Ceramic Tiles Defect Detection System

A Project Report
submitted in the partial fulfillment of the requirements for the award of the degree of

Bachelor of Technology in Computer Science & Engineering

Faculty of Engineering
Submitted by

NAME Enrollment ID

Amandip Singh ADTU/2020-24/BCS/026

Chintu Kalita ADTU/2020-24/BCS/021

Vivek Kalita ADTU/2020-24/BCS/036

Under the Guidance of

Dr. Aniruddha Deka
Associate Professor

Faculty of Engineering

Assam down town University

Sankar Madhab Path, Gandhi Nagar Panikhaiti, Guwahati,

Assam 781026

May, 2024

1
 Assam down town University
Faculty of Engineering

CERTIFICATE OF APPROVAL

This is to certify that the project report entitled “Ceramic Tiles Defect Detection System” is
hereby approved as a creditable study of the project and has been submitted by Chintu
Kalita bearing Enrollment No. ADTU/2020-24/BCS/021, Amandip Singh bearing Enrollment
No. ADTU/2020-24/BCS/026, and Vivek Kalita bearing Enrollment No.
ADTU/202024/BCS/036 is in satisfactory manner to warrant its acceptance as a
prerequisite to the degree of Bachelor of Technology in Computer Science and
Engineering under Assam downtown University for which it has been submitted.

It is understood that by this approval, the undersigned do not necessarily endorse any
conclusion drawn or opinion expression therein, but approve the report for the purpose for
which it is submitted.

Dr. Utpal Barman

Dean, Faculty of Computer Technology & Dean (Additional Charge),
Computer Science & Engineering. Faculty of Engineering & Technology,
Assam down town University, Guwahati, Assam, India

2
 Assam Down Town University
Faculty of Engineering

CERTIFICATE FROM SUPERVISOR

This is to certify that the project report entitled "Ceramic Tiles

Defect Detection," submitted by Amandip Singh, Chintu Kalita, and
Vivek Kalita, is being presented in partial fulfillment of the
requirements for the award of the degree of Bachelor of Computer
Science & Engineering in the Faculty of Engineering at Assam
down town University. This report is a record of bona fide work
carried out under effective guidance and supervision. The results
embodied in this report are original and have not been copied from
any other department, university, or institute.

Signature of the Supervisor

Dr. MRINAL GOSWAMI
Associate Professor
Programme of Computer Science & Engineering
Faculty of Engineering & Technology
Assam down town University
Guwahati. Assam, India

3
 Assam down town University
Faculty of Engineering

CERTIFICATE FROM GUIDE

This is to certify that the project report entitled “Ceramic Tiles Defect Detection System ” is
hereby approved as a creditable study of the project and has been submitted by Chintu Kalita
bearing Enrollment No. ADTU/202024/BCS/021, Amandip Singh bearing Enrollment No.
ADTU/2020-24/BCS/026, and Vivek Kalita bearing Enrollment No. ADTU/202024/BCS/036 is
in satisfactory manner to warrant it’s acceptance as prerequisite to the degree of Bachelor of
Technology in Computer Science And Engineering under Assam down town University is a
bonafide project work carried out by them under my supervision and guidance found to be
satisfactory.

I recommend the thesis for consideration for the award of the degree of Bachelor of Technology in
Computer Science and Engineering under Assam downtown University.

Date: Dr. Aniruddha Deka

Place: Guwahati, Assam Associate Professor

Faculty of Engineering & Technology

Assam down town University

4
 Faculty of Engineering

CERTIFICATE FROM EXTERNAL EXAMINER

This is to certify that the project report entitled “Ceramic Tiles Defect Detection System ” is
hereby approved as a creditable study of the project and has been submitted by Chintu Kalita
bearing Enrollment No. ADTU/202024/BCS/021, Amandip Singh bearing Enrollment No.
ADTU/2020-24/BCS/026, and Vivek Kalita bearing Enrollment No. ADTU/202024/BCS/036 is
in satisfactory manner to warrant it’s acceptance as prerequisite to the degree of Bachelor of
Technology in Computer Science And Engineering under Assam Downtown University is a
bonafide project work carried out by them under the supervision and guidance of Dr. Aniruddha
Deka, Associate Professor, has been examined by me and found to be satisfactory.

I recommend the thesis for consideration for the award of the degree of Bachelor of Technology in
Computer Science and Engineering under Assam down town University.

Date:

Place: Guwahati, Assam External Examiner

5
 Assam down town University
Faculty of Engineering

DECLARATION

The Project Report entitled “Ceramic Tiles Defect Detection” is a record of Bonafide work
of “ Chintu Kalita bearing Roll No. ADTU/2020-24/BCS/021, Amandip Singh bearing
Roll No. ADTU/2020-24/BCS/026, Vivek Kalita bearing Roll No.
ADTU/2020-24/BCS/036,” submitted in partial fulfillment for the award of Bachelor of
Technology in Computer Science and Engineering in the Faculty of Engineering in
Assam downtown University.

The results embodied in this report have not been copied from any other
departments/University/Institute.

Date: Chintu Kalita

Place: Guwahati, Assam Roll No: ADTU/2020-24/BCS /021

Amandip Singh

Roll No: ADTU/2020-24/BCS /026

Vivek Kalita

Roll No: ADTU/2020-24/BCS /036

6
 ACKNOWLEDGEMENT

Foremost, we express our sincere gratitude to our project Guide, Dr. Anirudha Deka. Sir’s
invaluable guidance, encouragement, and expertise were instrumental in shaping the
direction of this research and keeping us on track. His willingness to dedicate time to
answer our questions and address our challenges was pivotal for our research.

We would also like to extend our thanks to the Department of CSE under Faculty of
Engineering, Assam down town University. Their assistance in providing access to essential
equipment and resources was crucial in conducting experiments and collecting data. Their
technical support proved to be invaluable throughout the project.

