Problem Statement

Waste sorting is largely manual, inefficient, and error-prone. Incorrect sorting leads to contamination of recyclable materials, reducing recycling effectiveness and increasing environmental impact.

Solution

Developed a CNN-based image classification system using transfer learning (ResNet50) fine-tuned on a curated waste dataset. Implemented data augmentation to improve generalization across varied lighting and backgrounds.

Key Features

• ▸6-class waste categorization (glass, paper, cardboard, plastic, metal, trash)
• ▸Transfer learning with ResNet50 backbone
• ▸Extensive data augmentation pipeline
• ▸Real-time inference capability
• ▸Confusion matrix and per-class accuracy reporting
• ▸Model export for edge deployment

Challenges

• ⚡High intra-class variability in waste appearance
• ⚡Class imbalance in training data
• ⚡Distinguishing visually similar categories (e.g., plastic vs. glass)
• ⚡Achieving robust performance across varied lighting conditions

Results & Metrics

Lessons Learned

• 💡Transfer learning dramatically reduces training time and data requirements
• 💡Data augmentation is crucial for real-world CV application robustness
• 💡Confusion matrix analysis reveals systematic errors that guide model improvement