Goal : Develop a robust deep learning model that accurately classifies CIFAR-10 images while maintaining high performance under different types of image noise and corruption.
What Was Done
Built a custom ResNet-based CNN using TensorFlow with advanced data augmentation (MixUp, CutMix, CutOut), label smoothing, AdamW optimization, warmup cosine learning rate scheduling, early stopping, Test-Time Augmentation (TTA), and robustness evaluation against Gaussian, Salt & Pepper, and Blur noise.
Results
91.75% Clean Accuracy
92.84% TTA Accuracy
82.11% Gaussian Noise
81.69% Salt & Pepper Noise
Tech Stack
-Python • TensorFlow/Keras • NumPy • Scikit-learn • SciPy • Matplotlib • Google Colab, Git & GitHub