Finally, we would like to express our appreciation to the administration of Assam down
town University as well as the teaching and non-teaching staff of the department of CSE for
their unwavering encouragement and support throughout this journey.

Chintu Kalita
ADTU/2020-24/BCS/021
Amandip Singh
ADTU/2020-24/BCS/026
Vivek Kalita
ADTU/2020-24/BCS/036

7
INDEX
● Certificate of Approval………………………….….2
● Certificate from Guide……………………………...3
● Certificate from external examiner………………....4
● Declaration……………………………………….…5
● Acknowledgment……………………………….…..6
● Index………………………………………………...7
● Abstract……………………………………………..9

CHAPTERS TITLE PAGE NO

1. Introduction………………………………….………………………………………..10
1.1. Motivation…………………………….………..11
1.2. Scope…………………………………..………...11
1.3. Problem Statement.………………….….12
1.4. Objective…………………………………………12
1.5. Benefits of proposed system………12

2. Literature Review………………………………………………………………………14
2.1. Trend Analysis………………………………………….18
2.2. Key Findings……………………………………………..18
2.3. Survey on Existing System……………………..19

3.System analysis and Design……………………….……………..……….…..21

3.1. Proposed system………………………………………21
3.2. Advantages and Disadvantages ………….…22
3.3. Flow chart of proposed system……………...23
3.4. Gantt Chart…………………………………………..…….25

4.Implementation and Results discussion………………………………..27

4.1. Project Implementation…………………………….27
4.2. Description of the software used………….…35
4.3. Software development Approach…………….36

8
 4.4. Sample Source code………………..….....37
4.5. Process Flow Diagram…………………….39
4.6. Results……………………………………………….40

5. Conclusion and Future Scope…………………………………41

6. Reference…………………………………………………………………..43

9
 Abstract
The high-volume ceramic tile industry grapples with the inefficiency and subjectivity
inherent in manual quality control. Traditional methods rely on human inspectors to
identify defects like cracks, holes, edge defects and color variations on each individual tile.
This process is not only time-consuming but also prone to errors due to fatigue and the
reliance on experience-based judgment. In actual production, the difficulty in extracting
texture ceramic tile features and the small size of defects lead to low detection accuracy
and inefficiency. To address these limitations, this paper proposes utilizing digital image
processing (DIP) techniques and for training the model we used YOLO (YOLOv5 and
YOLOv8l model). The model has been trained for 50 epochs, resulting in a final validation
on the weights saved as best.pt This YOLOv8 model, utilizing a Tesla T4 GPU, comprises
268 layers with a total of 68,126,457 parameters and operates at 257.4 GFLOPs (Giga
Floating-point Operation Per Second). For the problem of complex background and small
defect size in the dataset, we improve the model in turn. We are further conducting a
comprehensive review to enhance and classify existing defects.

10
 Chapter 1

Introduction
Defect detection or anomaly detection in industrial processes is an important
procedure. Currently, manual and visual inspections performed by experts are expensive
due to the high cost of human labor during working hours, the possibility of material
waste, and the degraded quality of shipped products. In contrast, the use of machine
learning algorithms for automatic defect detection reduces labor consumption. In recent
years, automatic defect detection has played a critical role in the industry’s inspection
process. Current non-destructive testing methods for ceramic tiles mainly include
ultrasonic testing, machine vision inspection, and X-rays. However, this method is costly
and inefficient, which is not suitable for the needs of tile inspection in the industry.
Hanzaei et al. proposed an automatic image processing system. The defect edge detection
is performed by using statistical methods for the defect statistics of different diameters.
However, their system requires a high radius size for defects, and when the radius size is
too large or too small, the detection error reaches over 35%.
To address the limitations of existing methods, in this study, ceramic tiles with
different backgrounds are used as research objects, and YOLOv5 is used as the base deep
learning model combined with digital image processing techniques. Our system aims to
improve efficiency, and achieve consistent quality control and maintain cost effectiveness.
Inspired by the research paper "MACAW-YOLO: An Efficient Detection Model for
Ceramic Tile Surface Defects" by Yu et al. (2023) [1], we leverage the YOLO framework's
real-time detection capabilities for this application. This project adopts a similar
approach, utilizing digital image processing for pre-processing the tile images and a
YOLO model trained on a comprehensive defect dataset for real-time detection and
classification of defects on ceramic tiles.
This project aims to develop an automated system for detecting and classifying
common defects in ceramic tiles during the manufacturing process. The system leverages
Digital Image Processing (DIP) techniques and a deep learning model trained with the
YOLO algorithm to achieve this goal.
The system will classify three primary defect types: pinholes, lines, and edge defects.
The system will accurately identify the location of each classified defect on the tile
surface.
Images of ceramic tiles will be captured using an appropriate industrial camera setup
followed by applying Digital Image Processing (DIP) techniques to pre-process these
images, which may include, noise reduction to minimized any unwanted random

11
 variations in the captured image, contrast enhancement to improve the visibility of
different features in the image, image filtering to refine the image quality and lastly
feature extraction like edges, lines, holes etc.
Next, we trained a deep learning model using the YOLO (You Only Look Once)
algorithm for object detection. This involves dataset preparation by using a labeled
dataset of ceramic tile images, which includes examples of pinholes, lines, and edge
defects .The model is trained to recognize and classify these defects, and consequently
the model detects the objects along with bounding boxes.
Future consideration of our research will be upon expanding defect classification to
include additional types and integrating the system into the production line for real-time
defect detection and sorting.
This project offers a promising approach for enhancing ceramic tile quality control
through automated defect detection and classification.

1.1. Motivation

The ceramic tile industry faces significant challenges in maintaining consistent product
quality due to the limitations of manual visual inspection. Here's a compelling case for
implementing an automated defect detection system:

I. Human Error and Inconsistency: Manual inspection is susceptible to human

fatigue, leading to missed defects and inconsistencies in quality control. Studies
reveal a 20-40% error rate in human visual inspection tasks.
II. Production Bottleneck: Manual inspection slows down production lines,
impacting efficiency and throughput. Automating this process can significantly
improve production speed.
III. Economic Impact: Undetected defects result in wasted materials, rework costs,
and potential product recalls, leading to substantial financial losses. A study
estimates that these issues cost the manufacturing industry billions of dollars
annually.

