基于残差网络双数据增强的贺兰山动物岩画分类研究OA

As As prehistoric non-textual historical records,Helan Mountain rock art bears the profound cultural memories of northern nomadic peoples,and its classification and organization are the fundamental prerequisites for interpreting its cultural connotations and social connections.To address the classification difficulties caused by severe weathering,morphological similarity between categories,and the subjective limitations of manual interpretation in traditional classification methods of Helan Mountain animal rock art,and under the challenges of small sample sizes and high inter-class similarity of rock art images,this paper systematically validates the effectiveness of a synergistic augmentation strategy combining RandAugment and CutMix/MixUp in the study of specific rock art images,and provides a reference for model selection and hyperparameter configuration.With four typical types of animal rock art in Helan Mountain-cattle,sheep,deer,and tigers-selected as research objects,ResNet18 is first determined as the optimal backbone network through hyperparameter optimization and model depth comparison.Subsequently,a dual augmentation strategy is introduced to expand sample diversity and enhance local feature extraction.Experimental results indicate that an average F1 score of 95.73%is achieved on the test set,which shows a significant improvement in classification accuracy compared to the baseline model.The performance is superior to ResNet10 and ResNet50,and the confusion between similar species such as sheep and deer is effectively resolved.This study demonstrates the application potential of deep learning technology in the organization of prehistoricnon-textual historical materials,and provides effective methodological support for the construction of large-scale,standar dized digital archives and digital protection of rock art.