EcoSmart An AIBased Waste Segregation, Recommendation, and Analytics System

Improper waste disposal poses severe environmental and health risks, calling for automated, accurate sorting systems. This project presents an end‑to‑end web‑based waste classification system that integrates two deep learning approaches: a Siamese neural network trained from scratch to learn similarity embeddings, and a fine‑tuned EfficientNetV2L model equipped with focal loss, RandAugment, and test‑time augmentation (TTA). The Siamese architecture compares image pairs to produce discriminative embeddings; classification is performed by matching a query embedding against class prototypes, achieving 99.9% validation accuracy. The EfficientNetV2L model, trained on the same dataset with two‑stage fine‑tuning and modern regularisation techniques, achieved 80.03% .

Both models are deployed as plug‑in components in a Django web application secured by JWT authentication. Users can register, log in, upload waste images, and receive real‑time predictions of the appropriate bin colour (Green, Blue, Black), bio‑type (Biodegradable or Non‑Biodegradable), and a disposal suggestion (e.g., compost, recycle, hazardous waste facility). All predictions are stored, enabling users to view historical records, per‑class analytics (biodegradable/non‑biodegradable counts, bin usage percentages), and a profile management page. The system also includes a leaderships section and a logout functionality.

Experimental evaluation demonstrates that the Siamese network excels at capturing fine‑grained similarities, while the EfficientNetV2L provides robust performance with advanced augmentation. The integrated web application offers a practical, scalable solution for households, municipalities, and recycling centres, promoting sustainable waste management through artificial intelligence.

Keywords: Waste classification, Siamese network, EfficientNetV2L, focal loss, RandAugment, test‑time augmentation, Django, JWT authentication.