面向形态多变垃圾的轻量化视觉语言模型细粒度分类方法研究

Research on Fine-Grained Classification Method of Lightweight Vision-Language Model for Morphologically Variable Waste

摘要: 针对垃圾分类场景下垃圾形态多变、细粒度类别特征相近、训练样本稀缺以及边缘部署算力不足等难题，本文基于轻量化视觉语言模型SmolVLM-256，开展小样本细粒度垃圾分类的性能优化研究。结合真实场景中垃圾挤压、遮挡、复杂背景干扰等特点，本文构建了包含8类生活垃圾的细粒度数据集并搭建标准化实验评估框架。通过超参数调优、语义提示引导、数据增强与分类头结构改良等多维度优化实验，探究各策略对模型分类效果的影响。实验结果显示，优化后的SmolVLM-256分类准确率可达87.92% ± 1.56%，相较基线模型性能大幅提升。横向对比ViT、CLIP、BLIP、Qwen3-VL等模型可知，该轻量化模型在保障优良分类精度的同时，具备参数量小、推理成本低的优势，更适配边缘设备部署。研究表明，合理的训练与结构优化策略，能够有效提升轻量化视觉语言模型在复杂小样本垃圾场景的识别能力，可为边缘智能垃圾分类系统的研发与落地提供技术参考。

Abstract: Aiming at the practical challenges of variable garbage morphology, similar fine-grained feature distribution, limited training samples, and insufficient computing resources for edge deployment in garbage classification scenarios, this paper conducts a performance optimization study on small-sample fine-grained garbage classification based on the lightweight vision-language model SmolVLM-256. Considering real-world interferences such as garbage extrusion, occlusion, and complex background environments, a fine-grained dataset containing eight types of domestic garbage is constructed, and a standardized experimental evaluation framework is established. Multiple optimization experiments including hyperparameter tuning, semantic prompt guidance, data augmentation, and classification head structure improvement are carried out to explore the influence of different strategies on model classification performance. The experimental results show that the optimized SmolVLM-256 achieves a classification accuracy of 87.92% ± 1.56%, which obtains a significant performance improvement compared with the baseline model. Comparative experiments with ViT, CLIP, BLIP, and Qwen3-VL demonstrate that the lightweight model maintains competitive classification accuracy with fewer parameters and lower inference cost, making it more suitable for edge deployment. The research verifies that reasonable training strategies and structural optimization can effectively enhance the recognition capability of lightweight vision-language models in complex small-sample garbage scenarios, providing technical references for the design and practical application of edge intelligent garbage classification systems.