MA/01

Deep Learning / Case study

Waste Classification with VGG16

A transfer-learning classifier that separates organic and recyclable waste using feature extraction and fine-tuning.

View repository

Problem

Manual waste sorting is labor-intensive and can contaminate recyclable material streams.

Solution

Adapted ImageNet-pretrained VGG16 with augmentation, a custom dense head, and fine-tuning of the top convolutional block.

Architecture

Augmented image data → frozen / partially unfrozen VGG16 → dense classifier → organic or recyclable class.

Technical implementation

Trained on an approximately 1,200-image IBM Skills Network dataset. The repository reports 76% test accuracy for frozen feature extraction and 79% after fine-tuning.

Key challenges

Improve generalization on a relatively small binary image dataset while managing class-specific recall trade-offs.

Lessons learned

Fine-tuning improved overall accuracy and organic-class recall, demonstrating the value of task-specific feature adaptation.