HES-Net: An Efficient and Scalable Deep Learning Framework for Waste Image Classification

Abstract

Automated waste classification has become an essential component of intelligent waste management systems and sustainable environmental monitoring. Although deep convolutional neural networks (CNNs) achieve high classification accuracy, many existing architectures require significant computational resources, limiting their deployment In resource-limited systems like embedded systems, IoT devices and mobile devices. To address this challenge, this study proposes HybridEffiShuffleNet (HES-Net), a lightweight hybrid deep learning framework for efficient multi-class waste image classification. The proposed architecture integrates EfficientNet-B0 and ShuffleNetV2 into a dual-branch feature-extraction framework to capture complementary semantic and structural representations from waste images. An attention-based feature fusion mechanism is introduced to enhance discriminative feature learning and improve classification robustness. The proposed model was evaluated on a five-class waste image dataset comprising glass, metal, organic, paper, and plastic. Extensive preprocessing and data augmentation techniques were applied to improve model generalization. Experimental results demonstrate that HES-Net achieved an overall classification accuracy of 96.34\%, outperforming EfficientNet-B0 (95.20\%), ShuffleNetV2 (94.05\%), ResNet-18 (91.48\%), and ShuffleNetV2-S (76.05\%) under identical training conditions. Furthermore, the proposed framework exhibited consistent performance across all waste categories while maintaining computational efficiency suitable for real-world deployment. The findings indicate that HES-Net provides an effective and scalable solution for intelligent waste classification and smart recycling applications.