An improved lightweight multi-granularity feature fusion method for waste image classification

Abstract

Lightweight waste image classification models often have limited feature representation ability, especially when distinguishing visually similar waste categories. In contrast, many high-accuracy deep learning models require large numbers of parameters and high computational costs, which limits their use on resource-constrained platforms such as smart waste bins, mobile devices, and embedded recycling terminals. To address this problem, this thesis proposes an improved Lightweight Multi-Granularity Feature Fusion Network (LWMF-Net) for waste image classification.

LWMF-Net uses MobileNetV3-Large as the backbone and introduces a multi-granularity feature fusion module to combine low-level texture features, middle-level structural features, and high-level semantic features from different network stages. This design strengthens the representation ability of the lightweight backbone while maintaining a compact structure. Generalized mean pooling is also adopted to improve spatial feature aggregation and help the model focus on discriminative image regions.

The proposed model is evaluated through a quantitative experimental approach using a fixed training-validation-test split, ablation experiments, and controlled comparison with representative models. The dataset includes six waste classes: cardboard, glass, metal, paper, plastic, and trash. Compared with ResNet101, VGG16, ViT-B/16, Inception-V3, ShuffleNetV2-x1.0, and SqueezeNet1.1, LWMF-Net achieves the best overall performance, with 94.21% accuracy and a 93.35% macro F1-score on the test set. With only 4.17 million parameters and a model size of 16.13 MB, the results show that LWMF-Net provides a good balance between classification accuracy and deployment efficiency.