Generative Line Drawings

The bulk of computer vision research conceptualizes the image exclusively in spatial dimensions. But in the history of drawing, time has also been a significant of a factor in how we recorded the world. A linear approach towards drawing is particularly prevalent in ancient Chinese art, where the shape, motion and direction of the *line* dominated the aesthetic. The unity and harmony of Eastern art, with natural, flowing shapes, ties into the focus on the line.  It birthed a concentration on form and connection. Therefore, I was interested in generative artwork derived from line drawings rather than pixel-based images.

I produced a linearized survey of the Metropolitan Museum’s collection of early sculpture figurines from a variety of cultures, particularly early East and South Asian societies. I focused on the depiction of women.

These humanoid forms generated by a machine mirror early culture’s attempts to record the human body, particularly the female form. The drawings were made using a sequence of (1) background removal (2) edge detection (3) image vectorization (see post here) (4) centerline tracing and (5) drawing style application, on images scraped based on the Met’s public domaine collection.

In the image vectorization step, I applied the Ramer-Douglas-Peucker algorithm for line simplification, which is a method for reducing the number of points in a line while maintaining general shape. Adjusting this simplification parameter produces vectorized results of varying abstraction:

The underlying dataset was collected from nearly 4,000 of the Metropolitan’s Public Domain artworks depicting figurines, focusing on statue figures from 8000-2000 B.C.

The resulting images provide an abstracted lens onto human figure drawing across thousands of years. The machine’s perspective on the human form is linearized, offering a reduction of the body’s shape into closed-form strokes. It is a computerized version of the process through which Kanji was developed — morphing images into lines into symbols.

Adjusting various parameters, I achieved varying linear reproductions of the statuettes, ranging from simplified to more complex.

Simple: 10-75 points (to define the total objects)

Moderate: 75-150 points

Moderate: 75-150 points with different edge detection method

Complex: 150-300 points with varying edge detection methods

In order to achieve the moderate silhouette form (third example), it was crucial to apply a clean background removal technique, then to apply centerline tracing in the vectorization of the image. In the future, this technique may be applied to machine generate abstract sculptures or line-based drawings.