Curve Fitting-Based Feature Extraction for Offline Handwritten Word Recognition

Offline handwritten word recognition for Indic scripts such as Gurmukhi remains challenging due to complex structural variations and writing styles. Although deep learning models achieve high accuracy, their reliance on large datasets and significant computational resources limits deployment in resource-constrained environments. This paper presents a lightweight, segmentation-free framework based on zone-wise curve-fitting feature extraction for offline handwritten Gurmukhi word recognition. The proposed method applies parabolic and power curve models to horizontally partitioned word images, capturing geometric and structural characteristics without explicit character segmentation. The extracted features were evaluated using Decision Tree, Multilayer Perceptron (MLP), and Random Forest classifiers on a benchmark dataset of 10,000 handwritten Gurmukhi city names. Random Forest achieved the highest accuracy of 85.57% using power curve features under a 70/30 train–test split and demonstrated stable performance in 5-fold cross-validation. A sensitivity analysis confirmed that dividing images into eight horizontal zones provides an optimal balance between discriminative representation and computational efficiency. To further enhance performance, a hybrid model combining curve-fitting features with lightweight CNN-based deep features achieved an improved accuracy of 91.24%. With feature extraction time below 0.01 s per image, the proposed approach offers an effective trade-off between accuracy, interpretability, and computational cost for real-world deployment.