A

PROJECT REPORT

ON

REAL TIME STEEL SURFACE DEFECT DETECTION USING

DCNN

BY

Names of Student: - ID No

Atharva M. Dharwa 2017AAPS0266G

AT

CEERI, Pilani

A Practice School-II station

BIRLA INSTITUTE OF TECHNOLOGY & SCIENCE, PILANI

(October, 2020)

1
 A

PROJECT REPORT

ON

REAL TIME STEEL SURFACE DEFECT DETECTION USING

DCNN

BY

Names of Student: - ID No Discipline

Atharva M. Dharwa 2017AAPS0266G ECE

Prepared in partial fulfilment of the

Practice School-II Course No.

AT

CEERI, Pilani

A Practice School-II station

BIRLA INSTITUTE OF TECHNOLOGY & SCIENCE, PILANI

(October, 2020)

2
 Acknowledgement

I express sincere regards to my mentor, Dr. Dhiraj from CEERI for guiding me

right from the start. I thank him for extending his valuable guidance, support for &

critical reviews of the project and above all, the moral support and motivation he

provided me with during all stages of the project.

And my special thanks to Prof. Pawan Sharma, Faculty-in-charge, BITS-Pilani

PS division, for guiding me to work in the esteemed institution and providing a

platform for growth and learning.

3
4
 BIRLA INSTITUTE OF TECHNOLOGY & SCIENCE

PILANI (RAJASTHAN)

Practice School Division

Station: CEERI Centre: Pilani

Duration From: 24th July, 2020 To: 21st December, 2020

Date of Submission: 6th October, 2020

Title of the project: Real time Steel surface defect detection using DCNN

Names of Students: - ID No Disciplines

Atharva 2017AAPS0266G ECE

Name of expert: Dr. Dhiraj

Designation: Senior scientist at CEERI

Name of the PS Faculty: Prof. Pawan Sharma

Keywords: steel defect detection, deep learning, CNN, logistic regression, random forest, SVM

Project Area(s): Deep learning, Machine Learning

(Signature of Student(s)) (Signature of PS Faculty)

Date: Date:

5
 Abstract

Steel is one of the most successful alloys with wide range of applications in

several domains. However, the product of usable quality steel has never been

simple. The process involves the influence of raw materials, system control,

rolling process, etc. which results in several defects over its surface. Because of

the requirement of defect free steel, it is critically important to detect these

defects in as lesser time as possible to boost the production speeds. Once the

defect is detected, it can be used to treat this steel piece further to remove it

based on the class of defect. However, the steel could have several types of

defects such as scratches, scars, iron scales, black burn, bright prints, inclusion,

rolled-in scale, patches, crazing, pitted surface, pollution and other defects. Apart

from the appearance, these defects reduce the properties such as abrasion

resistance, corrosion resistance, and fatigue strength which causes huge loss to

the industries.

Until recent times, the detection process has been quite manual and slow.

However, due to the changing technologies, it is now possible to detect such

changes using artificial intelligence in real time. Due to the fast processors, it is

possible to reduce human involvement and increase detection speeds. The main

6
objective of the project is to improve the error detection rates of these systems by

experimenting with several algorithms.

7
 Table of Contents
Acknowledgement............................................................................................................3

Abstract............................................................................................................................5

Table of Contents.............................................................................................................6

Introduction......................................................................................................................7

Aim of the project.............................................................................................................8

Current Progress..............................................................................................................8

Week 1:.........................................................................................................................9
Week 2:.......................................................................................................................10
Week 3:.......................................................................................................................11
Week 4:.......................................................................................................................13
Week 5:.......................................................................................................................15
Week 6:.......................................................................................................................17
Week 7:.......................................................................................................................19
Week 8:.......................................................................................................................20
Week 9:.......................................................................................................................21
Week 10:.....................................................................................................................23
Conclusion.....................................................................................................................25

Recommendation...........................................................................................................26

References.....................................................................................................................27

Glossary.........................................................................................................................28

8
 Introduction

Central Electronics Engineering Research Institute (CEERI), located at Pilani,

Jhunjhunu District, Rajasthan and Chennai, Tamil Nadu is a research institute in

India and a constituent laboratory of Council of Scientific and Industrial Research

(CSIR India), New Delhi. CEERI Pilani was established on 21st September, 1953

in Pilani, Jhunjhunu District, Rajasthan for advanced research and development

in the field of Electronics. It was established with the vision of Mr. G. D. Birla.

CEERI excels in research in various domains such as IoT, cognitive computing,

Integrated systems, etc.

Since its inception, it has been working for the growth of electronics in the

country and has established the required infrastructure and well experienced

manpower for undertaking R&D in the following major areas:

Cyber Physical Systems

Microwave Tubes

Smart Sensors

The Chennai center focuses on process control instrumentation and automation

as well as machine vision technologies.

