Detect Text Regions (CRAFT)

How it fits

Typical backends

Caveats

• Hard constraintThe bound weights directory is scanned for `*.pth` files and the first match is used, so place a single checkpoint in the slot — a multi-file directory is ambiguous, and an empty one fails at startup. mag_ratio and canvas_size must both be greater than zero or startup aborts.
• I/O contractThe output is [Polygon<Double>], not an annotated image; pair with a visualization component for a rendered preview. Polygon mode is on, and when a curved polygon cannot be formed for a region the component falls back to that region's 4-point box. Regions with fewer than three vertices are silently dropped.
• Parameter interactioncanvas_size caps the resized long side before a 32-pixel pad — raise it for small text in high-resolution scans, lower it to bound memory and latency — while text_threshold, link_threshold, and low_text are CRAFT cutoffs that interact non-linearly and are best tuned on a validation set.
• Fallbackdevice starting with `cuda` silently falls back to CPU when CUDA is unavailable at startup; CPU inference is impractical for live video on this detector.
• State lifetimeAll configuration keys and the loaded weights are captured once at startup; runtime changes have no effect and require a redeploy.