1. Dolega, B., Agam, G., & Argamon, S. (2008). Stroke frequency descriptors for handwriting-based writer identification (B. A. Yanikoglu & K. Berkner, Eds.; pp. 68150I-68150I – 8). https://doi.org/10.1117/12.767227
1.1. STRENGTHS
1.1.1. Stroke-frequency features are relatively lightweight,making them computationally efficient for real-world applications.
1.1.2. Feasibility of using these descriptors through experiments on benchmark handwriting datasets
1.1.3. Captures frequency-domain characteristics like changes in writing speed or pen pressure
1.1.4. Unique way of capturing handwriting individuality, moving beyond traditional geometric or texture-based methods.
1.2. LIMITATIONS
1.2.1. The method has been surpassed by deep learning approaches that achieve higher accuracy.
1.2.2. Mainly focuses on English Handwriting
1.2.3. Noise, pen lifts, or scanning errors can affect results.
1.2.4. Cannot handle large datasets limited only for small handwriting datasets
2. Okawa, M. (2016). Offline signature verification based on bag-of-visual-words model using KAZE features and weighting schemes. 2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 121–128. IEEE. https://doi.org/10.1109/CVPRW.2016.24
2.1. Strengths
2.1.1. Patch-based - focuses on local handwriting strokes instead of whole images.
2.1.2. Robust - KAZE descriptors are resistant to scale, rotation, slant, and pressure variations.
2.1.3. Explainable - BoVW histograms and tf–idf weighting make the process interpretable.
2.1.4. No CNN required - avoids large datasets, retraining, and black-box models.
2.1.5. Generalizable approach - method can be applied to other handwriting or document analysis tasks.
2.2. Limitations
2.2.1. BoVW ignores spatial layout - stroke arrangement is lost when converting to histograms.
2.2.2. Parameter sensitivity - results depend heavily on vocabulary size, cluster count, and weighting scheme.
2.2.3. Handcrafted features only - lacks the adaptability of modern feature learning methods.
2.2.4. Computational overhead - KAZE extraction and BoVW encoding are resource-intensive.
2.2.5. Forgery challenge → skilled forgeries may still produce similar visual-word distributions, causing errors.
3. Hangrage, M., & Veershetty, V. (2019) Word Spotting in Handwritten Document Images based on Multiple Features.
3.1. Strengths
3.1.1. Multi-Feature Descriptors. This paper uses Gabor, HOG, LBP, morphological, and texture filters which captures local texture, edges and shape features
3.1.2. Patch-based analysis: Each word image is divided into patches which preserves local detail
3.1.3. Its tested on multiple datasets: George Washington, Kannada Handwritten, Camera-captured, Heterogeneous (Mixed of printed & handwritten )
3.1.4. Retrieval Acuracy is 97.6% mAP with texture + SVM on GW dataset
3.2. Limitations
3.2.1. Performance dropped significantly 30% lower on camera-captured documents due to noise, lighting , device issues.
3.2.2. Its focused mainly on English and Kannada. Generalizability to other languages not tested.
3.2.3. Relies on cosine distance and classification; does not define a universal similairty metric.
4. Sowmya, T. B., & Malathesh, S. H. (2015). Implementation of offline text-independent writer identification using SIFT and partial structure model. International Journal of Combined Research & Development (IJCRD), 4(3), Article 53045.
4.1. STRENGHTS
4.1.1. Patch-Based Descriptor Extraction Handwriting images are divided into patches; SIFT descriptors extracted from each patch to preserve local detail.
4.1.2. Hybrid Feature Strategy Combines SIFT descriptors with histograms orientation and partial structural modeling to capture both texture and stroke direction.
4.1.3. Text-Independent Identification Works without relying on specific words or content—focuses purely on visual handwriting traits.
4.1.4. Simple Implementation Uses traditional CV techniques—no CNNs or deep learning required.
4.1.5. Local + Global Feature Fusion Integrates fine-grained patch-level features with broader structural cues, improving writer differentiation.
4.2. LIMITATIONS
4.2.1. No Probabilistic Scoring Relies on structural similarity and classification; does not use forensic likelihood ratios or statistical confidence.
4.2.2. No Spatial Arrangement Modeling Patch extraction does not explicitly encode spatial relationships between strokes or characters.
4.2.3. Vague Performance Metrics Claims high accuracy but lacks detailed reporting (no precision, recall, or benchmark comparisons).
4.2.4. Structural Model Complexity Partial structure modeling adds interpretability challenges and may reduce reproducibility.