9
 Aim of the project

The aim of the project is to propose a superior algorithm to increase the accuracy

for the steel defect detection and classification. Various feature extractors such

as LBP, LTP, CLBP, and AECLBP can be used with a combination with several

classifiers such as SVM, logistic regression, random forest, and CNN to find

optimum combination with maximum accuracy.

10
 Current Progress

Several concepts have been studied to understand various concepts in Machine

Learning and Deep Learning. These concepts are important to build the

backbone of the project. Only after learning the basics, understanding the

concepts of already adopted by methods can be understood. By applying those

methods, it is easier to find the gaps in the current implementation and solve

propose for a superior method to detect and classify the defects in real time with

higher accuracy. Almost all the concepts that have been studied in machine

learning have been tried out with our dataset.

NEU dataset for steel defect has been planned to be used for the project for the

initial part. Upon using satisfactory results, various other datasets can be

explored.

The following timeline has been followed till now:

11
Week 1:

Aim

Get better understanding of the topic and get stronger on the basics.

Skills Gained (Technical as well as soft skills)

1. Gradient Descent

2. Linear Regression

3. Supervised and unsupervised learning basics

Major Learning & Project Milestones

1. Idea of topic given

2. Basic of ML started

Summary:

I got introduced to the core idea of teaching a computer to learn concepts using

data—without being explicitly programmed. I started by covering linear

regression with one variable. Linear regression predicts a real-valued output

based on an input value. Worked on the application of linear regression to

12
housing price prediction, present the notion of a cost function, and introduce the

gradient descent method for learning.

13
Week 2:

Aim

To improve the MATLAB skills for ML.

Skills Gained (Technical as well as soft skills)

1. MATLAB and Octave basics

2. Polynomial Regression

3. Gradient descent for multiple variables

Major Learning & Project Milestones

1. Details of the project confirmed

2. Using multiple variables to detect defects

Summary:

I have covered linear regression with multiple variables. I have seen how linear

regression can be extended to accommodate multiple input features. I have

explored best practices for implementing linear regression.

14
I went over how to use MATLAB. I have worked on programming course

assignments designed to help me understand how to implement the learning

algorithms in practice.

15
Week 3:

Aim

Focus on implementation of logistic regression. Multiclass classification shall be

studied keeping in mind the steel defect classification that has to be

implemented.

Skills Gained (Technical as well as soft skills)

1. Multiclass Classification

2. Multiclass classification: one vs all

3. Logistic regression

Major Learning & Project Milestones

1. Read through several papers and summarized it.

2. From CNN to computer vision has already been implemented in this domain.

3. Logistic Regression in not just basic but also used widely.

Summary

16
This week, I have looked into covering logistic regression. Logistic regression is a

method for classifying data into discrete outcomes. For example, we might use

logistic regression to classify an email as spam or not spam. I got introduced to

the notion of classification, the cost function for logistic regression, and the

application of logistic regression to multi-class classification.

I have also looked over regularization. Machine learning models need to

generalize well to new examples that the model has not seen in practice. We’ll

introduce regularization, which helps prevent models from overfitting the training

data.

17
Week 4:

Aim

Explore Neural Networks. Learn to test the data using one vs rest method for our

steel defect detection.

Skills Gained (Technical as well as soft skills)

1. Multiclass Classification

2. Neural Network basics

3. Non-linear hypothesis

Major Learning & Project Milestones

1. Project flow decided

2. Understanding CNN is important to implement it.

3. Basic model to learn CNN implemented

Summary:

I have covered neural networks this week. Neural networks is a model inspired

by how the brain works. It is widely used today in many applications: when the

phone interprets and understand the voice commands, it is likely that a neural

18
network is helping to understand the speech; when we cash a check, the

machines that automatically read the digits also use neural networks.

19
Week 5:

Aim

Learn to train the data the Neural Network (NN). Algorithms like back

propagation play an important role to improve the accuracy of our Neural

Network model.

Skills Gained (Technical as well as soft skills)

1. Learned to train the neural network

2. Learned back propagation to improve the weights of the algorithm

Major Learning & Project Milestones

1. The project flow decided.

2. Scope of internship finalized.

3. Studied some important architectures in ML.

Summary:

I have learned how to train Neural Networks. The Neural Network is one of the

most powerful learning algorithms (when a linear classifier doesn't work, this is

what people usually turn to), and the 'backpropagation' algorithm for training was

20
used to train the models. I have also gotten a chance to implement this algorithm

and see it work for myself.

21
Week 6:

Aim

Learn to improve the algorithm accuracy that is required to tune our model.

