Deep Neural Networks Based Offline Writer Identification

Offline writer identification is one of key researches in computer vision and pattern recognition,and it plays an important role on forensic,document protection,historical document analysis and other fields.Although offline writer identification has a long history and many researchers contributed much attention to this field,there are still some weaknesses to be overcame.For instance,it depends heavily on data augmentation and global feature encoding,the existence of common features in the learned local features,and so forth.To relax the above dependences,we investigate offline writer identification from the perspectives of robustness,generalization and practicality.Firstly,a robust offline writer identification(DLS-CNN)is proposed.It bases on statistical line-segmentation and deep convolutional neural network.DLS-CNN segments handwriting documents into image patches using statistical document line-segmentation method and sliding window approach at first,then an optimized residual neural network is used for feature learning,and finally the mean of all local features of each handwriting document is encoded into a global feature.Experimental results on the two well-known benchmark datasets(ICDAR2013 and CVL)show that,due to the learned robust local features,DLS-CNN achieves higher identification rate with fewer handwriting content and better retrieval results without data augmentation and good global encoding methods,compared with other popular offline writer identification.Secondly,a semi-supervised feature learning method is proposed to improve writer identification.This method bases on weighted label smoothing regularization(WLSR).It directs neural network to make an incorrect prediction towards extra unlabeled data,thus the classifier will be penalized and the baseline will be regularized effectively on the one hand.On the other hand,it significantly filters common features that are learned from baseline and reduces the adverse effect of common features,thus it improves the discriminability of features learned from neural network.The experimental results confirm the efficacy of the proposed method.To the best of our knowledge,this is a pioneering work using semi-supervised feature learning to automatically learn discriminative features to improve writer identification,and provides new insights into offline writer identification.