Research About Bronze Inscription Image Retrieval Based-on Deep Learning

Bronze inscriptions are inscriptions on bronze wares.They record information about the social forms and cultural life of the Shang and Zhou dynasties.For researchers of ancient characters and history,the bronze inscriptions have high research value.However,when interpreting the inscriptions according to the existing method,the inscription researcher needs to manually query the inscription data,and the interpretation of inscriptions is inefficient and error-prone.The dataset of bronze inscriptions studied in this thesis comes from the rubbings of bronze inscriptions,and the images often have problems such as lots of noise and complex background.At the same time,the distribution of samples of each category in the bronze inscription dataset is uneven,and there is a long-tail problem.Aiming at the deficiencies in the existing inscription retrieval methods and the problems in the inscription sample collection,this thesis conducted the following research:(1)Classification-based image retrieval method for inscriptions: Use the transfer training method to train the model to solve the problem of fewer samples of inscriptions.Aiming at the long tail problem in the inscription sample set,a multi-class Focal Loss is defined.When searching,use the trained model to classify the input inscription images first,and then search within the class.(2)The inscription image retrieval method based on feature fusion: through the visual analysis of the inscription image features output by each convolutional layer,it is found that the features output by the shallow convolutional layer contain a lot of details of the inscription image.The features output by deep convolutional layer contain rich semantic information.According to the characteristics of the output features from different convolutional layers,the feature fusion algorithm is researched.(3)Bronze inscription image retrieval method based on text-area detection and feature fusion: To solve the problem of complex background and noise of the inscription sample,the input inscription image should be detected by text-area before feature extraction to remove complex image background and some noise.Then,the convolutional neural network is used to extract the features of the bronze inscription text-area in the image for fusion.To improve the quality of extracting bronze inscription features,the Res Net-50 network structure is improved,and the attention module CBAM is added to the network.During training,the features output by the Contrastive Loss training network are used to expand the inter-class distance between features of different categories and reduce the intra-class distance between features of the same category.Experiments show that the optimal model has achieved better retrieval accuracy on the bronze inscription dataset,and can solve the actual problems in the retrieval of bronze inscriptions.(4)Bronze Inscription image retrieval system: To facilitate the use of users,this thesis designed and implemented the bronze inscription image retrieval system.The system consists of three core parts: bronze inscription database,server and client.Experiments have proved that:(1)The method based on classification has better retrieval results when the recognition accuracy of bronze inscription image is high;(2)The method based on feature fusion enhances the diversity of information in the features,and is more robust than the method based on classification;(3)The method based on text-region detection and feature fusion is an improvement on the feature fusion method.This method removes the complex background of bronze inscription images and achieved the highest retrieval accuracy.