YOLOv8 Screw Detection Performance variation with the Composition of Real and Synthetic Training Data
Deep learning object detectors facilitate automated inspection but require labelled images, making synthetic training data attractive. This study evaluates how real and synthetic image quantities affect a YOLOv8s detector for M3 ×8 and M3 ×12 screws. Eight valid combinations of 0%, 20% and 50% real data and 0%, 50% and 100% synthetic data were evaluated using two random seeds anda fixed 30% real test set. Preliminary hyperparameter selection identified SGD, an initial learning rate of 1 ×10−3 and a batch size of 32. The highest two-seed mean overall detection-classification score was 83.74% with 50% real and 100% synthetic training data. Synthetic-only training yielded substantially lower scores, indicating a synthetic-to-real gap. At 50% real data, additional syntheticimages slightly increased the mean score; at 20% real data, increasing synthetic data reduced it. In a second stage, Fuzzy C-Means (FCM) initialized an Adaptive Neuro-Fuzzy Inference System (ANFIS) using seven features from matched YOLO detections. The three-rule hybrid achieved 82.35% matched-detection classification accuracy and an F1 score of 80.85%, compared with 80.39% and79.45% for the YOLO baseline. However, McNemar’s exact test did not find the accuracy difference statistically significant (p = 0.629). Synthetic data may help, but its effect depends on real-data availability and the domain gap.
Authors
- Amir Nemati (ORCID: https://orcid.org/0000-0002-1998-4893)
- Robert Weidner (ORCID: https://orcid.org/0000-0003-1286-4458)
- Dennis Bäcker
- Krishna Kumar Murali
Institutions
- TU Bergakademie Freiberg (DE)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11600216/v1
- Primary Topic
- Industrial Vision Systems and Defect Detection
- Type
- preprint