Research On Dongba Few-shot Classification Based On Graph Neural Network

Dongba painting of Naxi nationality is a treasure of Chinese minority culture.It is also the most original and unique ethnic cultural works in Southwest China.In Dongba paintings,there are figures,animals,ghosts,gods and other Dongba artistic images.They are typical images of Dongba paintings,which originate from the daily life and worship of Naxi people.However,Dongba painting art works are located in remote areas and the local economy is relatively backward.These works have a long history but lack of protection,so there are few of them in existence.The classification of Dongba paintings is of great significance to the protection of ethnic minority art works and the artistic archiving and digital processing of Dongba paintings.Existing image classification methods are either based on the underlying features of the image,which can only classify the specific texture style but not the specific image,or based on the deep neural network,which requires a large number of data sets as training set,and is not applicable to the Dongba painting image which only have small sample size and low data volume.So,in view of the outstanding lines and clear colors of the Naxi Dongba paintings,we propose an end-to-end classification method based on graph neural network.Firstly,in order to solve the problem of insufficient information extraction or information loss when the original image is converted into embedded nodes,the features of Dongba painting are obtained by using multi-resolution and multi-scale image feature extraction network,fused with edge features then embedded into the graph neural network as nodes.Secondly,in the process of graph neural network classification,in order to more comprehensively represent and save the relationship information required by node classification,the two-dimensional edge weight feature is used to replace the traditional one-dimensional edge weight as the classification basis,and the two-dimensional edge feature can save the relationship information between the two kinds of nodes at the same time,that is to better retain the intra class similarity and inter class dissimilarity required by node classification.Thirdly,a method combining self-attention mechanism and feature saliency attention mechanism is proposed to guide the updating of node features,which solves the shortcomings of insufficient relationship expression between nodes and lack of saliency judgment and filtering of each feature in the process of feature aggregation updating of neighborhood nodes,which makes the process of node feature updating more targeted and effective.Then,through the episodic training method,the intra class similarity,inter class dissimilarity,feature saliency and node features are transferred to update the edgelabeling features,so that the edge-labeling features can completely reflect the inter class relationship between nodes,and obtain the edge-labeling features among all nodes.Then the final classification is obtained by the weight of edge-labeling features between querry nodes and other nodes.The experimental results show that the method can better extract the features of Dongba painting image and more sufficiently retain local details and structural features required for image classification of Dongba painting.the classification accuracy is better on the low-data set of Dongba painting image.