Research on Waste Image Classification Algorithm Based on Improved YOLOv8

With the acceleration of urbanization, the production of household waste continues to increase, while traditional manual sorting methods suffer from low efficiency, high cost, and unstable accuracy. Deep-learning-based image classification technology provides an effective solution for automated waste classification. However, household waste images present challenges such as large variations in object scale, complex backgrounds, densely packed small objects, and easily confusable categories, making it difficult for existing models to meet practical classification requirements. To address these issues, this paper proposes BE-YOLOv8, an improved household waste image classification model based on YOLOv8, which integrates multiple strategies including data augmentation, attention mechanisms, and edge feature enhancement. First, to tackle the problems of limited training samples and class imbalance, an improved LMix data augmentation method is proposed. By introducing a label smoothing strategy, dynamically correcting mixed label weights, and adding a regularization penalty term to the loss function, the generalization ability of the model is effectively improved. Second, an Edge-Guided Multi-Scale Hybrid (EGMSH) attention mechanism is designed, which enhances the model′s perception of edge contours and multi-scale texture features through online edge computation, multi-scale depthwise separable convolutions, and adaptive gating fusion. Finally, a learnable BoundaryEdge feature enhancement module is proposed, which utilizes a trainable color projection layer and fixed-weight Sobel operators to generate high-quality edge features online and embeds them into the network via residual connections, significantly improving the discrimination of shape-similar and easily confusable categories. Experiments are conducted on a household waste image dataset containing 26,994 images across 20 categories. The results demonstrate that BE-YOLOv8 achieves a Top-1 accuracy of 83.5% on the test set, improving by 1.1% over the baseline YOLOv8. The hazardous waste category exhibits the most significant improvement, with a Top-1 accuracy of 92.5%. It is demonstrated that the model has excellent robustness in scenarios with complex backgrounds and easily confusable categories, providing a high-precision technical solution for practical waste classification applications.