Deep Transfer Learning for writer identification in medieval books

Bria, Alessandro; Cilia, Nicole Dalia; Stefano, Claudio De; Fontanella, Francesco; Marrocco, Claudio; Molinara, Mario; Tortorella, Francesco

doi:10.1109/metroarchaeo43810.2018.9089780

Cited by 8 publications

(2 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…On the contrary, where classification learning generally uses supervised learning, Christlein et al [12] used ResNet DL architecture with unsupervised learning and SIFT keypoint location of databases ICDAR17 on historical documents and ICFHR16 on historical Latin script documents demonstrating promising results. This research is similar to Bria et al [13], who also experimented using different DL architectures, namely VGG19, ResNet50, InceptionV3, InceptionResNetV2, and NASNetLarge. However, instead of end-to-end, they are used as transfer learning on medieval books for paleograph writer identification, and InceptionResNetV2 demonstrates better performance among others.…”

Section: Introductionsupporting

confidence: 79%

“…Most researchers from previous studies mainly focusing developing additional methods or specific approaches, e.g. [7][8][9][10][11][12][13][14][15][16][17][18][19] with arbitrary DL selected to get the best accuracy in writer identification with TopN, SoftN, or HardN evaluation metrics performance. On the contrary, our work mainly focuses on getting a scientific explanation by comparative study without any additional method or specific approach to the handwriting writer identification problem.…”

Section: B Performance Comparison With Previous Researchmentioning

confidence: 99%

See 1 more Smart Citation

End-To-End Evaluation of Deep Learning Architectures for Off-Line Handwriting Writer Identification: A Comparative Study

Suteddy

Agustini

Adiwilaga

et al. 2023

JOIV : Int. J. Inform. Visualization

View full text Add to dashboard Cite

Identifying writers using their handwriting is particularly challenging for a machine, given that a person’s writing can serve as their distinguishing characteristic. The process of identification using handcrafted features has shown promising results, but the intra-class variability between authors still needs further development. Almost all computer vision-related tasks use Deep learning (DL) nowadays, and as a result, researchers are developing many DL architectures with their respective methods. In addition, feature extraction, usually accomplished using handcrafted algorithms, can now be automatically conducted using convolutional neural networks. With the various developments of the DL method, it is necessary to evaluate the suitable DL for the problem we are aiming at, namely the classification of writer identification. This comparative study evaluated several DL architectures such as VGG16, ResNet50, MobileNet, Xception, and EfficientNet end-to-end to examine their advantages to offline handwriting for writer identification problems with IAM and CVL databases. Each architecture compared its respective process to the training and validation metrics accuracy, demonstrating that ResNet50 DL had the highest train accuracy of 98.86%. However, Xception DL performed slightly better due to the convergence gap for validation accuracy compared to all the other architectures, which were 21.79% and 15.12% for IAM and CVL. Also, the smallest gap of convergence between training and validation accuracy for the IAM and CVL datasets were 19.13% and 16.49%, respectively. The results of these findings serve as the basis for DL architecture selection and open up overfitting problems for future work.

show abstract

Section: Introductionsupporting

confidence: 79%