Multiclass Image Classification using CNN

Overview

Built and deployed a lightweight CNN architecture achieving 73% accuracy with 15.17s training time, optimized for edge devices and real-world deployment.

Technical Architecture

• Model: Custom-designed lightweight CNN (TinyVGG architecture) optimized for edge deployment
• Framework: PyTorch with CUDA acceleration
• Dataset: Curated subset of Food101 dataset (3 classes, 75 training/25 testing images per class)
• Input Pipeline: Automated image processing with data augmentation
• Performance: 73% classification accuracy with 15.17s training time on CPU

Results

Challenges and Solutions

• Overfitting Mitigation: Addressed overfitting by incorporating data augmentation techniques such as TrivialAugmentWide, improving generalization on unseen data.
• Scalability: Designed the system to scale efficiently with larger datasets by leveraging PyTorch’s modular design for easy expansion.

Learning Outcomes

• Gained expertise in designing and implementing custom CNN architectures tailored for specific tasks.
• Developed skills in data preprocessing and augmentation to enhance model robustness.
• Learned advanced PyTorch techniques for building scalable and efficient machine learning pipelines.

To view the project code, visit the GitHub repository