TriFusionNet: A Heterogeneous Feature Fusion Approach for Multi-Class Waste Image Classification

Abstract

The growing waste volume and the variety of waste materials demand reliable, automatic waste classification systems. While deep learning techniques have been effective for waste image classification, existing methods are limited to a single feature extraction backbone, potentially missing the rich set of visual attributes of waste materials. To solve this problem, this paper introduces TriFusionNet, a three-branch heterogeneous feature-fusion network consisting of EfficientNet-B0, ResNet18 and a lightweight custom CNN. The framework extracts high-level, residual, and task-specific local feature representations from the input images and combines them into a unified 1920-dimensional feature representation for classification. The images are integrated from both the TrashNet and Waste Classification Data datasets to create a balanced 5-class waste dataset of 29,085 images, including glass, metal, organic, paper and plastic. The model is based on the idea of using pre-trained backbones with transfer learning and data augmentation, and regularization to enhance generalization. Accuracy, precision, recall, F1-score, confusion matrix and ROC-AUC are used to assess the effectiveness of the proposed framework. A systematic ablation study involving seven configurations is conducted to investigate the contribution of the individual and combined feature-extraction branches. The total classification accuracy of TriFusionNet is 97.16%, and the ablatio results show that the full three-branch configuration yields the highest accuracy with respect to the different configurations evaluated. In addition, the image regions that support the model’s predictions are qualitatively analyzed using Grad-CAM visualization. The results show that the proposed heterogeneous feature fusion approach demonstrates the potential of heterogeneous feature fusion for robust five-class waste image classification.