIST and CIFAR-10 respectively. The Simple_CNN_Model is highly recommended for simpler tasks like digit recognition, while the VGG16_Model is better suited for more complex image classification tasks. Despite the lower accuracy, MobileNet_Model offers a good balance between performance and computational efficiency, making it ideal for deployment in resource-constrained environments. The results also highlight the importance of selecting the appropriate architecture based on the specific dataset and task at hand.

✒️ Your Signature

Utsav Singhal
LinkedIn | GitHub | Email