Skip to content
项目

ShuaiLYU/Deep-Learning-Approach-for-Surface-Defect-Detection

在 GitHub 查看

ShuaiLYU/Deep-Learning-Approach-for-Surface-Defect-Detection has 750 GitHub stars and 257 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-19 18:59 UTC. Last push 2020-05-14.

这些数字如何统计

实测星标增长

24小时
0
7天
正在收集历史
15天
正在收集历史
1个月
正在收集历史

各周期内的星标净变化,周期相互重叠,并非每日新增。缺失的测量保持空白。星标代表关注度,不代表软件质量。

关于

A Tensorflow implementation of "Segmentation-Based Deep-Learning Approach for Surface-Defect Detection"

机器学习

README.md

原文

Deep-Learning-Approach-for-Surface-Defect-Detection

A Tensorflow implementation of "Segmentation-Based Deep-Learning Approach for Surface-Defect Detection" The author submitted the paper to Journal of Intelligent Manufacturing (https://link.springer.com/article/10.1007/s10845-019-01476-x), where it was published In May 2019 .

The test environment

python 3.6
cuda 9.0
cudnn 7.1.4
Tensorflow 1.12

You should know

I used the Dataset used in the papar, you can download KolektorSDD here. If you train you own datset ,you should change the dataset interfence for you dataset.

You can refer to the paper for details of the experiment.

my experimental results on KolektorSDD

Notes: the first 30 subfolders are used as training sets, the remaining 20 for testing. Although, I did not strictly follow the params of the papar , I still got a good result.

2019-05-21 09:20:54,634 - utils - INFO -  total number of testing samples = 160
2019-05-21 09:20:54,634 - utils - INFO - positive = 22
2019-05-21 09:20:54,634 - utils - INFO - negative = 138
2019-05-21 09:20:54,634 - utils - INFO - TP = 21
2019-05-21 09:20:54,634 - utils - INFO - NP = 0
2019-05-21 09:20:54,634 - utils - INFO - TN = 138
2019-05-21 09:20:54,635 - utils - INFO - FN = 1
2019-05-21 09:20:54,635 - utils - INFO - accuracy(准确率) = 0.9938
2019-05-21 09:20:54,635 - utils - INFO - prescision(查准率) = 1.0000
2019-05-21 09:20:54,635 - utils - INFO - recall(查全率) = 0.9545

visualization: kos49_Part4.jpg

testing the KolektorSDD

After downloading the KolektorSDD and changing the param[data_dir]

python run.py --test

Then you can find the result in the "/visulaiation/test" and "Log/*.txt"

training the KolektorSDD

First, only the segmentation network is independently trained, then the weights for the segmentation network are frozen and only the decision network layers are trained.

training the segment network

python run.py --train_segment

training the decision network

python run.py  --train_decision

training the total network( not good)

python run.py  --train_total