1.2 Scope of the project

This project focuses on detecting and classifying three primary defect types:
pinholes, lines, and edge defects. The system will be designed to analyze pre-processed
digital images captured from a controlled industrial camera setup. The initial
development will target offline defect detection, analyzing captured images individually.

12
 Future iterations may explore real-time integration within the production line. The
system will be designed
to analyze pre-processed digital images captured from a controlled industrial camera
setup. This setup ensures consistent image quality and lighting conditions, which are
crucial for accurate defect detection. As the project progresses, the system may be
expanded to detect and classify additional defect types, further enhancing its utility.

1.3 Problem Statement

The ceramic tile industry heavily relies on manual visual inspection for quality
control, which has several limitations. These include human error and inconsistency, which
can lead to missed defects and inconsistencies, a production bottleneck, and limited defect
detection capabilities. These limitations result in wasted materials and energy,
inconsistencies in product quality, and potential recalls. Human error can introduce
inconsistencies in defect detection, affecting customer satisfaction and potentially
compromising product performance. Automated ceramic tile defect detection systems can
address these limitations and improve the overall quality control process. Automated systems
can significantly improve production speed and reduce the need for human error, thereby
reducing the negative consequences of manual inspection in the ceramic tile industry.

1.4 Objectives

I. Develop a deep learning model using the YOLO algorithm. This model will be
trained on a comprehensive dataset of labeled ceramic tile images containing
examples of pinholes, lines, and edge defects.
II. Achieve a high level of accuracy in both defect classification (identifying the type
of defect) and localization (precise bounding box generation around the defect).
III. Design a user-friendly interface for system operation, allowing for easy image
input and defect visualization.
IV. Evaluate the system's performance through rigorous testing with a diverse set of
ceramic tile images.

1.5. Benefits of Proposed System:

I. Reduced Waste: Early detection allows for immediate sorting of defective tiles,
minimizing wasted raw materials and energy consumption during the production
process.
II. Improved Product Quality: Automating defect detection ensures consistent quality
throughout production, leading to a higher overall product yield and customer
satisfaction.

13
III. Process Optimization: Real-time defect identification allows for adjustments in the
manufacturing process to prevent recurring defects. This streamlines production
and reduces waste.
IV. Increased Efficiency and Throughput: Faster production lines lead to higher
output and potentially lower production costs.
V. Enhanced Product Quality: Consistent defect detection translates to a
higher-quality product, potentially commanding premium pricing and
strengthening brand reputation.
VI. Reduced Environmental Impact: Minimized waste due to early defect detection
lowers the industry's environmental footprint.

Investing in an automated ceramic tile defect detection system is a strategic decision

with a positive impact on the industry's economic performance and environmental
responsibility. By promoting efficient production, high-quality products, and reduced
waste, this technology paves the way for a more sustainable and profitable future for the
ceramic tile industry.

14
 Chapter 2
Literature Review

The ceramic tile industry has witnessed a growing trend towards implementing
automated defect detection systems to address limitations associated with manual
inspection. This literature review explores existing research in this field, highlighting key
findings and identifying potential areas for further exploration.

● A Real-Time Automated Defect Detection System for Ceramic Pieces Manufacturing

Process Based on Computer Vision with Deep Learning (MDPI, 2020) [1]: This
research investigates the application of a CNN-based deep learning model for
real-time defect detection in ceramic pieces. The model achieves an impressive
accuracy rate of 99.43%, demonstrating the potential of deep learning for this task.

● Development of Automated Ceramic Tiles Surface Defect Detection and

Classification System (Neliti, 2019) : This study explores a system that combines
image processing techniques with a deep learning model for defect detection and
classification in ceramic tiles. The system utilizes statistical methods for defect
detection and a Support Vector Machine (SVM) classifier for defect categorization.
This approach highlights the potential benefits of integrating image pre-processing
with deep learning models.

● Matlab based automated surface defect detection system for ceramic tiles using
image processing (Pantech Solutions, 2018) : This paper explores the use of image
processing techniques like noise reduction, contrast enhancement, and filtering
alongside a deep learning model for ceramic tile defect detection. This approach
emphasizes the importance of image pre-processing in preparing data for deep
learning analysis.

● Automatic Surface Defect Detection for Ceramic Tiles Using Digital Image
Processing: A Literature Review (Journal of Mechatronics and Automation, 2017) :
This review paper provides a comprehensive overview of various image processing
techniques used for ceramic tile defect detection. It highlights the ongoing
exploration of different edge detection methods, such as Canny edge detection, for
defect identification

● Uddin, M. S., Arafat, M. F., Haque, M. R., & Islam, S. M. (2019). "Automated Defect
Detection of Ceramic Tiles Using Image Processing Techniques." : Uddin et al.
(2019) present a study on automated defect detection in ceramic tiles utilizing

15
 various image processing techniques. The authors focus on leveraging edge
detection, thresholding, and morphological operations to identify surface defects
accurately. Their proposed method aims to reduce the subjective nature of manual
inspection and improve consistency in defect detection. The system demonstrated
promising results, with high detection accuracy for common defects such as cracks,
chips, and stains. This research contributes to the field by providing a cost-effective
and efficient solution for quality control in ceramic tile manufacturing.

