Evaluating Deep Learning Architectures for Waste Image Classification: From Cross-Domain Training to Edge Deployment

Abstract

The continuous increase in municipal waste generation presents a significant challenge for sustainable development, emphasizing the importance of intelligent waste classification systems. Deep learning-based classification systems have achieved reliable results, enabling the automatic identification of waste categories through image analysis. This paper presents a systematic performance evaluation of three convolutional neural network architectures, ResNet-50, MobileNetV2, and EfficientNet-B0, for six-class waste image classification across three dataset configurations: TrashNet-only (T), TACO-only (A), and a merged multi-domain dataset (T+A). Results show that EfficientNet-B0 demonstrates the highest classification accuracy across all configurations, reaching 95% on TrashNet, 83% on TACO, and 92% on the combined dataset. Multi-domain training on (T+A) consistently outperforms TACO-only training across all architectures, demonstrating the benefit of merging studio and real-world imagery for enhanced generalization. The highest-performing architecture is deployed on a Raspberry Pi 4 Model B using TensorFlow Lite, achieving 90.72% accuracy, an average prediction latency of 304.7 ms, and a lightweight model footprint of 15.92 MB, demonstrating the feasibility of accurate waste classification on a system.