The work here outlines a regression-based approach for assessing handwriting quality at the word level based on structural, perceptual, and fringe features. The work is based on a diversified corpus of more than 1296 unruled pages of 65 + writers and 1160 ruled samples from various sessions. Random Forest and XGBoost were the best-performing models among various models tried, with an excellent R2 value of 0.996. An active learning approach enhanced model training through the selection of uncertain samples, which performed better than random sampling. Perceptual attributes like neatness and readability were shown to have the most influence, and structural and marginal attributes were also factors. It should be noted that it also properly breaks down handwriting and spots tendencies with regards to low-quality manuscripts. A Diverse and comprehensive data set contains more than 1296 un-ruled pages from 65 + writers, and 1160 ruled samples obtained from multiple sessions. With high predictive accuracy, there is outstanding performance with R2 = 0.996 on the entire feature set for Random Forest regression, indicating a very strong relationship with manually assigned quality values. Efficient model refinement with an active learning technique speeds up convergence and performs better compared to random sampling.
Koushik et al. (Tue,) studied this question.