A new approach for feature extraction with applications to Automatic Writer Recognition

Automatic Writer Recognition based on scanned images of handwriting is a behavioral biometric method which has applications in forensic and historical document analysis. In this paper, an efficient method for feature extraction from handwritten images is presented. In our proposed method, the normal vectors of the outer contour points of each connected-component are calculated and the sequence of obtained normal vectors is encoded to be rotation invariant and scale invariant. Also, two weighted histograms are designed to generate the feature vector of the input image by putting together the probability mass functions obtained using these histograms. A dataset consisted of 100 people's Persian handwriting have been gathered to evaluate the proposed method. The experimental results are satisfactory and the accuracy of the proposed method is 97% on our dataset.