Handwritten Text Recognition Using Linear Clustering

Abstract

Handwritten Text Recognition (HTR) presents a persistent challenge in pattern recognition and machine learning owing to inherent variability in individual writing styles, character distortions, and noise present in scanned or captured documents. This paper proposes a lightweight and interpretable HTR system that avoids dependency on computationally expensive deep learning architectures such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). The proposed methodology follows a systematic pipeline comprising input image acquisition, preprocessing, segmentation, handcrafted feature extraction, linear clustering, classification, post-processing, and final output generation. During preprocessing, noise removal, normalization, and binarization are applied to enhance image quality, followed by segmentation into individual characters or words. Meaningful features—including Histogram of Oriented Gradients (HOG), zoning features, pixel intensity distributions, and geometric attributes—are extracted and mapped into a feature space. Linear clustering techniques, particularly K-Means clustering, then group similar handwriting patterns to reduce intra-class variation and improve feature separability. A classical machine learning classifier—Support Vector Machine (SVM) or Logistic Regression—subsequently assigns labels to the clustered feature representations. Experimental evaluation demonstrates competitive recognition performance with significantly reduced computational overhead, making the system particularly suitable for resource-constrained environments and applications requiring transparent, explainable decision-making.