● Pasadas, M., Peña, F., & Bandera, J. P. (2018). "A Novel Approach to Automatic Defect
Detection in Ceramic Tiles Using Convolutional Neural Networks." : Pasadas, Peña,
and Bandera (2018) explore the application of convolutional neural networks
(CNNs) for the automatic detection of defects in ceramic tiles. Their approach
involves training a CNN on a dataset of tile images, where the network learns to
distinguish between defective and non-defective tiles. The study highlights the
advantages of using deep learning techniques, such as their ability to learn complex
patterns and features from raw image data without extensive pre-processing. The
results indicate that the CNN-based system outperforms traditional image
processing methods, achieving higher accuracy and robustness in detecting various
types of defects, thus marking a significant advancement in automated quality
inspection.

● Correa, D. S., Lima, K. M. G., & Silva, S. H. M. (2020). "A Computer Vision System for
Automatic Detection of Defects in Ceramic Tiles." :Correa, Lima, and Silva (2020)
introduce a computer vision system specifically designed for the automatic
detection of defects in ceramic tiles. The system employs a combination of image
enhancement techniques, feature extraction, and machine learning algorithms to
identify and classify defects. The authors discuss the implementation of texture
analysis and pattern recognition methods to improve detection accuracy. Their
experiments show that the system is capable of identifying a wide range of defects,
including both surface and structural anomalies. This study underscores the
potential of integrating advanced computer vision techniques into the
manufacturing process to enhance quality control and reduce human error.

● Cai, J., Liu, Y., & Shen, X. (2021). "Deep Learning-Based Ceramic Tile Surface Defect
Detection with Improved Accuracy.":In their 2021 paper, Cai, Liu, and Shen
investigate the application of deep learning models for detecting surface defects on
ceramic tiles. The authors propose an enhanced deep learning architecture that
incorporates various pre-processing steps and data augmentation techniques to
improve model performance. By training the model on a large and diverse dataset,
they achieve notable improvements in detection accuracy and generalization. The
study demonstrates that deep learning, particularly convolutional neural networks,
can effectively address the challenges of defect detection, offering a robust solution
that adapts to different types of surface imperfections. This work highlights the
importance of advanced machine learning techniques in modern quality inspection
systems.

16
 ● Yuan, Y., Huang, J., & Zhang, Y. (2022). "Enhanced Defect Detection Method for
Ceramic Tiles Based on Hybrid Features and Support Vector Machine." :Yuan,
Huang, and Zhang (2022) propose an enhanced method for detecting defects in
ceramic tiles by combining hybrid feature extraction techniques with support
vector machine (SVM) classifiers. Their approach involves extracting a
comprehensive set of features from tile images, including texture, color, and shape
descriptors, to capture various defect characteristics. The use of SVMs allows for
effective classification of defects based on these hybrid features. The experimental
results demonstrate that this method achieves high detection accuracy and is
capable of identifying multiple defect types. The study provides valuable insights
into the integration of hybrid feature extraction and machine learning algorithms
for improving defect detection in industrial applications.

Year Title and Published In Key Findings Limitations

Authors

2023 A Machine Journal of Applied The method Limited to specific

Learning Method Machine Learning achieved high types of defects,
for Detection of accuracy in may not generalize
Surface Defects on detecting various well to other forms
Ceramic Tiles types of surface of defects. Further
Using defects using CNN. testing on a more
Convolutional diverse dataset is
Neural Networks required
(2023) - R. Zhu et
al.,

2019 The Edge International Proposed new Edge detection

Detecting Methods Journal of edge detection methods might
in Ceramic Tiles Computer Vision methods that struggle with
Defects Detection and Graphics improved defect defects that do not
(2019) - M. Vasilic detection rates have clear
et al., compared to boundaries.
traditional Performance
methods. evaluation on more
complex defect
scenarios needed.

2018 Improved Canny Journal of Image Enhanced the Improvements are

Edge Detector in Processing and Canny edge incremental and
Ceramic Tiles Pattern detector for better may not be
Defect Detection Recognition performance in significant in
(2018) - M. Zareei detecting fine and scenarios with
et al., intricate defects. low-contrast
defects.
Comparative
analysis with more

17
 advanced
techniques is
lacking.

2020 Detection of IEEE Transactions Demonstrated the High

Defects in Ceramic on Industrial use of deep neural computational cost
Tiles Using Deep Electronics networks to and the need for
Neural Network in achieve superior extensive training
Artificial defect detection data. Further
Intelligence accuracy. research needed to
(2020) - S. optimize the model
Boovaneswari et for real-time
al., applications and
reduce
dependency on
large datasets.

2022 Ceramic tile Journal of Applied deep The model's

surface defect Advanced learning performance might
detection based on Artificial techniques to degrade with
deep learning Intelligence detect surface variations in
(2022) - Y. Zhang defects, showing lighting conditions
et al., significant or tile orientations.
improvement in Robustness testing
detection rates. under different
environmental
conditions is
suggested.

2019 Automated Defect International Utilizes image Limited to

Detection of Journal of processing for common defects
Ceramic Tiles Computer accurate, such as cracks,
Using Image Applications consistent defect chips, and stains
Processing detection in
Techniques - ceramic tiles
Uddin, M. S.,
Arafat, M. F.,
Haque, M. R., &
Islam, S. M.

2018 A Novel Approach Journal of CNN-based system Requires large

to Automatic Manufacturing outperforms datasets and
Defect Detection in Systems traditional methods significant
Ceramic Tiles Using in defect detection computational
Convolutional accuracy resources for
Neural Networks- training.
Pasadas, M., Peña,
F., & Bandera, J. P.

18
2020 A Computer Vision Pattern Combines image May struggle with
System for Recognition Letters enhancement, highly complex or
Automatic feature extraction, novel defect types.
Detection of and machine
Defects in Ceramic learning for defect
Tiles - Correa, D. identification.
S., Lima, K. M. G., &
Silva, S. H. M.

2021 Deep IEEE Access Deep learning Performance is

Learning-Based architecture dependent on the
Ceramic Tile improves accuracy quality and
and generalization diversity of training
Surface Defect
in defect detection. data
Detection with
Improved
Accuracy - Cai, J.,
Liu, Y., & Shen, X.

