MORE | Fall 2024

Improving YOLOv8 Transferability via Model Fine-Tuning with Domain-Specific Manufacturing Data

Pretrained Deep Learning models frequently confront performance bottlenecks on data from manufacturing applications, due to the models’ limited knowledge about the domain-specific data. YOLOv8, as a popular computer vision model for object recognition, has been adopted in many fields such as aerospace, marine, and construction for crack detection and quality evaluation. It provides an automatic solution compared with the traditional manual inspection. To enhance the crack detection performance of YOLOv8 and enable its transfer across various manufacturing datasets, this work will develop a model fine-tuning method to improve YOLOv8’s domain knowledge and overcome its performance bottlenecks on manufacturing data. We will use the fine-tuned YOLOv8 model to detect crack patterns on weld surfaces of metal sheets, evaluating its transferability enhancement as a result of the model fine-tuning.