Text Independent Offline Hand Writer Recognition Using Machine Learning

Handwriting is a behavioural biometric that an individual learns and develops over time and automated writer identification systems can be developed by identifying these behavioural aspects of an individual’s writing style. These writer recognition systems greatly assist forensic experts by facilitating them with semi-automated tools that segment the text, narrow down the search, help with visualization and finally assist in the final identification of an unknown handwritten sample. Handwriting, as a behavioural characteristic, has been a subject of interest for researchers for many decades and intensive research performed in this field has resulted in the development of multiple methods and algorithms. However, automated writer identification is still a challenging problem. Difficulties in segmenting text and the deviation of an individual from his or her unique writing style is the reason for ongoing research in this field. This thesis aims to investigate the problems faced in automated writer identification and propose novel techniques of segmentation and classification that would contribute to the field of writer identification. This has led to four different contributions. First a novel segmentation algorithm is proposed for segmenting sub-words within hand written Arabic words, the proposed method outperforms the previously used projection profile method. The second proposed method offers a segmentation free multi-scale Local Ternary Pattern Histogram for text independent writer identification. Local ternary patterns are applied at various scales to produce a predictor model for its respective scale while the high dimensionality problem of a multi-scale approach has been tackled with dimensionality reduction using SR-KDA. The third contribution tackles the problem of writer identification in noisy conditions. A robust offline text independent writer identification system is proposed using Bagged Discrete Cosine Transforms. The proposed system effectively utilizes discrete cosine transform for writer identification while avoiding problems of high dimensionality and memory limitations. Finally, in the fourth contribution a dissimilarity Gaussian mixture model is proposed for describing the contrast between different writers of a dataset. Furthermore, a weighted histogram approach is also proposed that penalizes bad prediction scores with a cost function to significantly enhance the identification rate.