A NOVEL MULTICLASS-MULTISTAGE WRITER VERIFICATION USING HYBRID APPROACH IN SPATIAL AND TRANSFORM DOMAIN.

摘要

Writer verification system compares the two test samples and gives the decision about whether the two samples are written by the same or a different person. The decision is based on the threshold value of the claimed writer. It is observed that the style of a writer and the type of writing instrument used, greatly affects the handwriting. The ink width variation in the handwritten document image depends on the writing instrument, pen pressure and pen control. It also decides the ink distribution in the handwritten document image. Ink distribution causes variation in handwriting and it results into high error rates. It is difficult to get high accuracy of writer identification and verification under different ink width conditions. The existing off-line English handwriting databases of writer identification and verification ignored the ink width variations for same writer samples. Mostly, writer identification is dealt with, which find likely list of Top 1 and Top 10 writers from the known dataset. Very less work is reported in writer verification. Researchers evaluated the performance of their methods on smaller dataset(s) except Srihari (used Centre of Excellence in Document Analysis and Recognition, CEDAR database of 1000 writers) and Bulacu (used IAM database of 650 writers). Researchers used experimental datasets containing only 2-4 samples of each writer. In such case, it is not possible to achieve False Acceptance Rate (FAR) equal to False Rejection Rate (FRR) and Equal Error Rate (EER) cannot be used as a performance parameter for writer verification. It is found that EER method of threshold calculation using only 2/3 samples in the dataset is inappropriate. Only one sample was used for training and testing of writer identification and verification. Most of the work is reported in Chinese and Arabic language. This motivated us to design the off-line text-independent writer verification system in English handwritten document under different ink width conditions using new features/methods which will provide low false acceptance error rate and less response time on largest dataset. The existing dataset samples do not contain ink distribution for same writer samples. Therefore, a new dataset is created which contains 2 samples using ball pen and 2 samples using sketch pen of each 1000 writers. The handwriting samples were collected from males and females of different age groups, different professions and belonging to different places. The handwriting samples were scanned using professional hp scanner as 8 bit colour JPEG image with 300 dpi resolution. The newly created database is named as SLK Database. It is made publically available and freely downloadable from homepage www.sharadakore.com. In this research work, the writer verification methods used slant, pixel distribution in different directions and Entropy feature in the handwritten document image. Features are extracted based on contour using chain code method. Mmultistage writer verification improved accuracy further, using combination of chain code based features. To reduce verification time (T) and improve accuracy further, classification is done before multistage writer verification. Writer classification is done based on directional ink distribution information. It is calculated using Entropy of wavelet decomposed sub-band images. Experimental result has shown that the classification prior to multistage writer verification gives FAR of 1.92%, verification accuracy of 91.49 % and takes verification time of 21.672000secs for 1000 writers in the SLK dataset 2. The presented methods are text and ink width independent. A generalized model which works under different text and ink width is presented for writer verification. This automatic writer verification system can be used to verify person based on handwriting. It is useful in determining the authentication in case of handwritten documents such as Patents, Will deed, personal notes, examination answer sheets, death threats, suicide notes, ransom notes etc.