By Guizhong Fu, Peize Sun, Wenbin Zhu, Jiangxin Yang, Yanlong Cao, Michael Ying Yang and Yanpeng Cao.
The paper is available at |[PDF Download]
In this paper, we present a compact yet effective convolutional neural network (CNN) model, which emphasizes the training of low-level features and incorporates multiple receptive fields, to achieve fast and accurate steel surface defect classification.Our proposed method adopts the pre-trained SqueezeNet as the backbone architecture.We also construct a diversity-enhanced testing dataset of steel surface defects to evaluate the robustness of classification models. The dataset contains severe camera noise, non-uniform illumination, and motion blur.
The enhanced dataset is build on NEU dataset:
A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects
neural networks | |[Download]
The enhanced dataset can be download at:
Google Drive: https://drive.google.com/file/d/16HjqGvnr_OUfTF0HSN1XyNg9XPr4Edfw/view?usp=sharing
Baidu Cloud: https://pan.baidu.com/s/18OowNbcJmBIi92fTWKI2wg Password:fcmh
This code is based on Caffe. Thanks to the contributors of Caffe. Caffe: https://github.com/BVLC/caffe
Our Model also uses the pretrain model SqueezeNet-1.0 SqueezeNet: https://github.com/DeepScale/SqueezeNet
Method | Running time | Model size | Accuracy on NEU dataset | Accuracy on enhanced dataset |
---|---|---|---|---|
ETE | 5.3ms | 1.9MB | 95.8% | 80.3% |
DECAF+MLR | 10.3ms | 244MB | 99.7% | 91.3% |
SDC-SN-ELF+MRF | 8.0ms | 3.1MB | 100% | 97.5% |
[ETE] An end-to-end steel strip surface defects recognition system based on convolutional
neural networks | |[pdf]
[DECAF+MLR] A generic deep-learning-based approach for automated surface inspection | |[pdf]
If you have any questions, feel free to contact:
- Guizhong Fu (fuguizhongchina@163.com)
- Yanpeng Cao (caoyp@zju.edu.cn)
Copyright(c) Guizhong Fu and Yanpeng Cao All rights reserved.