Research On Detection Of Tibetan Opera Mask Images Based On Few Shot

Tibetan opera is a kind of drama and art with a long history.As the main carrier of Tibetan opera,Tibetan opera mask is a window to explore the traditional Tibetan culture.Its artistic expression language,unique national style and artistic charm have become the artistic treasures of the Tibetan people and even the whole world.The detection of Tibetan opera mask images can help people interpret the content of Tibetan opera and feel the Tibetan culture.Because of the large pixel differences,different feature mapping space domains and low similarity between similar classes and high similarity between some different classes,the target detection of Tibetan opera mask images has raised higher requirements.In this paper,we propose a detection method for Tibetan opera masks with small samples mainly from two aspects: feature extraction network and model improvement.The main research contents are as follows:(1)Constructing a Tibetan opera mask dataset and analyzing the dataset characteristics.In order to understand the characteristics of the dataset and explore the multidimensional properties of the Tibetan opera mask dataset,this paper analyzes the dataset mathematically,and finds that the number of images in each class is unbalanced and there is a high similarity between classes and a low similarity between classes.To expand the limited number of Tibetan opera mask images,random rotation,scaling,cropping and Mosaic operations are performed on the dataset.The experiments prove that the data expanded dataset can achieve better detection results under the same model.(2)A low-connected multi-scale fusion feature extraction network for small-sample Tibetan opera masks was built.Since the number of neural network model parameters is too large,the detection accuracy is too low in small sample data sets.In this paper,we reduce the number of parameters and design convolutional kernels to fuse the feature information of different layers to obtain a multi-scale feature map,so that the model has better feature expression capability.The results of Darknet53 network and the feature extraction network proposed in this paper are compared on a selfbuilt Tibetan opera mask dataset.The experiments show that the feature extraction network can effectively improve the average accuracy of the detection of Tibetan opera masks.(3)A channel attention object detection model based on Loss＿recall for Tibetan opera masks is designed.First,the channel attention module is introduced into the network model,and based on the proposed feature extraction network,the attention of the network model to the key target is strengthened,and the attention of the model to the background is reduced,so as to improve the target detection accuracy of the model;second,for the problem of low recall due to low similarity between classes of Tibetan opera masks,the Loss＿recall loss function is designed to improve the detection model’s recall of the detection model is improved.Finally,a neural network connection strategy is established to prevent the model from overfitting and improve the detection robustness of the model.The experiments show that the Loss＿recall-based attentional object detection model for Tibetan opera mask channel achieves better results on the Tibetan opera mask dataset.