Application of Huangling mianhua image recognition and classification system based on deep learning

Traditional methods for recognizing Huangling mianhua images suffer from poor generalization, insufficient local-detail perception and weak robustness, because they still rely on hand-crafted features or standard convolutional neural networks. As a typical intangible cultural heritage (ICH) pattern that integrates regional culture and folk beliefs, the intelligent recognition of Huangling mianhua (dough ornament) poses a higher challenge to model accuracy and robustness. To this end, this article proposes a mianhua image recognition system that integrates convolutional neural network (CNN) and CBAM (Convolutional Block Attention Module) to enhance the model’s ability to focus on key information and improve image discrimination performance. During the research process, this study first collects multi-angle Huangling mianhua original image data through professional imaging equipment and constructs a standardized dataset through Laplacian clarity screening and GrabCut interactive foreground segmentation; then, the improved ResNet-18 network architecture is designed, and the CBAM attention module is embedded in each residual block to jointly drive the discriminative expression of key features through channel-spatial attention. A systematic data enhancement strategy is deployed simultaneously and combined with WarmUp startup, Cosine annealing learning rate scheduling, and L2 regularization to optimize the training process to enhance the model’s generalization ability; finally, by building four types of baseline model comparison test frameworks and ablation experiment systems, a closed-loop verification of model accuracy, robustness, and computational efficiency is achieved. The experimental results show that the CBAM-ResNet-18 system constructed in this article achieves a classification accuracy of 93.67% and an F1-Score of 93.0% on the standard test set, and its overall performance is better than that of various control systems. In the robustness test, the average recognition accuracy of the proposed system reaches 85.68%; in the sub-test set constructed for typical extreme scenes such as blur, occlusion, and strong light interference, it shows good adaptability and robustness. In addition, the average single inference time of the system is only 0.06 s, which is significantly better than the inference efficiency of other control models. In summary, the recognition system proposed in this study has outstanding performance in terms of accuracy, robustness, and real-time performance, providing a feasible and high-performance technical path for the intelligent recognition of complex ICH images, which has important research value and practical application prospects.