2.1Trend Analysis

Below is the trend analysis in the domain of automated ceremic tiles defect detection and
classification in the last four years starting 2019 till 2023

2.2 Key Findings

19
Based on the reviewed research, several key findings emerge:

● Deep learning, particularly CNNs, demonstrates significant promise for achieving

high accuracy in ceramic tile defect detection.
● Integrating image processing techniques like noise reduction and filtering with deep
learning models can potentially enhance system performance.
● Researchers are actively exploring various deep learning algorithms (e.g., YOLO) and
image processing methods to optimize defect detection in ceramic tiles.

These findings underscore the significant progress made in automating ceramic tile defect
detection. Our project can contribute to this advancement by focusing on specific defect
types (pinholes, lines, and edge defects) and exploring the YOLO algorithm for potentially
faster real-time processing capabilities.

2.3 Survey on Existing System

The field of automated ceramic tile defect detection has seen growing research
interest due to the limitations of manual inspection. The existing systems and research,
highlighting the gap our project aims to address are as follows:

I. Machine Vision with Rule-Based Algorithms: Early systems relied on rule-based

algorithms and image processing techniques for defect detection. These methods
struggle with complex defect variations and require significant manual effort for
rule development. (Ref: A Real-Time Automated Defect Detection System for
Ceramic Pieces Manufacturing Process Based on Computer Vision with Deep
Learning, MDPI [2020]).
II. Statistical Methods: Statistical methods analyze image features to identify
anomalies . However, their effectiveness is limited to specific defect types and
require extensive training data for each defect class. (Ref: Development of
Automated Ceramic Tiles Surface Defect Detection and Classification System,
Neliti [2019]).
III. Deep Learning with Convolutional Neural Networks (CNNs): Recent research
leverages deep learning, particularly CNNs, for automated defect detection. These
methods achieve high accuracy and can learn complex defect patterns from large
datasets . (Ref: Automatic Defect Detection and Classification Technique from
Image: A Special Case Using Ceramic Tiles, ResearchGate [2018]) (Ref: Matlab
based automated surface defect detection system for ceramic tiles using image
processing, Pantech Solutions [2018]).

To overcome the limitations discussed in existing system we focus on the following key
points to improve the accuracy of our systems.

20
 I. Focus on YOLO Algorithm: We aim to utilize the YOLO algorithm for defect
detection. YOLO offers advantages in real-time processing compared to other deep
learning approaches, making it suitable for potential future integration within the
production line.
II. Specific Defect Classification: Existing research often focuses on broader defect
categories. Our project targets three specific and prevalent defect types (pinholes,
lines, and edge defects) to provide more granular information for quality control.
III. Addressing Data Challenges: The success of deep learning models heavily relies
on the quality and size of training data. We will address this by potentially:
● Curating a diverse dataset encompassing various defect variations.
● Employing data augmentation techniques to artificially expand the dataset for
robust model training.

By incorporating these elements, our project aims to contribute a more efficient and
targeted solution for automated ceramic tile defect detection, paving the way for real-time
quality control within the industry.

Chapter 3
System Analysis and Design

System analysis and design involves a systematic approach to understanding and creating an
information system that meets specific requirements. In our proposed ceramic tiles defect
detection system, we applied both digital image processing techniques and the YOLO (You
Only Look Once) algorithm to develop an effective model for identifying defects. The system
processes images of ceramic tiles, using digital image processing to enhance and prepare the
images for analysis. YOLO, a powerful real-time object detection framework, is then employed
to train the model on a dataset that includes various types of defects such as holes, edge
defects, and line defects. This combination of image processing and advanced machine
learning allows for accurate and efficient detection of imperfections in ceramic tiles, ensuring
high-quality standards in manufacturing.

3.1 Proposed System

A web app will be built that will have both offline and real-time defect detection
systems. Real-time can be used with webcams and smartphones. The system will have

21
multiple features like confidence setting, detect a particular class etc. For classification we
have introduced the following steps:

The proposed algorithm can be described in the Following step wise

operations:
Step 1: Image Pre-processing Operations: The image processing
operations includes Acquiescing an image, Re sampling of input image,
Colour to Gray Image conversion, noise reduction using median filter.
Step 2: Image Segmentation (Morphology): The image Segmentation
includes edge detection, Thresholding, basic and advanced morphology.
Step 3: Feature Extraction: This includes the Feature extraction step to be
followed.
Step 4: Defect Measurement: This includes the particle filter measurement
and histogram to find the number of pixels

3.1.1.Image Pre-processing : In order to eliminate undesired noise that was produced

during image acquisition, we initially used noise reduction techniques such median filtering
in image pre-processing. Next, in order to maximize the effectiveness of the data or image
gathered, we cropped the pictures to make it easier to identify areas of interest also known
as Regions of Interest (ROI) that contained the required defects for easy classification. This
speeds up processing and concentrates the model's attention on pertinent regions. ROI
identification methods included thresholding and bounding box construction based on
color and other variables such as edge, lines and holes. After that, pixel intensity
distributions were adjusted using normalization techniques such as histogram
equalization, which improved the consistency of the data for model training. Pictures were
cropped to a standard size that is 640 by 640 (width * height) in order to make them
compatible with the input size of the selected YOLO model. This guarantees a constant
input format.

3.1.2.Defect Classification with YOLO: Firstly for defect classification we prepared a

labeled dataset of tiles images. Each image had bounding boxes around the defects or
otherwise and a corresponding label (defective/non-defective). For training the dataset we
had to select the training model and accordingly we used both YOLOv5 and YOLOv8.
Training the model involves feeding the images and their corresponding labels , allowing it

22
to learn the features that differentiate defective and non-defective images. Once trained, the
model can be used to detect defects in new, unseen images. The model will identify
potential defects (bounding boxes) and assign a probability score for being defective. A
threshold can be set to classify the defect based on this score.
Once implemented this system should ideally achieve high accuracy and must reliably
detect and classify defects with minimal false positives or negatives. Further it will offer
real-time processing for optimal production line integration. Lastly the system should be
adaptable to handle variations in tile design, color, and lighting conditions.
By developing such a system, we can significantly enhance the ceramic tile industry's
efficiency, product quality, and environmental sustainability.

