Analyze an image
This first version does not claim to “recognize beauty.” It creates a visible, debuggable diagram of the image: regions, edges, candidate centers, scale levels, and positive-space proxies. It is meant to be wrong in interesting ways, so the next phase can calibrate it against human judgment.
What is actually stored?
The app stores a downscaled JPEG, the computed metrics, SVG diagnostic diagrams, and the generated progress-report HTML in Datastore. That keeps the first deployment inside the three-tier architecture you asked for.
What is actually computed?
The server computes Otsu thresholding, connected components, edge density, quasi-convexity of foreground and background regions, a component-size scale ladder, local symmetry proxies, and center candidates.
Why this phase?
Alexander says positive space is not leftover background: both figure and ground should be well-shaped. So the system must look at both object regions and the spaces between them.
