| With the impetus of digital technology,the cultural and creative industries have gradually formed a digital cultural and creative ecology by combining with artificial intelligence,big data and other digital technologies.Nowadays,the digital manufacturing process of cultural and creative industry needs to strengthen the protection of cultural creative design,which is usually described by multiple view images for the appearance of three-dimensional objects,thus the research of multi-view image classification method for creative design comes into being.The traditional single-view image classification method is difficult to describe the relationship between the views of three-dimensional objects.In contrast,multi-view images often contain more spatial location information between them,which can be efficiently classified by using the correlation between images.Therefore,effectively identifying and classifying the objects corresponding to multi-view images is a worthy research topic.In this thesis,the research is carried out from multi-view image classification methods,and the main contributions of the thesis can be summarized as follows:(1)In the multi-view image classification method based on multi-head self-attention,we first propose a Multi-View Convolutional Neural Network based on Self-Attention(MVCNNSA),and construct a network model using the multi-head self-attention module alone by evaluating the correlation between different views using the multi-headed attention mechanism in the Vision Transformer(ViT)model.Then we propose a Multi-view image classification model based on depth-wise convolution Vision Transformer(MViT),which solves the problem of CNN multi-view classification model in perceiving the spatial location relationship between images by using the ViT architecture and its multi-head attention with the property of capturing global features.Meanwhile,the lack of inductive bias in ViT results in a poor ability to capture local features of image blocks.Thus,we design a depth-convolution-based ViT module DViT(Depth-wise convolution ViT module)to capture the local features by introducing a depthconvolution mechanism.In order to further expand the distance of different samples and reduce the distance between feature expressions of the same class of samples,contrastive loss is introduced to further improve the classification of the model.(2)In the multi-view classification method based on wavelet transformer and global filtering module,we first propose a Multi-view image classification model based on the Global filtering Vision Transformer(MGViT).A global filtering module is designed by using the property that the Fourier transform can learn the interactions between spatial positions in the frequency domain.The long-distance dependencies within the image are mined using position encoding and the Fourier transform acting on the frequency domain.The model complexity is reduced to log-linear complexity by following the principle that ViT does not set the induction bias.Then we propose a Multi-view image classification model with joint Wavelet-mix Vision Transformer(MWViT),using the feature of wavelet decomposition with down-sampling to implement feature map pooling,which can effectively reduce the loss of feature information.The hybrid wavelet module splices and fuses the three detailed high-frequency components of wavelet decomposition and then splices them with the low-frequency components.Finally,the overall model can learn both air domain features and frequency domain features by fusion strategy on the basis of MGViT model.In this thesis,the experiments were conducted on the public dataset Model Net40/10 and a design patent dataset Patent-MNIST of Guangdong Key Laboratory of Intellectual Property &Big Data for classification tasks and ablation experiments.The results show the effectiveness of the method in this thesis. |