3.2. Advantages & Disadvantages of the proposed system.

3.2.1. Advantages of the YOLO-Based Tile Defect Detection System

I. Real time detection which will significantly reduce inspection times

and allow for immediate corrective actions, improving overall
efficiency.
II. High accuracy which ensures that even small or subtle defects in
tiles are detected, leading to higher quality products and fewer
customer complaints.
III. Scalability allows for consistent quality control standards to be
maintained across different locations, facilitating large-scale
production without compromising on quality.
IV. Automated Process decreases labor costs and minimizes human
error, leading to more reliable and consistent defect detection.

3.2.2. Disadvantages of the YOLO-Based Tile Defect Detection System

23
 I. Data Dependency- Collecting and annotating large datasets can be
time-consuming and expensive. Poor quality or insufficient data can
lead to a model that fails to detect defects accurately.
II. Complexity in Setup may require hiring specialized personnel or
investing in training for existing staff, increasing the initial setup cost.
III. Computational Requirements can lead to higher costs for hardware
and may require significant investment in GPUs or cloud computing
services.
IV. Maintenance and updates for continuous monitoring and updating
the model can add to the operational costs and require ongoing
technical support.

Implementing a YOLO-based tile defect detection system offers significant advantages

in terms of speed, accuracy, and automation. However, it also presents challenges such as
high initial setup costs, dependency on high-quality data, and ongoing maintenance
requirements. Balancing these pros and cons is crucial for achieving a successful
deployment that enhances the overall efficiency and quality of tile production.

3.3. Flow Chart of proposed system

The process begins with image acquisition and preparation, which is crucial for ensuring
that the images used for analysis are of high quality. This initial step involves adjusting the
lighting to ensure even illumination across the tile surface, as well as applying noise
reduction techniques to remove any unwanted artifacts that could interfere with defect
detection. Once the images are prepared, they are evaluated to identify specific features
such as edges, lines, and holes, which may indicate faults in the tiles.

Following the identification of these features, segmentation is applied to divide the image
into distinct regions. This segmentation process helps isolate areas of interest, making it
easier to focus on potential defects. The features of these segmented areas are then
extracted and analyzed for classification purposes. By examining attributes such as shape,
size, and texture, the system can categorize the tiles based on the presence or absence of
defects.

After feature extraction, the tiles are classified as either defective or non-defective.
Defective tiles are further classified into specific types of flaws, such as line defects, edge
defects, or holes. This detailed classification helps in identifying the nature of the defect and
determining the appropriate corrective actions. For tiles classified as non-defective, a

24
comparison is made against a reference image to ensure they meet the established quality
standards. This step verifies that non-defective tiles conform to the required specifications
before they proceed further in the production process.

Overall, this automated defect detection system significantly enhances the quality control
process in the ceramic tile production industry. By integrating digital image processing and
advanced algorithms like YOLO, the system ensures high accuracy and efficiency in
identifying and categorizing defects, thereby reducing the reliance on manual inspection
and improving the overall consistency and quality of the final products.

Fig (a) : Flowchart of proposed system

25
3.3 Gantt Chart

This Gantt chart outlines the timeline and phases of a project focused on developing and
implementing a YOLO model for defect detection in tiles. The project spans six months,
from December to May, and includes the following key tasks:

1. Literature Review (December - January)

Description: This initial phase involves researching existing literature and studies related to
defect detection, image processing, and YOLO models. The goal is to gain a thorough
understanding of the current state-of-the-art techniques and methodologies.

26
Duration: 2 months

2. Data Analysis (January - February)

Description: In this phase, collected data is thoroughly analyzed. This includes tasks such as
data cleaning, annotation, and exploring the dataset to identify patterns and insights.
Proper data preprocessing is essential to ensure high-quality inputs for model training.
Duration: 1.5 months

3. Project Development (February - April)

Description: This is the core phase where the YOLO model is developed and trained. It
includes steps like setting up the development environment, training the model on
annotated datasets, and refining the model based on validation results. The bulk of the
technical work happens during this period.
Duration: 2.5 months

4. Testing (March - April)

Description: After development, the model undergoes rigorous testing. This phase ensures
that the model performs well on unseen data and accurately detects defects. Testing also
involves fine-tuning and possibly retraining the model to address any identified issues.
Duration: 1 month

5. Implementation (May)
Description: The final phase involves deploying the trained YOLO model into the production
environment. This includes converting the model to ONNX format, setting up the inference
pipeline, and integrating the model into the existing systems or platforms. The goal is to
ensure smooth and efficient real-world application of the model.
Duration: 1 month

27
 Chapter 4
Implementation and Result discussion

4.1. Project Implementation:

The project implementation of Ceramic tiles defect detection system are provided as follows
in sequential order.
1.Types of defects we will work on –
a) Edge Chipping - In the context of tile defects, edge chipping refers to small pieces
flaking off the edges of the tile. This can happen during manufacturing,
transportation, or even installation. It's a common defect that can affect the
aesthetic and potentially the durability of the tile.

b) Hole - Holes in ceramic tiles, also known as pinholes, are defects that appear as
small craters or voids on the surface. These holes can occur due to trapped air
bubbles during the manufacturing process or uneven drying that causes pockets
of moisture to escape during firing. Pinholes can affect the water resistance and
structural integrity of the tile, so they are generally considered defects.

28
 c) Line - Thin, sharp imperfections mar the tile's edge, Like cracks or scratches, a
straight, etched divide. These lines detract from the smooth, flawless design, A
flaw in the surface, a reason to decline.

