Research On The Classification And Few Shot Object Detection Algorithm Of Thangka Yidam Based On Deep Learning

Thangka,as the main carrier of Tibetan culture,is a window to explore the traditional Tibetan culture and religious customs.The recognition of the main categories and the detection of key objects in Thangka Yidam can help us obtain important semantic information of Thangka images,thus further interpreting image content and feel the living Tibetan culture.In recent years,with the rapid development of computer technology,more and more image processing technology has been applied in the protection and inheritance of Thangka culture.This paper studies thangka images based on deep learning,and the main research work and innovations are as follows:(1)The research on the classification and recognition method of Thangka Yidam.The unique cascade mode of DenseNet can alleviate the gradient disappearance during deep training,but it also leads to a large number of redundant information reuse.In addition,this method does not learn the weight correlation between feature channels from the feature dimension,so the efficiency of feature extraction is low.To solve the above problems,a SMAD model based on DenseNet and SENet is proposed for Thangka image classification in this paper.Specifically,on the basis of retaining the advantages of DenseNet structure,this model incorporates SE module for weight adaptive calibration between each dense block and the transition layer.At the same time,the Leaky Re LU is used as the nonlinear activation function of DenseNet bottleneck layer,and the pooling function of SENet is improved.Finally,the validity of the model is verified on the constructed Thangka dataset.The experimental results show that the performance of the improved model is better than that of the traditional DenseNet,and the accuracy reaches 95.51%.(2)The research on the key objects detection method of Thangka Yidam.The traditional data-driven detection model does not have the ability to adapt to target deformation and target overlap,and cannot accurately locate and accurately identify the key objects of the Thangka Yidam.Therefore,in this paper,a small sample detection method based on Res Net and deformable convolution is proposed for the study of the headwear and sits of Thangka Yidam.Firstly,the hierarchical structure of Res Net network is adjusted and optimized to reduce the depth of the middle layer network and increase the depth of the shallow layer network,so that the network pays more attention to the details of the Thangka image,reduces the image feature loss and reduces the amount of calculation.Secondly,the deformable convolution is integrated into the feature extraction stage to adapt to the deformation of the target and focus on the key feature information of the target,so as to achieve the purpose of accurately extracting the features of the Thangka image.Finally,the proposed dual-threshold non-maximum suppression algorithm is used to detect overlapping targets in the border regression stage,and the missed and repeated detection rates of targets are reduced to further improve the detection accuracy.The experimental results on the Thangka dataset show that the proposed method has better performance than the FSOD model.The AP and AP50 under the 2-way 5-shot task are 33.3% and 71.2%,respectively,which are 4.7% and 5.3% higher than those under the traditional model,respectively.