Abstract
Non-biodegradable garbage threatens terrestrial, coastal, and marine ecosystems. Manual labor-based waste detection and management approaches are limited by scalability, cost, and operational feasibility, especially in remote or hazardous environments. Computer vision and deep learning have made automated litter identification possible. Due to class imbalance in litter datasets, object detection systems often struggle with accurate recognition of underrepresented waste types like batteries, shattered glass, and aerosol cans. This points out the need of creating detection systems that stay efficient in several litter types and environmental settings.This work makes use of an open-source, community-curated repository produced for litter detection activities based on Trash annotations in Context (TACO) dataset. Among 1,500 annotated images TACO provides from various real-world sites are urban streets, beaches, forests, and industrial areas. The dataset has more than sixty waste categories, each tagged with bounding boxes. These observations give contextual relevance and spatial accuracy. Over datasets including TrashNet, Open Litter Map, and the Waste Images Dataset TACO offers good environmental diversity and annotation granularity. The long-tailored distribution and imbalance of the dataset capture the actual difficulties of waste management and offer a suitable benchmark for assessing detection models.
The research evaluated an object detection system based on Faster R-CNN with ResNet-FPN backbones when applied to the TACO dataset through various experimental conditions. Techniques such as Stratified sampling, targeted data augmentation techniques, and weighted loss functions were used to address the imbalance issues during training. This project also deployed a web-based prototype using a real-world trash image dataset in California, demonstrating basic functionalities of trash recognition, summarization, and visualization.