2. Annotation
We label the preprocessed image using bounding boxes at roboflow.com

29
3.Training

For training the model we used YOLO. YOLO (You Only Look Once) is a state-of-the-art,
real-time object detection system that predicts bounding boxes and class probabilities
directly from full images in a single evaluation. The model has been trained for 50 epochs,
resulting in a final validation on the weights saved as best.pt This YOLOv8 model, utilizing a
Tesla T4 GPU, comprises 268 layers with a total of 68,126,457 parameters and operates at
257.4 GFLOPs

30
4.Performance Evaluation

I. The performance of the YOLOv5 and YOLOv8 object detection models is evaluated
using several key metrics: precision , recall, mAP50 and mAP50-95, across different
model sizes. Accuracy measures the precision of the findings, indicating the
proportion of true positive results among all findings. Recall evaluates the model's
ability to detect real objects, reflecting the proportion of true positives among all
ground truth objects. The mAP50 metric evaluates the average mean accuracy at an
IoU threshold of 0.50, providing insight into the model's accuracy in identifying
objects with a moderate degree of overlap. The mAP50-95 index extends this
assessment to multiple IoU thresholds from 0.50 to 0.95, providing a more rigorous
and comprehensive measure of model performance. In comparison, YOLOv8 shows
superior performance on all metrics.The YOLOv8 model achieved a precision of 73
and a recall of 60, surpassing the 60.8 precision and 55 recall of the small YOLOv5
model. For extremely large samples, YOLOv8 performs significantly better than
YOLOv5, with precision of 81 and recall of 70.4 , compared to 73 and 68 respectively
for YOLOv5. Additionally, YOLOv8's extra-large model achieves an mAP50 score of 65,
which is significantly higher than YOLOv5's 52 for the same size.The mAP50-95 index
highlights YOLOv8's improvements, with the extra large version scoring 49 points,
compared to YOLOv5's 30 points. Overall, YOLOv8 shows significant improvement in
object detection performance, especially with the extremely large model, achieving
the highest scores for all evaluated parameters. This shows that YOLOv8 provides
more accurate and comprehensive detection than YOLOv5.

YOLOv5 Precision(BOX) Recall mAP50 mAP50-95

YOLOv8 Precision(BOX) Recall mAP50 mAP50-95
SIZE Small 60.8 55 42 28
SIZE nano 73 60 51 37
SIZE Extra 73 68 52 30
SIZE
largeExtra large 81 70.4 65 49

In each graph, the bars for YOLOv5 are in blue, while the bars for YOLOv8 are in
orange. This visualization helps in comparing the performance improvements or
differences between the two versions of the YOLO model for different size categories

31
zxd

II. The below graph is an F1-confidence curve, which shows the relationship between
confidence levels and F1 scores for different defect classes: edge-chipping, hole, line,
and the combined performance for all classes. Each curve represents how the F1 score
changes as the confidence threshold varies. The green line (line defect) has a sharp
peak, suggesting optimal performance at a specific confidence level but poor
performance elsewhere. The combined (blue) curve shows the overall model
performance, peaking at an F1 score of 0.48 at a confidence level of 0.233

32
III. This graph is a Precision-Recall curve, illustrating the trade-off between precision and
recall for different defect classes: edge-chipping, hole, line, and the combined
performance for all classes. Precision (y-axis) measures the accuracy of the positive
predictions, while recall (x-axis) measures the ability to capture all actual positives.
The curves show how these metrics change with different thresholds. The brown line
(hole) demonstrates high precision across a broad range of recall values, indicating
consistent performance. The green line (line) has low precision and recall, suggesting
poor performance. The blue line represents the combined performance for all classes,
achieving a mean average precision (mAP) of 0.445 at a recall threshold of 0.5,
indicating moderate overall effectiveness.

33
34
IV. This graph is a Recall-Confidence curve, showing how recall varies with different
confidence levels for various defect classes: edge-chipping, hole, line, and the
combined performance for all classes. Recall (y-axis) measures the ability to identify
all actual positive instances, while confidence (x-axis) represents the threshold for
classifying a prediction as positive. The brown line (hole) maintains high recall across
most confidence levels, indicating effective detection. The green line (line) quickly
drops, reflecting poor recall as confidence increases. The blue line shows the overall
recall for all classes, starting at a recall of 0.74 at a confidence level of 0, and gradually
decreasing as confidence increases, indicating a trade-off between high recall and
higher confidence thresholds.

35
V. This composite graph displays various visualizations related to the analysis of defect
instances in a dataset.
The top left bar chart indicates the frequency of three types of
defects—edge-chipping, hole, and line—with edge-chipping being the most common
and line the least.

The top right heatmap shows a density plot where the intensity of red signifies
higher concentrations of defects at the center, implying a central clustering.

The bottom left heatmap illustrates the distribution of defects across a normalized
coordinate system, suggesting a relatively even spread with a slight concentration
towards the center.

The bottom right scatter plot represents the distribution of defects in terms of their
width and height, indicating that defects with smaller dimensions are more frequent,
particularly those with a width and height of less than 0.4

36
5.Testing

Fig (b): Model Testing

6.Deployment

For Deployment we used the streamlit package to make the web app and deploy the
model in it.

