Research on Damage Defect Detection Based on Computer Vision
Research on Damage Defect Detection Based on Computer Vision
Volume 9, Issue 12, December – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14613871
Research on Damage Defect Detection Based on
Computer Vision
Chhavi Rajput
PGCP from Indian Institute of Technology,
Guwahati in Artificial and Deep Learning
Abstract:- When an customer places an order online, I. INTRODUCTION
they expect a fast and accurate delivery. Customer
demand for a seamless experience from placing an order Operational development is a driving force of
to receiving an undamaged order in the hands. To automation and improving customer experience. Computer
provide this seamless experience to our customers, large vision based automated damage defect detection systems
level of industrialization is happening on the backend can help ecommerce enterprises achieve this goal by
from picking each product, scanning the barcode and improving customer experience, product quality and market
putting the order on the conveyor belt after packaging competitiveness by reducing the defect. At the same time,
and shipping the order at the right address. However, this process will reduce the cost of manual inspection and
automation comes with certain risks of mis-sortation of improve production. This damage defect detection computer
packages, damage defects during packaging the product, vision algorithm has the capability to identify the damages
barcode sticker alignment and the received product can by looking at the image coming from cameras and sensors.
be hampered due to liquid spillage, open damage box, Our world is moving fast towards computer vision and
uncovered tape and other factors. industries need advance computer vision based systems to
improve productivity and reduce packaging defects and
Therefore, this research is an effort to identify the enhance capabilities of future industrial development [1].
damages and defective products before delivering the
order. With the help of computer vision technology, This paper will explore the computer vision, image
cameras are placed on the top of each conveyor belt and processing methods to identify the damages and defective
camera will share the images at every 3-5 seconds and products, revealing its potential from implementation point
advance algorithm will be used to identify the defects or of view and prospects in ecommerce industry. Firstly, it will
damage packages. This paper will cover the computer review the image processing, computer vision and deep
vision algorithm along with image processing learning aspects that are part of this research of
normalization techniques to identifying the damages due identification of damage defect detection. In addition to
to human interaction and leading late deliveries and introducing the automated damage defect detection system
poor customer experience. based on computer vision, this paper also aims to stimulate
thinking about the future application of computer
Keywords:- Image Processing, Computer Vision, Defect technology. In this fast moving era, computer vision is also
Detection. helping the world to drive autonomous car driving by
identifying and segment the each object in the frame and
take the decision accordingly[2]. By merging
industrialisation with advance technology, it breaks the
traditional methodology and provides people seamless
experience of services and intelligent way of thinking.
Fig. 1 High Level Diagram of Damage Defect Detection Computer Vision Model
IJISRT24DEC2040 www.ijisrt.com 2616
Volume 9, Issue 12, December – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14613871
II. MODELS AND ALGORITHMS FOR DAMAGE useful features from the images and categorize them into
DEFECT DETECTION BASED ON pre-defined classes. ImageNet[3] is a large scale data set of
COMPUTER VISION images related to each category, this help researchers to use
pre-trained models directly and researchers can start their
In any industry, the biggest challenge is to transform work with small dataset. ImageNet contains 14 million
business problem into machine learning and AI problem images with 20000+ categories. Image classification[4]
statement. There are several examples in real time where involves categorizing or labelling images or specific areas
automation and ML can be applied. Framing an ML problem within them. This process offers a comprehensive
is the most creative and approachable way to solve any understanding of the image's content and can be carried out
machine learning problem. In this paper, our main focus is using either traditional methods or deep learning techniques.
to identify the defects and no-defects, to improve the Popular deep learning CNN architectures include Alexnet,
customer experience by not delivering the damaged product ResNet20, VGGNet16, and ResNet50. In this paper,
and identify the damage packages at first place. This will be experimentation based on transfer learning and ResNet50
an image classification problem, where binary classifier will pre-trained models will be presented. Since it has lots of
predict the defect and no-defect categories with probability advantages compare to other deep neural network
or confidence score. Image processing and image architectures, in term of training time, learning pattern etc.
classification is a core filed of computer vision and deep Transfer learning is a reliable source for experimenting with
learning. It gives an advance opportunity of rapid small dataset of images and transfer learning models are
development towards deep learning and machine learning. already trained on large dataset to give the desire output by
Advance image processing algorithm helps to read the tuning hyperparameters and architectural changes.
