Berthe Morisot: Zig-Zag Detection

open_access
Open Access
Author:
Singh, Aneesh
Area of Honors:
Data Sciences
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
James Z Wang, Thesis Supervisor
Rebecca Jane Passonneau, Thesis Honors Advisor
Keywords:
Computer Vision
Machine Learning
Art
Paintings
UNet
Brushstroke Detection
Abstract:
Berthe Morisot’s paintings are marked by a distinctive zig-zag brushstroke technique that highlights her individuality within the broader Impressionist movement. This thesis develops a computational approach to detect and analyze these zig-zag strokes using computer vision and machine learning. Classical image-processing methods, gradient-based feature extraction, structure tensor analysis, edge and corner detection, streamline tracing, and clustering with DBSCAN are first explored as building blocks for isolating stroke-like structures. While these methods help reveal local orientation and curvature patterns, they prove too noisy on their own to produce clean, painting-wide zig-zag maps. The final system therefore uses these handcrafted features as input to a U-Net–style segmentation network. Multi-channel tiles (RGB plus gradient, structure-tensor, and corner-based features) are extracted from digitized paintings, and the U-Net is trained on a combination of carefully hand-labeled zig-zag masks and noisier pseudo masks generated from earlier algorithms. The network outputs per-pixel zig-zag probabilities, which are stitched back together into full-painting maps and summarized with quantitative measures such as zig-zag density, stroke length, and average angularity. Qualitative results on key works by Morisot show that the model successfully highlights many of the zig-zag bands in water, foliage, and clothing that art historians associate with her style, though residual noise, missed strokes, and weaker performance on paintings by contemporaries like Monet point to limited data and some overfitting. Overall, this work demonstrates that Morisot’s zig-zag brushwork leaves measurable traces that can be recovered with modern computer vision tools, and it lays a foundation for future studies that use richer data, improved models, and broader artist comparisons to deepen the computational study of painterly style.