Deep Learning: CNN Image Classification & Transfer Learning

Build robust image-classification pipelines with CNNs, transfer learning, fine-tuning, data augmentation, and API-based deployment.

Categories

Tech Used

Problem

Image-classification systems must handle variation in lighting, image quality, class imbalance, backgrounds, and visual ambiguity while still producing reliable predictions on unseen data.

Approach

• Built end-to-end preprocessing, normalization, contrast-enhancement, and data-augmentation pipelines
• Trained CNN classifiers and transfer-learning models using established architectures including AlexNet, GoogLeNet, and ResNet
• Fine-tuned pretrained networks and compared model behavior across architectures
• Applied class-balancing and threshold-aware evaluation to improve robustness on imbalanced datasets
• Connected the trained model with a FastAPI-based application workflow for practical inference

Results

• Created a reusable deep-learning image-classification pipeline from preprocessing through deployment
• Improved model robustness for noisy, low-quality, and visually challenging image samples
• Built a transferable workflow suitable for medical imaging, quality inspection, product classification, and other image-recognition tasks

Demo Videos