songguorong / edrnet

⚡EDRNet:Encoder-Decoder Residual Network for Salient Object Detection of Strip Steel Surface Defects

EDRNet

source code for our IEEE TIM 2020 paper entitled EDRNet: Encoder-Decoder Residual Network for Salient Object Detection of Strip Steel Surface Defects (DOI：10.1109/TIM.2020.3002277) by Guorong Song, Kechen Song and Yunhui Yan.

Requirement

• Python 3.6
• Pytorch 0.4.1 or 1.0.1(default)
• numpy
• torchvision
• glob
• PIL
• scikit-image

This code is tested on Ubuntu 16.04.

Training

1. cd to ./Data, and Unzip the file of trainingDataset.zip into this folder.
2. path of training images:./Data/trainingDataset/imgs_train/ path of training labels:./Data/trainingDataset/masks_train/
3. runpython edrnet_train.py to start training
4. the trained model will be saved in ./trained_models

Testing

1. download the test dataset SD-saliency-900.zip, then Unzip it to the directory of ./Data
2. download the pre-trained model EDRNet_epoch_600.pth, then put it to the directory of ./trained_models
3. path of testing dataset: ./Data/SD-saliency-900/imgs/ path of pre-trained model: ./trained_models/EDRNet_epoch_600.pth
4. runpython edrnet_test.py to start testing
5. the predict results will be saved in ./Data/test_results/

Note: If you use SD-saliency-900 dataset in your paper, please cite Saliency detection for strip steel surface defects using multiple constraints and improved texture features

Results

We also provide the experimental results of all the comparative methods in our paper.(Results)

You can also download all the files including SD-saliency-900.zip, EDRNet_epoch_600.pth, Results in BaiduYun Drive.(link:https://pan.baidu.com/s/1RSgkzNKxXA11ajtoFnk6Mw code: z91m)

Supplement

Here, we provide the results tested on Noisy Images with Salt and Pepper noise. (GoogleDrive) BaiduYun Drive: (link:https://pan.baidu.com/s/1jw8jHEpa_AWgf2rMpmsebQ code:c9gb)

• mat_Results.zip
• NoisyImages.zip
• NoisyTestResults.zip

Performance Preview

Visual comparison

Quantitative comparison

Citation

@InProceedings{SGR_2020_TIM,
author = {Song, Guorong and Song, Kechen and Yan, Yunhui},
title = {EDRNet: Encoder-Decoder Residual Network for Salient Object Detection of Strip Steel Surface Defects},
booktitle = {IEEE Transactions on Instrumentation & Measurement (IEEE TIM)},
month = {June},
year = {2020}
}

edrnet's Issues

Image Size

Hello,
Thanks for your job.
Can you please tell me is there a limit of the size of my images in my training dataset and testing dataset? And is the code has the image resize?

test dataset

作者，您好！能分享一下10%噪声和20%噪声的数据集吗？如果可以的话，还有这两个噪声数据集对应于论文中15个模型的测试结果。感谢！

GPU

你好，
感谢你所做的贡献，我想请问一下啊，如果我想用CPU代替GPU进行模型训练，我应该更改哪个参数

Input

你好，
请问你知道如何才能把edrnet_test文件夹图片输入改为单张图片输入的模式吗？

测试结果

你好我想问一下 表2 的三列 上面的噪声比例 是指训练集的还是测试集的 可否将不同噪声的数据集发一下谢谢！！！

Evaluation Metrics

作者，您好！我对您论文中的评价指标比较感兴趣，可否上传一下论文中测试那5个指标的代码，万分感谢！

train test

为什么训练集 和 测试集的数据有重复