Research On Object Detection Model Of Ancient Character Rubbings

Ancient characters are important materials for historical research.Generally,they are attached to inscriptions or steles,which are processed into rubbings.The first step in the study of ancient characters is to label,which is time-consuming and laborious.This paper introduces Deep Learning into this process to realize the automatic detection of ancient characters.This paper mainly studies the detection of ancient characters on rubbings of Bronze Inscriptions and rubbings of stone inscription of Wei,Jin and Southern and Northern Dynasties.Based on the analysis of rubbings data set of ancient characters,a series of problems such as complex rubbing background,irregular characters distribution,dense small characters,and overlapping detection are found in the detection process.Different from modern Character Detection,this paper regards Ancient Character Detection as Object Detection.A Novel Detection Network for the Detection of Ancient Characters on Rubbings Combined with the characteristics of rubbings data set of ancient characters,the improved feature extraction network based on the Dark Net-53 has stronger feature extraction capabilities.Based on the existing Bi FPN(Bi-Directional Feature Pyramid Network)and PANet(Path Aggregation Network),dual-FPN(dual Feature Pyramid Network),a repeated feature pyramid network,is proposed to integrate the feature maps of different scales in the model more effectively.Aiming at the overlapping detection problem in the rubbings detection of ancient characters,ANMS(Ancient Non-Maximum Suppression)is proposed,which effectively solves the overlapping box problem in the rubbing detection results.At the same time,a novel Non-Maximum Suppression Loss Term,ANMS-Loss(Ancient Non-Maximum Suppression Loss),is proposed to increase the position loss of the target box and the prediction box,which is added into the Loss Function of the model to make the Gradient Descent direction more in line with expectations.The m AP@0.5 : 0.95 indicator of the detection network is up to 62.7%,the m AP@0.5 indicator is up to 96.1%,and the FPS is 63.Speed-Oriented Feature Extraction Network This paper designs feature extraction network from the following three aspects.A-Dark Net(Ancient Dark Net)is proposed,which simplify the feature extraction network structure.In the detection network,the m AP@0.5 : 0.95 indicator is 62.7%,and the FPS is up to 278.Based on the simplified feature extraction network,TR-Dark Net(TRansformer Dark Net)with Transformer encoder structure is proposed.The m AP@0.5 : 0.95 indicator of the detection network is up to 62.9%,and the FPS is 256.Combining the above two feature extraction networks,dual-Dark Net(dual Dark Net),a repeated feature extraction network,is proposed.The m AP@0.5 : 0.95 indicator of the detection network is 62.9%,and the FPS is 278.