Lee, P.-X.P.-X.LeeDing, J.-J.J.-J.DingWang, T.-C.T.-C.WangLee, Y.-C.Y.-C.LeeJIAN-JIUN DING2020-06-042020-06-042018https://scholars.lib.ntu.edu.tw/handle/123456789/497042Writer verification is to identify whether the script was written by a person himself. However, although there are many advanced machine learning methods now, automatic writer verification is still a very challenging work since the training data in the false case (forged case) is usually very hard to acquire. In practice, to avoid being convicted, a criminal may write scripts that are fully different from what he wrote on the forged document. In this manuscript, we adopt global and semi-global information, including log-Gabor features, advanced moments features, and co-occurrence matrix features for writer verification. In addition, a more flexible classifier is proposed. Though the convolutional neural network is popular, its performance is limited when the training data is not enough. Therefore, another classifier based on the weighted squared Euclidean distance is adopted. Simulations show that the proposed algorithm outperforms other methods and will be very helpful for identifying forged scripts. © 2018 IEEE.end-point information; feature extraction; forensic image processing; Handwriting; pattern recognition[SDGs]SDG16Character recognition; Feature extraction; Image processing; Neural networks; Pattern recognition; Adaptive classifiers; Co-occurrence-matrix; Convolutional neural network; End points; Handwriting; Machine learning methods; Squared euclidean distance; Writer verification; Learning systemsconference paper10.1109/ICMEW.2018.85515662-s2.0-85059982759https://www.scopus.com/inward/record.uri?eid=2-s2.0-85059982759&doi=10.1109%2fICMEW.2018.8551566&partnerID=40&md5=30e79fcae47c38518dd9886d7bd308de