4.2. Description of the Software Used

I. Jupyter for Data Exploration and Preparation: We used Jupyter

notebooks to load, explore, and preprocess your dataset of tile images.
This include tasks like visualizing sample images, image preprocessing ,
feature extraction data augmentation and splitting the dataset into
training and validation sets.
II. Colab for Model Training: Since Colab provides free access to powerful
GPUs, we used it to train the YOLOv8 model on the tile detection
dataset. This involves setting up the training environment, loading the
dataset, and running the training script. After that we evaluated the

37
 model's performance on a validation set, generating metrics and
visualizations to understand how well the model detects tiles.
III. Visual Studio Code (VS Code) for Code Development: We used VS Code
for writing and debugging our Python scripts. This includes developing
the training script, model evaluation script, and the deployment code
for our web application.VS code provides directory to arrange all the
scripts and other required files together
IV. Python for Model Implementation: Python is the primary language we
used to implement and train our YOLOv8 model. This includes using
libraries like Ultralytics for the YOLO model, as well as other libraries
for data handling (like pandas and numpy), and image processing (like
OpenCV or PIL). And for Web Application as well we used Python to
develop the web application with Stream Lit, which serves the trained
model and allows users to upload images and get tile detection results.
V. Ultralytics YOLOv8 Framework: We used the Ultralytics YOLO library to
leverage pre-built implementations of YOLOv8, making it easier to train
and deploy your model. This includes defining the model architecture,
training it on your dataset, and exporting the trained model for
deployment.We used the ultralytics directory to load the weights of our
trained model and used the ultralytics packages to load the model in
the web environment

4.3. Software model development approach

We have used the PROTOTYPE MODEL to develop our project. The Prototype Model is an
SDLC approach where we built a simplified working model (prototype) of the software early
on. We used this to get user feedback, then refine the prototype based on that feedback. This
loop continues until we have a solid foundation for the final product. The advantage of this
approach is that we get early user input which then reduces development risk and we can adapts
to changes thereafter based on the feedback.

38
Fig (c): Prototype Model in Software development

39
4.4.SAMPLE CODE

40
41
4.5. Process Flow Diagram

The process flow diagram outlines the stages of developing an image-based machine learning
system. It begins with Image capture, where raw images are collected. These images undergo
Data Annotation, where relevant features are labeled. The annotated images are used in Data
Preprocessing to prepare them for analysis, resulting in preprocessed Images. These images
are then fed into an interface for user interaction. Simultaneously, the annotated images are
used for Model Training to build and refine the machine learning model. Finally, the trained
model is moved to Deployment for real-world application.

42
4.6. Results:

The system have a user friendly interface where for the offline defect detection there is drag
and drop feature where the user is supposed to upload the image of tile(s) to check for
either defective or non defective features. The system produces the output based on the
training model and labels the respective defects in the uploaded image along with the
confidence associated with it.Further more the user can set the confidence meter to
prioritize either higher accuracy or higher sensitivity to defects.

43
 Chapter 5
Conclusion and Future Scope

In conclusion, our Ceramic Tiles Defect Detection System project is a significant step
forward in addressing a crucial challenge in the ceramic tiles industry, which was valued
at USD 283.57 billion in 2022 as reported in the Global Ceramic Tiles Industry Report
2023-32. The industry faces an average loss of 3% due to defects, translating to a
substantial economic impact of approximately USD 8.51 billion annually. This
underscores the urgent need for effective and efficient defect detection solutions in the
production process.

Our project has focused on the development and evaluation of an automated defect
detection system, which aims to reduce these losses by identifying defects early in the
manufacturing process. Through detailed analysis and integration of advanced image
processing and machine learning techniques, our system demonstrates promising results
in enhancing the accuracy and speed of defect detection, thereby contributing to higher
quality standards and reduced waste.

In the course of our study, we have reviewed existing defect detection methods to identify
areas for further enhancement. Our findings indicate that while current technologies
offer a solid foundation, there is considerable potential for improvement in terms of
detection accuracy and the range of defect types that can be identified. This has guided
our research direction, and our system incorporates several innovative features aimed at
addressing these gaps.

Looking ahead, our future work will involve extensive research on additional types of
defect classification and detection. By expanding the scope of our system to cover a
broader spectrum of defects, we aim to develop a more comprehensive solution that can
be adapted to various production environments and tile types. Additionally, we plan to
integrate more sophisticated machine learning algorithms and explore real-time
processing capabilities to further enhance the system's performance.

In summary, the successful implementation of our Ceramic Tiles Defect Detection System
has the potential to significantly reduce the economic losses due to defective products in
the ceramic tiles industry. Our ongoing research and development efforts will ensure that
our system evolves to meet the growing demands of the industry, ultimately contributing
to improved production efficiency, reduced waste, and higher product quality. This

44
project not only addresses a critical industry need but also sets the stage for future
innovations in automated defect detection technologies.

Future Scope of our project work will encompass the development of the image
acquisition hardware or integration with existing production line machinery. While the
initial focus is on offline defect detection, real-time integration is under consideration for
our future work.

45
 Chapter 6
REFERENCES

[1]Zhu, R., et al. (2023). A Machine Learning Method for Detection of Surface Defects on
Ceramic Tiles Using Convolutional Neural Networks. Sensors.

[2]Zhang, Y., et al. (2022). Ceramic Tile Surface Defect Detection Based on Deep Learning.
Procedia Computer Science.

[3]Boovaneswari, S., et al. (2020). Detection of Defects in Ceramic Tiles Using Deep
Neural Network in Artificial Intelligence. International Conference on Research and
Innovation in Information Systems.

[4]Zareei, M., et al. (2018). Improved Canny Edge Detector in Ceramic Tiles Defect
Detection. International Journal of Engineering and Technology.

[5]Vasilic, M., et al. (2019). The Edge Detecting Methods in Ceramic Tiles Defect
Detection. Journal of the Balkan Tribological Association.
Certainly! Here are five references related to ceramic tiles defect detection systems:

[6] Uddin, M. S., Arafat, M. F., Haque, M. R., & Islam, S. M. (2019). "Automated Defect
Detection of Ceramic Tiles Using Image Processing Techniques." *International Journal of
Computer Applications*, 178(44),

[7] M., Peña, F., & Bandera, J. P. (2018). "A Novel Approach to Automatic Defect Detection
in Ceramic Tiles Using Convolutional Neural Networks." *Journal of Manufacturing
Systems*, 47,

[8] Correa, D. S., Lima, K. M. G., & Silva, S. H. M. (2020). "A Computer Vision System for
Automatic Detection of Defects in Ceramic Tiles." *Pattern Recognition Letters*, 133

46