Research on Location of Chinese Handwritten Signature Based on EfficientDet
Fazheng Wang, Yanwei Yu, Ding Shuiyuan, Wang Zhihao, Wang Yibo, Yuanping Song · 2021
Nowadays, most of the time we need to obtain the location information of the signature in the documents for signature verification or other tasks. Manual processing can be completed for a small amount of documents, if the amount of documents is large, it will consume huge manpower and material resources to do it .When using deep learning methods to intelligently locate signatures, but there are several difficulties: First, for documents such as Chinese contracts, files, etc. the signature is covered by seal coverage or other interference in most cases, which brings a lot of interference to the location of the signature.Second , it is hard to get real-scene contracts, bills which can be used as training sets.Third, the signature itself has fewer features, this brings certain difficulties to the intelligent location. This paper proposes an improved feature extraction network based on EfficientNet to adapt to detection targets with fewer features and more interference. Training datasets mainly comes from algorithm synthesis.Finally we tested on 1000 scanned documents containing 2048 handwritten signatures,the test results on real-scenes data set containing Chinese handwritten signatures can achieve more than 90% of AP, Precision, and Recall.