Writer identification problem is the problem of assigning a known writer to an unknown text script from a group of possible writers. This is useful in many areas of interest such as forensics, security, historical documents analysis and many more. Writer identification has seen a plethora of research and attention from researchers in recent years, however, it still remains a difficult problem to solve due to variations in writing style and conditions, and the problems arising from variations in image quality due to document degradation and other incidental factors.
The pipeline, shown in the figure below, shows the main modules that the data goes through before making a decision: Preprocessing, Feature Extraction and Classification.
First, each image is processed to retrieve the handwritten paragraph only to reduce the overall computational complexity and remove other irrelevant information from the image so that only the handwritten part affects the decision making process.
In feature extraction module, the implementation was based on the methodology in [1] as their feature vector was based on Local Binary Pattern (LBP) which is used in many computer vision problems due to its robustness against noise. Authors in [1] introduced their own variation of LBP, Sparse Radial Sampling Local Binary Pattern (SRS-LBP), which allows for sampling of the circular patterns of traditional LBP at a lower computational cost. In addition, the comparison operator of the traditional LBP is changed to a threshold that is derived statistically from each image. Then the feature vector is formed as the histogram of each SRS-LBP at different radii, concatenated together after removing the 0 pattern which corresponds to foreground and background only patterns, thus providing no information.
Then the features are normalized using L2 normalization before applying a Nearest Neighbour classification to produce the resulting label.
We analyzed the performance of our algorithm based on IAM dataset. We worked with complete form text only.
Each iteration, three different randomwriters are chosen from the set. For each writer, two random pages are chosen for training and the others are used for testing.
Further info can be found here.
[1] A. Nicolaou, A. D. Bagdanov, M. Liwicki, and D. Karatzas, “Sparse radial sampling lbp for writer identification” ICDAR, 2015.