Classification of garbage based on lightweight deep learning multimodal method

Abstract

The rapid increase in household waste due to rising living standards has made efficient garbage classification a critical global challenge. Traditional manual sorting methods suffer from low efficiency, high labor intensity, and inconsistent accuracy. This paper presents a lightweight multimodal deep learning framework that integrates feature information from multiple models to improve waste categorization accuracy. Three advanced lightweight neural network architectures: EfficientNet-B0, MobileNetV3, and DenseNet121 are investigated as image feature extractors within a multimodal fusion framework. The dataset comprises 40 categories of common household waste, with 14,743 training images and 400 testing images before augmentation. After applying data augmentation techniques (rotation, flipping, and brightness adjustment), the dataset expands to 40,400 training images and 1,600 testing images. Experimental results demonstrate that the Feature Fusion method achieves superior performance with an accuracy of 0.95 and an F1-score of 0.94 on the test set. Notably, the Ensemble voting method achieve competitive accuracies of 0.92 and F1-score 0.91, respectively, with less training time. With approximately 18.7M parameters, the model is suitable for deployment on resource-constrained devices. These findings indicate that lightweight multimodal networks can effectively balance accuracy and computational efficiency for real-world automated waste sorting.