R&D
CENTER
KOENJA

Straighten first, then judge — Multi-Inspector's yolo_align mode

A new mode that runs YOLO on images that are only rotated and centered, without Canny preprocessing — reproducing training-time alignment on the line.

Multi-Inspector (EB) v2.2 adds a yolo_align mode. The existing yolo mode judged Canny-filtered, rotation-corrected images. The new mode feeds YOLO the live image with alignment only.

Why

Models are trained on images aligned by the Image Processing Tool. Line input has to be aligned the same way for training and inference to match. So the Augmentation Tool's FeatureAligner was ported as-is, and it reads the same alignment config file.

Flow

Live image (no Canny)
  └ Align + Center   ← config_align.ini [aligner]
  └ YOLO judgment    ← existing logic, unchanged

Per-mode settings

Thresholds and per-class thresholds can be set per mode number (modeN_*). The existing yolo mode keeps reading its own keys, so its behavior does not change.

modeN_type = yolo_align
modeN_threshold = 0.7
modeN_iou = 0.25
modeN_align_config = config_align.ini

Safeguards

If alignment fails (for example, no product detected), judgment continues on the original image. If the alignment module cannot load, the app still runs and only that mode falls back to the original.