This study explores the integration of AI-generated synthetic data into deep learning pipelines for an automated waste detection and classification in industrial recycling systems. Based on the YOLOv12 object detection framework and the publicly available WaRP dataset, the proposed model was trained for recognizing six major waste categories, including bottles, cans, cardboard, and glass. In order to address data scarcity and class imbalance, up to 10,000 synthetic images were generated by using ChatGPT’s image generation capabilities, compositing realistic waste objects onto clean conveyor-belt backgrounds. The experimental results demonstrated that augmenting the employed dataset with synthetic samples improved the proposed model`s detection and classification performance, which is proven by an increase from 0.593 (for the baseline model) to 0.622 for the mAP@50 metric and from 0.466 to 0.504 for the mAP@50:95 metric. The best results were achieved for the model augmented with 5,000 synthetic images, after which there was no further improvement in the performance of the employed model. These findings highlight the fact that high-quality synthetic data can effectively enhance deep learning models in waste sorting applications, reducing the dependence on extensive manual data collection. However, for further improvements it would be necessary to enhance the asset realism and diversity rather than simply increasing the dataset size. To sum up, the proposed approach underscores the potential of combining generative AI and computer vision for accelerating industrial automation.
*This article represents an extension of the conference paper: “Automated Waste Sorting Using Deep Learning and Synthetic Data”, presented at the CONAT 2024 International Congress of Automotive and Transport Engineering (Part Two: Automobile and Environment).
Deep learning, Synthetic data generation, Object detection, Waste classification, Recycling automation, Industrial computer vision, Generative AI, Data augmentation, Industry 4.0.
Iulian Alexandru OGREZEANU, Constantin SUCIU, Lucian Mihai ITU, "Enhancing Waste Sorting Models with Deep Learning and AI-Generated Synthetic Images*", Studies in Informatics and Control, ISSN 1220-1766, vol. 35(1), pp. 19-32, 2026. https://doi.org/10.24846/v35i1y202602
All articles published by Studies in Informatics and Control (SIC) are published in accordance with the terms and conditions of the Creative Commons Attribution License (CC BY-NC-SA 4.0).