| Petroglyphs are the carriers of ancient ancestors’ expressions of language,containing the cultural activities and social conditions of the time,and therefore petroglyphs have important research value and historical significance.Due to the differences in cultural expressions between ancient and modern times,the contents expressed by the ancestors in rock paintings are generally abstract,and it is difficult for people to understand the connotations without professional cultural knowledge.This paper combines computer vision and natural language processing to semantically associate the rock paintings of Helan Mountain,enabling people to understand the content expressed in the rock paintings through modern language,which is of great significance for the preservation and transmission of cultural heritage.In this paper,we study the semantic association of rock paintings in Helan Mountain based on CNN-LSTM model,The main work done in this paper are:(1)We obtained the rock painting pictures of Helan Mountain through two ways:field photography of the rock painting scenic area and interception of the ancient books of rock painting in Helan Mountain,and expanded the number of rock painting pictures of Helan Mountain through data enhancement,etc.Finally,we divided and labeled the rock paintings of Helan Mountain to make the data set required for the experiment of this paper.(2)The CNN-LSTM model was constructed to train and test the Heilanshan rock painting dataset.The experimental results show that the trained model can realize the semantic association of Heilanshan rock paintings,and the descriptive utterances generated after the test match the content of Heilanshan rock paintings,and the generated utterances are more natural.(3)The ReLU activation function of the convolutional layer in the vgg16 model is replaced by the LeakyReLU function to train,test and evaluate the Heilanshan petroglyph dataset,and the experimental results show that the statements generated by the improved model test are more accurate and more imageable than before the improvement,while the evaluation indexes BLEU1,BLEU2,BLEU3 and BLEU4 scores of the improved model are more accurate and more imageable than before the improvement The scores of the improved model evaluation indexes BLEU1,BLEU2,BLEU3,and BLEU4 are improved by 0.167,0.219,0.279,and 0.308,respectively,compared with those before improvement. |