Skills Gained (Technical as well as soft skills)

1. Learned to diagnose bias vs variance

2. Learned Error Analysis

3. Learned about the error metrics for skewed class

Major Learning & Project Milestones

1. Finished 6 weeks of course

2. Learned various useful algorithms for the project domain

3. Research for the details of flow of the project

Summary

I learned about systematically improving our learning algorithm. I have focused

on learning how to tell when a learning algorithm is doing poorly, and describe

the 'best practices' for how to 'debug' our learning algorithm and go about

improving its performance.

22
I have also covered machine learning system design. To optimize a machine

learning algorithm, it is important to first understand where the biggest

improvements can be made. I have learned how to understand the performance

of a machine learning system with multiple parts, and also how to deal with

skewed data.

23
Week 7:

Aim

Learn about Support Vector Machines (SVM). As one of the classifiers to test our

steel defect data on is SVM, it is important to learn the basics and ways to

improve it.

Skills Gained (Technical as well as soft skills)

1. Large Margin Classification

2. Kernels

3. SVM

Major Learning & Project Milestones

1. Finished 7 weeks of course content

2. Researched several methods to improve SVM classifier

Summary

This week, I learned about the support vector machine (SVM) algorithm. SVMs

are considered by many to be the most powerful 'black box' learning algorithm,

24
and by posing a cleverly-chosen optimization objective, one of the most widely

used learning algorithms today.

Week 8:

Aim

Implement the first part of the paper by Kechen Song, that is, implement the

dataset available using SVM and LBP method

Skills Gained (Technical as well as soft skills)

1. LBP

2. Kernels

3. SVM

Major Learning & Project Milestones

1. Finished first stage implementation using LBP

Summary:

LBP is an algorithm that considers an image’s neighbors pixel value to generate

a new image. If the corresponding neighbor has a greater than or equal to the

current pixel value, then the true or ‘1’ value is given to the pixel else, false or ‘0’

25
is given to it. By using SVM as a classifier provided with the self-

implemented LBP processed image, we were able to attain an accuracy of

93.33%.

Week 9:

Aim

Implement the complete paper using LTP, CLBP, and AECLBP and SVM on our

dataset.

Skills Gained (Technical as well as soft skills)

1. SVM

2. LBP

3. LTP

4. CLBP

5. AECLBP

Major Learning & Project Milestones

1. Finished complete paper implementation using LBP, LTP, CLBP, AECLBP

Summary

26
Since the codes for any of the methods such as LTP, CLBP, AECLBP are not

available (as they are relatively new and not being used in practical applications),

it was important to self-implement these codes in MATLAB to test the accuracy

on our dataset. After running these several times and tweaking its performance, I

was able to achieve an accuracy of 89.17%, 97.78% and 95.83% using LTP,

CLBP, and AECLBP respectively. All these images have been tested without

noise. Compared with 93.33% accuracy that was obtained by LBP, CLBP has

achieved maximum accuracy without noise on our dataset.

27
Week 10:

Aim

Implement the all stages of the paper including LBP, LTP, CLBP, AECLBP using

several classifiers like SVM, CNN, Logistic Regression, and Random Forest

Skills Gained (Technical as well as soft skills)

1. SVM

2. Logistic Regression

3. CNN

4. Random Forest

Major Learning & Project Milestones

1. Finished complete paper implementation of LBP, LTP, CLBP, AECLBP with

several classifiers

Summary

It took some time to understand and implement new algorithms for classifying

other than the SVM that has been previously used. New methods that were

28
implemented in MATLAB are logistic regression, CNN, and Random Forest.

However, even after trying these various algorithms for classifying, SVM appears

to be best classifying on our dataset.

29
 Conclusion

After trying several methods of feature extraction (LBP, LTP, CLBP, AECLBP) and classification (SVM,

Random Forest, Logistic Regression, CNN), the best result obtained yet was using CLBP and SVM.

However, more research has to be done to test these results under various conditions such as in

addition of noise.

30
 Recommendation

It is recommended to try various different settings to bring improvement to the current models. Also, a

greater number of classifiers and extraction features can be tried out to reach a better solution.

Different hurdles such as noise can be added to check the robustness of the conclusion.

31
 References

Song, Kechen, and Yunhui Yan. "A Noise Robust Method Based On Completed Local Binary Patterns For Hot-

Rolled Steel Strip Surface Defects". Applied Surface Science, vol 285, 2013, pp. 858-864. Elsevier BV,

doi:10.1016/j.apsusc.2013.09.002.

NG, Andrew. "Machine Learning". Coursera, 2020, https://www.coursera.org/learn/machine-learning.

32
 Glossary

 LBP : Local Binary Pattern

 LTP: Local Ternary Pattern

 CLBP: Completed Local Binary Pattern

 AECLBP: Adjacent Evaluation Completed Local Binary Pattern

33