Loading [a11y]/accessibility-menu.js
Automated ceramic plate defect detection using ScaledYOLOv4-large | IEEE Conference Publication | IEEE Xplore

Automated ceramic plate defect detection using ScaledYOLOv4-large


Abstract:

Automated visual inspection has become a popular topic of research in the last couple of decades, as computation power available grew exponentially. Judging by the fact t...Show More

Abstract:

Automated visual inspection has become a popular topic of research in the last couple of decades, as computation power available grew exponentially. Judging by the fact that visual inspection is a critical task for the quality of the products, it would be highly recommended that people only supervise the system. This paper proposes a low cost, rapid development defect detection system based on the Scaled-YOLOv4 object detection model. The original model achieves almost state of the art detection mAP on the COCO dataset with a mAP(mean average precision) of 56.0 for the largest variant, named YOLOv4-P7. Our version is derived from the ScaledYOLOv4-P5 model and is trained on ceramic plate defects and achieves 87.4 mAP at a intersection of union of 0.5, while comfortably processing a frame in 20ms on a consumer RTX3070 GPU. Thus, the real time constraint for the manufacturing system is fulfilled. Hence, the critical aspects of the development process are the: quick development process, fast training, rapid deployment on the factory floor, quick validation and feedback, using images acquired in the lab - not on the factory floor for first iteration and overall low cost of the automated inspection system.
Date of Conference: 01-03 July 2021
Date Added to IEEE Xplore: 23 August 2021
ISBN Information:
Conference Location: Pitesti, Romania

References

References is not available for this document.