AI Segmentation

Micro-Seg integrates AI-assisted segmentation through its plugin system. All inference runs locally via ONNX Runtime by default — your images never leave your machine. (Teams with a shared inference server can configure a remote backend instead.)

Interactive mode — `Ctrl`+`Shift`+`A`

Activate the AI tool from the toolbar or by shortcut, then guide the model with clicks:

Left-click places a positive point: “this is part of the object.”
Right-click places a negative point: “this is not.”
Each click updates the proposed mask; alternate positive and negative points to grow and trim the proposal until it matches the structure.
Hold Ctrl while clicking to place several points without triggering a prediction after each one.
Esc cancels the interaction.

The accepted result becomes a regular annotation you can refine with the normal editing tools. Consecutive interactions on the same object reuse the previous mask as a hint, so proposals improve iteratively.

Automatic mode

The auto mode scans the whole image from a grid of seed points (8×8 by default, expanded into multiscale passes) and materializes every accepted candidate as a regular annotation. Progress is reported live, e.g. Scanning seeds 126/1344 (accepted 12), and results appear incrementally — you don’t wait for the whole scan to finish before seeing objects.

Existing annotations are respected: seeds that land inside already-accepted objects probe for sub-structures instead of duplicating them.

Models are plugins

AI models are packaged as task plugins, so new models can be added without rebuilding the app. Micro-Seg ships with built-in HRNet-OCR-based plugins for interactive and automatic segmentation.

Which plugin handles Interactive and Auto annotation is configured per project, so different projects can use different models. The model catalog (local ONNX models, thresholds, optional remote backends) lives in the configuration file.

Performance notes

The first run on an image triggers a warm-up (model load + dummy inference); subsequent runs on the same image reuse the prepared runtime.
Long auto scans are chunked into bounded sessions, which keeps memory flat and makes cancellation responsive.