Source code used on "A robust and fast deep learning-based method for defect classification in steel surfaces" paper.
- NEU Experiments - ReadMe
- Copyright (C) 2018, Vicomtech (http://www.vicomtech.es/),
- (Spain) All rights reserved.
- fsaiz@vicomtech.org
- DESCRIPTION
This directory contains the source code used on "A robust and fast deep learning-based method for defect classification in steel surfaces" paper.
It contains the following scripts:
- preProcess: apply filters to highlight the defects. (i) imadjustn: Adjust image intensity values (achieves the best classification accuracy) (ii) histeq: Enhance contrast using histogram equalization (iii) localcotnrast: Edge-aware local contrast manipulation of images
- dataAugmentation: apply transformations
- generateOcclusion: generate a percentage of occlusions
- increaseBrightness: augment the brightness
- NEU_CNN_[train|test] with different architectures (1) AlexNet (2) GoogleNet (3) ResNet50
- SETTINGS
-
NEUpath: folder of the input images
-
architecture_net: 1 = AlexNet v2
2 = pretrained GoogLeNet (224x224x3)
3 = pretrained ResNet50 (224x224x3) -
k: k-fold cross validation
-
trainingOptions parameters
Please if you use the code, reference our work:
Saiz, F.A. , Serrano, I., Barandiran. I., Sanchez, J.R. A robust and fast deep learning-based method for defect classification in steel surfaces. IEE-IS18. pp. 455-460 (2018)