作成者 |
|
|
|
|
|
|
|
|
本文言語 |
|
出版者 |
|
|
発行日 |
|
収録物名 |
|
巻 |
|
号 |
|
開始ページ |
|
終了ページ |
|
出版タイプ |
|
アクセス権 |
|
Crossref DOI |
|
権利関係 |
|
概要 |
Checking the surface of steel is still primarily done manually and visually in the industrial business. Because of the high number and surface area of steel, inspection is time-intensive and of poor q...uality. Machine learning may be used to detect steel surfaces automatically by training a model with a large amount of data, and the model findings can then be used to detect other steel surfaces. Automatic detection will undoubtedly save control time, save prices, and improve checking quality. This research uses the core architecture of UNet and five versions of EfficientNet (B0 to B4) as the backbone (encoder). In this research, we use three types of training processes: the first is binary classification to predict the presence of defects; the second is multi-label classification to predict the type of defect; and the third is image segmentation to determine the location of defects according to their type. The results of this research show that the EfficientNet-B0 version can generally provide the best measurement results. Except for measuring recall in the binary classification process, it turns out that EfficientNet-B4 is the best, and measuring dice loss in the segmentation process for defect-4 turns out that EfficientNet-B2 gives the best results. Based on the accuracy value obtained from our study, it is still feasible to attempt employing a different type of architecture as a backbone (other than EfficientNet) to acquire even greater accuracy values in order to detect uncommon faults on the steel surface.続きを見る
|