Study on Recognition of Xixia Text Based on Improved DenseNet
Due to a large number of strokes,complex structure,high similarity,and the problems of missing characters,foxing,and fading in the ancient books of Xixia,it is still a difficult research to detect and recognize them at present,and the existing recognition studies mostly have problems such as suboptimal recognition accuracy,omission,misdiagnosis.Therefore,we propose an improved DenseNet-based Xixia text recognition method based on a comprehensive analysis of the current mainstream research.The proposed method replaces the traditional 3×3 convolution in the original model by introducing the spatial and channel reconstruction convolution,which mainly utilizes the channel reconstruction module and the spatial reconstruction module to reduce the redundancy between the feature maps in the training process of the network,and improves the feature representation capability of the network.Furthermore,it uses the mutual-channel loss instead of the cross-entropy loss in the loss function part,which further reduces the feature redundancy and improves the ability of the network to focus on the key recognition regions without introducing any external parameters.The results of the comparison experiments show that the accuracy of the proposed method is 97.08%and the parameters are 6.2 MB on 668 types of Xixia text recognition datasets,which is a more obvious improvement relative to the current mainstream methods,proving its effectiveness.
ancient books of Xixiatext recognitionchannel reconstructionspatial reconstructionmutual-channel loss
国家科技期刊平台
NSTL
万方数据