Table 1 Experimentation on Different Pre-Trained Models
Model Name Model Training Results Model Metrics
AlexNet Epoch 30/30 Specificity: 75%
[==============================] - 75s Sensitivity: 87%
709ms/step - loss: 0.5612 - accuracy: 0.7065 - ROC-AUC: 0.81
val_accuracy: 0.8133 Accuracy: 81%
Balanced Accuracy: 81%
Matthews Correlation Coefficient: 0.63
VGG16 18/50 [==============================] - Specificity: 0.86%
95s 902ms/step - loss: 0.7819 - accuracy: 0.9440 - Sensitivity: 0.85%
val_accuracy: 0.8504 ROC-AUC: 0.86
Accuracy: 0.86%
Balanced Accuracy: 86%
Matthew Correlation Coefficient: 0.72
ResNet50 with Image With 3 Dense and 2 Drop out layers106/106 Specificity: 87%
Augmentation [==============================] - 67s Sensitivity: 89%
634ms/step - loss: 0.0378 - accuracy: 0.9882 ROC-AUC: 0.88
val_accuracy: 0.8628 Accuracy: 88%
Balanced Accuracy: 88%
Matthew Correlation Coefficient: 0.77
III. IMAGE AUGMENTATION robust. In our case, the models will be highly accurate for
No Defects, but less accurate for all the different types of
Training a model on large datasets presents a defects. In this use case, image augmentation would be
significant real-time challenge in deep learning. Deep optimal approach to handle imbalance dataset. Image
learning and computer vision tasks have achieved augmentation involves creating additional examples from
remarkable results thanks to the use of convolutional neural the original dataset, but with slight alterations such as
networks (CNNs). However, diverse dataset for training has flipping images vertically or creating mirrored images. This
become a necessary part for computer vision models. Image will increase the number of images in the underrepresented
Augmentation [5] is a process to diversify the data based on classes and yield a more balanced dataset. Image
the existing dataset. Computer vision models rely on a large augmentation helps this experiment to augment the images
amount of data for training, but for certain types of images by rotating the image on different angles, along with
there aren’t enough examples to provide a robust cropping and blurring technique. Image augmentation helps
representation of the data. For example, if we are training a this model to increase 12% specificity and 2% sensitivity
model to identify images of defects in products, there are from AlexNet to ResNet50 with image augmentation. Model
many images of Empty and No Defect cases but relatively is capable to identify the 82% (441/536) of missorts defect
few images of different types of defects. Let’s say the split is and 61% (464/750) of the overall defects with 95% high
95% No Defects, and 5% Defects. This means that the confidence and 91% precision rate.
training process may lead to an inefficient representation of
the data, resulting in models that are highly specific and less
IJISRT24DEC2040 www.ijisrt.com 2617
Volume 9, Issue 12, December – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14613871
IV. CONCLUSION
This paper introduces a computer vision model for
damage and defect detection, designed to enhance the
quality and packaging processes in e-commerce fulfilment
centres by identifying damaged products. Leveraging three
key technologies — image augmentation and pre-
processing, computer vision, and defect recognition and
classification — the model operates with high efficiency and
accuracy to detect mis-sorted items, label defects, and other
damages, with the ultimate goal of achieving real-time
monitoring and automatic detection on production lines.
This approach significantly boosts detection efficiency and
accuracy, while minimizing human intervention and the risk
of misjudgement. At the core of the system are computer
vision techniques and deep learning algorithms, which play
a crucial role in enabling precise defect detection. Although
many deep learning models currently offer promising
results, providing the required detection accuracy and speed
for industrial production lines, challenges remain,
particularly around privacy and security concerns. However,
it is anticipated that these issues will be addressed in the
future. With the continued development of advanced
models, the system's detection accuracy, real-time
performance, and robustness will improve, broadening its
applications and prospects for future growth.
REFERENCES
[1]. Zhonghe Ren1 · Fengzhou Fang1,2 · Ning Yan1 · You
Wu1 . State of the Art in Defect Detection Based on
Machine Vision
[2]. Senay Cakir, Marcel Gauß, Kai Häppeler, Yassine
Ounajjar, Fabian Heinle, Reiner Marchthaler Semantic
Segmentation for Autonomous Driving: Model
Evaluation, Dataset Generation, Perspective
Comparison, and Real-Time Capability
[3]. Olga Russakovsky, Jia Deng, Hao Su, Jonathan
Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael
Bernstein, Alexander C. Berg, Li Fei-Fei - ImageNet
Large Scale Visual Recognition Challenge
[4]. Òscar Lorente, Ian Riera, Aditya Rana - Image
Classification with Classic and Deep Learning
Techniques
[5]. Suorong Yang, Weikang Xiao, Mengchen
Zhang, Suhan Guo, Jian Zhao, Furao Shen - Image
Data Augmentation for Deep Learning: A Survey
