Abstract: Traditional waste aluminum (Al) alloy sorting technologies face technical bottlenecks such as low efficiency and poor accuracy. Image-based classification methods struggle to accurately identify Al grades due to the similar external characteristics; laser-induced breakdown spectroscopy (LIBS) methods face the analytical accuracy and stability problem caused by the matrix effects and heterogeneity. Therefore, this paper proposes an improved ResNet18-SVM-RF fusion classification model using image-spectral bimodal data to improve accuracy and robustness. The model extracts image features by an improved lightweight ResNet18 network, integrates these features with LIBS spectral features processed by support vector machine (SVM), and subsequently utilizes a Random Forest classifier for the final classification. This approach achieves the purpose of efficient and accurate classification. Experimental results demonstrate that the proposed model exceeds 97% on accuracy and recall metrics, significantly outperforming both the ResNet18 model based on single image data and the SVM method based on single spectral data. Furthermore, it attains a 72% reduction in classification prediction time compared to the original ResNet18 model, demonstrating its computational efficiency. This proposed model provides an effective solution for waste Al alloy intelligent sorting and holds broad application prospects.