Classification Method of Surface Defects of Aluminum Profile Based on Transfer Learning
Lists
Only you see this
Abstract
In the field of industrial defect detection based on machine learning, small and unbalanced number of sample data has been a challenge to the defect classification task. In this paper, the number of aluminum profile defect pictures is increased by digital image processing methods such as rotation and flipping. Then, a VGG16 network is pre-trained on ImageNet dataset, in which the weights of feature extraction modules are transferred to the classification task of surface defects of aluminum profile. Furthermore, the classification modules of VGG16 are structurally fine-tuned, the number of neurons in the second and third fully connected layers of which is adjusted to 256 and 10 respectively. The experimental results show that the classification accuracy of 10 types of aluminum profile surface defects reaches 98.47%, and the training process is stable without any obvious overfitting phenomenon. Our results suggest that the proposed method for aluminum profile surface defect classification is smooth in feature transferring and structural fine-tuning and renders reliable results in defect classification.
BibTeX
@inproceedings{li2022classification,
author = {Bin Li and Fuji Ren and Hongjun Ni and Xin Kang and Shuaishuai Lv and Zhuangzhuang Hao},
title = {{Classification Method of Surface Defects of Aluminum Profile Based on Transfer Learning}},
booktitle = {International Conference on Machine Learning},
year = {2022},
doi = {10.1109/mlise57402.2022.00008},
}Loading related content...