발간년도 : [2021]
| 논문정보 |
|
| 논문명(한글) |
[Vol.16, No.5] Deep Learning based Surface Defect Detector on Metal Surface |
|
| 논문투고자 |
Sangmin Suh |
|
| 논문내용 |
Steel surface defect should be detected and repaired in steel industry. Therefore, automatic detection of the steel surface defects plays a vital role in the steel manufacturing process. For the defect detection, machine learning based classification methods have been widely used such as HAAR feature-based cascade classifiers and support vector machines (SVM). As deep learning methods have been popular, the neural network based surface defect detection has been recently introduced. As for the methods, many researchers, in general, adopt a trained neural network, which is mainly winner in the recent ILSVRC (ImageNet Large Scale Visual Recognition Challenge). Then, the weights and last layers are modified to be used for surface defect detection (SDD), which is called transfer learning. In the previous researches, ResNet152 (winner in ILSVRC 2015) was used and the resulting performances were F1=0.975 and F1=0.912 in two different studies, respectively. However, the neural network used in their research has very wide and deep. Therefore, huge memories to save the trained weights and many multiplier–accumulators (MAC) are necessary, which means expensive hardware systems are essential to predict surface defect on the steel surface. This paper suggests a small neural network dedicated to surface defect detection. The proposed network has only three convolution layers and two fully connected layers. From the experimental results, we obtained F1=0.931 and minimum AUC (area under the curve)=0.995. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
