Fine-Tuned Image Classifier

Problem

Training image classifiers from scratch is data- and compute-intensive, especially for smaller, domain-specific datasets.

Approach

Applied transfer learning with MobileNetV2, ResNet, and VGG16 backbones, fine-tuning each on a target dataset and comparing accuracy and efficiency trade-offs.

Tech highlights

Python with TensorFlow/Keras for model definition and training, OpenCV for image preprocessing, and a head-to-head comparison across the three architectures.

Results

Full write-up with accuracy comparisons across architectures coming soon — check back or follow the GitHub repo for progress.

Challenges

Avoiding overfitting when fine-tuning large ImageNet backbones on a small target dataset — tuning how many layers to unfreeze and how low to drop the learning rate so the pretrained features survived, and keeping the comparison fair when MobileNetV2, ResNet, and VGG16 have very different parameter counts and memory footprints.