Improving Convolutional Neural Network - Based Waste Classification Through Data Augmentation and Synthetic Data Generation
Blessing Oriaguaman · 2025
The global increase in waste production has necessitated efficient and cost-effective waste management solutions. Traditional waste sorting methods, such as those employed by Wolverhampton City Council, are often inefficient and expensive. This research investigates the enhancement of Convolutional Neural Network (CNN)-based waste classification through data augmentation and synthetic data generation to address these challenges. By improving the accuracy and robustness of CNN models in identifying diverse waste types, this study aims to contribute to more sustainable waste management practices. This research Implemented various data augmentation techniques and employed modified lightweight Generative Adversarial Networks (GANs) to generate synthetic waste images, to enhance the training datasets and address the limitations of existing models. Experimental results demonstrate that the augmented and synthetic data significantly improves the models' performance. The VGG16 model's accuracy increased from 63.30% to 78.28%, the ResNet50 model's accuracy improved from 79.63% to 80.47%, the MobileNetV2 model's accuracy increased from 35.94% to 83.18%, and the InceptionV3 model's accuracy improved from 69.19% to 73.92%. Notably, the MobileNetV2 model, after augmentation, achieves the highest accuracy of 83.18%, showcasing its potential for real-world waste classification applications. This research contributes to the growing body of knowledge on AI-driven waste management solutions. Findings suggest that adopting CNN-based waste classification can reduce operational costs and enhance the overall efficiency of waste management processes, aligning with sustainability goals.