Executive Summary : | Handwriting has been a cornerstone of human civilization. Even in the digital age, handwriting still plays a vital role in daily learning and communication and is an integral part of many activities (e.g., academic, offices, courts). Handwriting is considered solid behavioral biometrics and forensics evidence in many places, counting India. Inspecting a writer from its handwriting is challenging due to the inter-variation and intra-variation of writing strokes. Here, inter-variation denotes the variation of writing between two writers, and intra-variation represents the variation of writings of the same writer. Although research on inter-variation writer inspection can be seen easily, work on intra-variation is hardly found. However, such intra-variation in handwriting can be easily found due to mechanical (e.g., writing instrument, surface), physical (e.g., illness, aging), psychological (e.g., excitement, anger, mood) factors. Moreover, in multilingual countries like India, where people can write/read more than one language, the problem of writer inspection is more challenging due to intra-variation across the script. This project aims to facilitate intra-variation writer inspection to empower handwriting biometrics. More precisely, we focus on: 1) Intra-variation in the same script: The handwriting may change with the writing surface due to friction between the writing medium (paper) and tool (pen). The handwriting also changes with time and the person's age, or mood. Some diseases (e.g., Parkinson’s, Alzheimer’s) affect an individual's handwriting. In this project, we first aim to focus on the intra-variation of a person on the same script. We plan to use deep reinforcement learning-based strategy to find idiosyncratic patches for a writer, followed by transformer-based architecture to identify the writer from the patches. 2) Intra-variation across multilingual scripts: In multilingual countries like India, people can write multiple scripts. Therefore, script-independent writer inspection is required when a system is trained and tested on two different scripts. The past methods primarily focused on writer inspection from training/testing on the same script. Here, we plan to understand the writer while training on one script and testing on another. Here, we aim to extract classical features (e.g., margin space width, text-line skewness, word slant) and convolutional neural network (CNN)-based deep features to understand the intra-variation across the script. 3) Individuality of certain characters: Some characters of a script are more revealing than other characters. We intend to find a small set of characters that carry enough information to inspect a writer. This will help to identify a writer only by looking at the small set of characters. We plan to work with an Indic script. We will extract features from a CNN and use an attention mechanism to understand where/how much to focus on the character strokes to inspect the individuality. |