BIION-D Docs

Case 0: Plastic Discs — YOLO Counting & Face Detection

Training Exercise 0 — YOLO face detection and counting of industrial plastic discs.

Scenario Overview

Plastic discs arrive at an inspection station with either face visible to the camera. Use a YOLO Object Detection model to locate every visible disc, identify its face (Face A or Face B), and count the discs for each face and in total. This exercise focuses on detection and counting, not surface-defect inspection.

Assignment Tasks

  1. Camera and lighting: Set up a camera view and uniform illumination so that all discs and their distinguishing face features are visible.
  2. YOLO dataset and training: Collect and annotate examples of both faces, including realistic changes in position and orientation. Train a YOLO Object Detection model with classes disc_face_a and disc_face_b.
  3. Action Plan and pipeline: Configure image capture followed by an Object Detection module that reads the camera image from a previous block.
  4. Counting logic: Use a Plugin to count detections of each class and export count_a, count_b, and total_count to the Data Registry.
  5. Verification: Run images with different numbers and orientations of discs in Inference mode and compare the reported counts with the visible pieces.

System Specifications & Constraints

Parameter Specification
Inspection scope Detect and count visible discs by face; no anomaly detection.
Available camera 1 camera viewing the discs.
Available lighting Lighting sufficient to distinguish Face A from Face B.
Model type YOLO Object Detection with classes disc_face_a and disc_face_b.
Plugin output count_a, count_b, and total_count (count_a + count_b).

Tips & Practical Guidance

Include examples of both faces at different positions and rotations, with consistent class labels. Check that glare does not hide the features used to distinguish the faces.

Object Detection returns classes and bounding boxes but does not automatically set pipeline compliance. Use the detections in the Plugin for counting; add a separate rule only if a pass/fail decision is required.

Acceptance Criteria

  • [ ] YOLO detects the visible discs and assigns each detection to the correct face class.
  • [ ] The Plugin reports the correct count_a, count_b, and total_count for test images with different disc arrangements.
  • [ ] The camera image is available to Object Detection from a previous pipeline block.
  • [ ] No anomaly model, dashboard, or PLC integration is required in this case.