A Hair Drawing Evaluation Algorithm for Exactness Assessment Method in Portrait Drawing Learning Assistant System

Yue Zhang, Nobuo Funabiki, Erita Cicilia Febrianti, Amang Sudarsono, Chen-Chien HsuAlgorithms 2025 · DOI

TL;DR

This paper introduces a novel algorithm to evaluate the accuracy of hair drawings within a portrait learning assistant system (PDLAS). By analyzing hair texture using eigenvalues and eigenvectors, the system provides more comprehensive feedback to novice artists, improving the overall assessment of portrait drawings.

What Problem Does This Paper Address?

Portrait drawing is a popular hobby, but beginners often struggle with the foundational aspects of facial proportions and structure. Existing learning assistant systems, like PDLAS, aim to help by providing auxiliary lines for facial features. However, a significant limitation of these systems is the lack of a robust method to evaluate the accuracy of hair drawings, which constitute a large and crucial part of a portrait. Current assessment methods in PDLAS focus on other facial features but overlook hair, leaving a gap in providing complete feedback for artistic improvement.

Methods

The core of this paper is the development of a hair drawing evaluation algorithm that extends the existing Normalized Cross-Correlation (NCC) method used in PDLAS. The proposed approach involves several key steps:

  • Bounding Box Definition: A specific bounding box for the hair region is defined using keypoints extracted by the OpenPose library.
  • Texture Feature Extraction: To capture the intricate details of hair, the algorithm extracts texture features by computing eigenvalues and eigenvectors of the hair region. This is a novel step compared to the previous NCC application which relied solely on pixel values.
  • Grayscale Conversion and Edge Detection: The input image is converted to grayscale, and the Canny edge detection algorithm is applied to identify hair edges.
  • Morphological Processing: Dilation and erosion operations are used to refine the edge detection results.
  • Hair Region Isolation: The grayscale image is combined with a mask to isolate the hair region.
  • NCC Score Calculation: The NCC score is then calculated between the texture features of the original image and the user's drawing. This allows for a more nuanced comparison that accounts for the complexity of hair texture.

Datasets

The study utilized drawing results from 12 students at Okayama University, Japan, who used the PDLAS system. The face photos used as references were downloaded from a public source ([34]). No new datasets were introduced.

Key Insights and Ideas

  • Hair Texture is Crucial: The paper highlights that hair texture is a complex and challenging aspect of portraiture, significantly impacting the overall impression of a drawing.
  • Eigenvalue/Eigenvector Analysis for Texture: Utilizing eigenvalues and eigenvectors for texture feature extraction is a key innovation, allowing for a more detailed and robust comparison of hair patterns than simple pixel-wise correlation.
  • Improved Assessment for Novices: By specifically addressing hair evaluation, PDLAS can offer more comprehensive and actionable feedback, aiding novices in developing their drawing skills more effectively.
  • Iterative Improvement: The study demonstrates that by focusing on low-scoring facial parts (including hair) and re-evaluating, users can achieve significant improvements in their drawings.

Key Results

The proposed hair drawing evaluation algorithm was applied to the drawings of 12 students. The results showed that incorporating texture analysis led to hair NCC scores that were more comparable to the scores of other facial features (eyes, nose, mouth, eyebrows). For instance, the average NCC score for hair with texture analysis was 0.47, compared to 0.24 without it.

Furthermore, an iterative evaluation with three users demonstrated improvements in their drawings after modifying parts with low NCC scores. For User 1, the average NCC score improved by 0.04; for User 3, by 0.06; and for User 7, by 0.02. These improvements were often attributed to better representation of details like boundary lines, shading, and volume.

Limitations and Future Work

The authors acknowledge several limitations:

  • Lack of Specific Guidance: The current system only provides an NCC score for each face part, but lacks specific guidance or suggestions on how to improve a drawing.
  • Limited Subject Diversity: The experiments were conducted with university students, and a broader range of subjects (different ages, genders, professions, and countries) is needed for more generalizable results.

Future work could involve:

  • Developing explicit improvement suggestions: Moving beyond just scores to offer concrete advice.
  • Expanding subject demographics: To ensure the system's effectiveness across a wider user base.
  • Adapting for Cartoon/Anime Styles: OpenPose is not suitable for cartoon characters due to their distinct facial proportions. Future research could explore deep learning-based landmark detection tailored for these styles.

Summary updated

BibTeX
@article{zhang2025hair,
  author = {Yue Zhang and Nobuo Funabiki and Erita Cicilia Febrianti and Amang Sudarsono and Chen-Chien Hsu},
  title = {{A Hair Drawing Evaluation Algorithm for Exactness Assessment Method in Portrait Drawing Learning Assistant System}},
  journal = {Algorithms},
  year = {2025},
  doi = {10.3390/a18030143},
}
Figure