Ancient Chinese Character Image Retrieval Model Based On Attention Learning Network

The digitization of ancient Chinese books and documents has laid a foundation for efficient related research on them and documents with the help of information technology.However,due to the complexity of ancient Chinese characters,the flexible writing styles and the degradation of the handwriting layouts caused by a long time,it has brought difficult to realize them automatic retrieval of ancient Chinese character images.In response to this problem,deep learning theories and methods are introduced to study the ancient Chinese character image retrieval model through employing convolution neural network,attention mechanism and difference hash algorithm.(1)An ancient Chinese character image retrieval method(Inception Resnet V2?Ancient Chinese Image,IRV2?ACI)based on CNN(Convolutional Neural Network)is designed,which adapts the font structure of ancient Chinese characters through a multi-scale convolution kernel to improve the retrieval performance.Firstly,the combination of maximum pooling and average pooling is used to get the important information of ancient Chinese characters while preserving the detail information of Chinese characters,for utilizing its advantages of global and detailed features,respectively.Then,convolution layer and combination pooling are used to replace the full connection layer to realize the depth feature fusion of ancient Chinese character images,so as to alleviate the lack of expression of spatial structure caused by full connection and retain the original structure of Chinese characters as much as possible.Finally,the cross entropy loss function is used to train the neural network,and the Adam optimizer is used to optimize the network model to improve the stability of the model and accelerate the convergence speed of the model.The experimental results show that the image features of ancient Chinese characters extracted by the improved Inception Resnet V2 network have better expressiveness for ancient Chinese characters,and the precision and recall rate are improved by 4.15% and 7.72% respectively compared with the compared methods.(2)An ancient Chinese character image retrieval model is designed,which integrates spatial and channel attention network(Spatial Channel Ancient Image Hash Network,SCAIHNet).Firstly,according to the spatial structure characteristics of ancient Chinese characters,integrating the low-dimensional characteristics of space and the high-dimensional characteristics of channel for obtaining more abundant semantic feature information of ancient Chinese characters than combined pooling,and reducing the interference caused by the differences of writing styles of ancient Chinese characters and balancing the samples of data set under the supervision of weighted cross entropy function.Then,an Inception?residual structure module is constructed for enriching the edge and contour details of ancient Chinese character images and reducing the loss of context information between the internal structures of Chinese characters,and eliminating the semantic gap.Finally,the hash layer is added,the logarithmic loss function is designed to enhance the dependence between the classification layer and the hash layer,and the image features of ancient Chinese characters are differentially hash coded to realize the fast and effective retrieval of ancient Chinese character images.The experimental results on GJHZ data set show that under the same experimental environment,the average retrieval accuracy(Mean Average Precision,MAP)of ancient Chinese character images with down,lest and right,surrounding and single structures is improved,and its MAP is up to 80.75%,which shows that the retrieval model of ancient Chinese character image based on integrating spatial and channel attention network has better performance.