基于深度学习与统计模型的垃圾分类任务研究

A Study of Deep Learning and Statistical Models for Multiclass Classification of Household Waste

摘要: 随着城市生活垃圾产量持续增长，垃圾分类已成为推进资源循环利用与城市精细化治理的重要环节。针对传统人工分拣效率低、成本高以及智能分类模型在复杂场景下泛化能力不足的问题，本文基于华为云生活垃圾分类数据集，构建了一个融合统计模型与深度学习的多模型对比实验框架，对生活垃圾多分类任务展开研究。本文首先采用HOG-LBP-HSV多手工特征融合提取轮廓、纹理、色彩三个维度的图像特征，构建支持向量机(SVM)、随机森林(RF)和XGBoost三种传统统计基线模型。随后基于迁移学习策略，选取ResNet18、MobileNetV3-Large和Swin-Tiny三种深度学习模型进行微调训练。实验中对原始训练集按9:1划分为训练集与验证集，并将原始测试集作为独立测试集，统一采用准确率、宏精确率、宏召回率、宏F1值、ROC-AUC以及推理FPS、参数量、FLOPs算力指标等指标对各模型进行综合评估，并引入ANOVA方差分析验证性能差异显著性，结合归一化混淆矩阵与高频错分样本完成可解释分析。实验结果表明，深度学习模型在垃圾分类任务上显著优于传统统计模型。其中，Swin-Tiny在测试集上取得了最优的分类性能，ResNet18和MobileNetV3-Large也表现出较强的特征学习能力与泛化性能。相比之下，SVM、随机森林与XGBoost三个传统模型受限于手工特征表达能力，在复杂类别识别上性能相对较弱。进一步的混淆矩阵和错误案例分析表明，模型的主要误差集中于外观相似、材质接近的垃圾类别之间。综上，本文验证了迁移学习深度模型在生活垃圾多分类任务中的有效性，并为智能垃圾分类系统的模型选型与落地部署提供了实验依据。

Abstract: With the continuous growth of municipal solid waste, waste classification has become an important part of resource recycling and fine-grained urban governance. To address the low efficiency and high cost of manual sorting, as well as the limited generalization ability of intelligent classification models in complex scenarios, this study constructs a multi-model comparative framework based on a public waste image dataset and investigates the waste multi-classification problem using both statistical machine learning and deep learning methods. This paper first extracts multi-dimensional visual features—including contour, texture, and color information—by fusing HOG-LBP-HSV descriptors, and subsequently trains three classical models: Support Vector Machine (SVM), Random Forest (RF), and XGBoost. Then, three pretrained deep learning models, ResNet18, MobileNetV3-Large, and Swin-Tiny, are fine-tuned under a transfer learning strategy. The original training set is split into training and validation subsets with a ratio of 9:1, while the original test set is used as an independent test set. Comprehensive performance assessment is conducted using multiple metrics, including Accuracy, Macro-Precision, Macro-Recall, Macro-F1, ROC-AUC, as well as computational efficiency indicators such as inference FPS, parameter count, and FLOPs. ANOVA was introduced to verify the significance of performance differences, and an interpretable analysis was completed by combining the normalized confusion matrix and high-frequency misclassified samples. Experimental results show that deep learning models significantly outperform traditional statistical baselines in waste classification. Among all models, Swin-Tiny achieves the best overall performance on the test set, while ResNet18 and MobileNetV3-Large also demonstrate strong feature learning and generalization capabilities. In contrast, SVM, RF, and XGBoost are constrained by handcrafted feature representations and show relatively weaker performance in recognizing visually similar categories. Further confusion matrix and misclassification case analyses indicate that most errors occur between categories with similar appearance and material characteristics. Overall, this study validates the effectiveness of transfer learning-based deep models in municipal waste classification and provides useful evidence for model selection and practical deployment in intelligent waste sorting systems.