Defect detection and recognition based on ADABOOT-
SVM integrated model
ZhiKai Liang
Beijing Institute of Graphic
Communication
Beijing No. two, Xinghua street, China
13051379003, 102600
Liangzhikai0812@gmail.com
ShaoZhong Cao
Beijing Institute of Graphic
Communication
Beijing No. two, Xinghua street, China
13520127931, 102600
chaoshaozhong@gmail.com
YuKun Tan
Beijing Institute of Graphic
Communication
Beijing No. two, Xinghua street, China
18801017073, 102600
tanyukun@gmail.com
ABSTRACT
As the core component of printing machinery, the surface finish
and geometric accuracy of printing drum will have an important
impact on the quality of printed matter. However, the use of acid
ink, alcohol and other chemical raw materials corrode the drum,
leading to local collapse or spots. How to effectively identify the
types of drum defects has become an important issue. To solve
this problem, a defect detection and recognition framework based
on adaboot-SVM ensemble learning model is proposed. The
framework is composed of two parts: feature extraction and
classifier design. The first part is feature extraction from
directional gradient histogram (HOG). In the second part, we
construct an ensemble of different SVM classifiers to identify
defects. The validity of the proposed model is verified by nine
different defects. The results show that the integrated model of
adaboot SVM is helpful to improve the recognition accuracy of
defects.
Keywords
Defect recognition; ensemble learning; multiple classifier sets;
SVM
1. INTRODUCTION
As an important branch of machine vision technology, visual
inspection is a hot research direction in the field of product
nondestructive testing in China. Data shows that the domestic
market in 2015 has reached 350 million US dollars, accounting
for 8.3% of the global market, growth rate is 22.2%, ranking first
in the world, China has become the world's third largest machine
vision market after the United States and Japan. From 2016 to
2020, the growth rate of China's machine vision market is
expected to remain above 20%, and will reach a billion dollar
market space. Visual sensor has many advantages, such as large
amount of information, non-contact with workpiece, high
sensitivity and accuracy, strong anti-electromagnetic interference
ability, and so on. It is a hot research field. As the core component
of printing machinery, the surface finish and geometric accuracy
of printing drum will have an important impact on the quality of
printed matter. However, the use of acid ink, alcohol and other
chemical raw materials corrode the drum, leading to local collapse
or spots. How to effectively identify the types of drum defects has
become an important issue.
Scholars at home and abroad have studied and proposed a variety
of new feature extraction algorithms from the analysis of feature
extraction, machine learning, depth learning and other methods,
and achieved good recognition results. However, there are many
kinds of defects, and some defects are very similar. The selection
of effective feature combination usually depends on the
experience of experts. The accuracy of recognition is difficult to
guarantee. Therefore, it is of great significance to find a data-
based self-learning method to improve the performance of defect
recognition. With the rapid development of artificial intelligence
technology in recent years machine learning and its application
had a hot topic in the field of artificial intelligence. At present,
machine learning has achieved good results in some pattern
recognition fields [15-19], such as speech recognition, image
classification, natural language processing and so on. At the same
time, scholars at home and abroad are committed to introducing
machine learning into defect recognition [20-24]. Therefore, this
paper uses the gradient histogram of Oriented Gradient proposed
in reference [25] to extract the features of defects. At the same
time, considering that the idea of ensemble learning can be used to
construct an effective combined classifier model, it can be used to
replace the soft Max classifier used in most depth learning
applications. Based on this, this paper proposes a defect
identification method combining HOG and ensemble learning.
The specific process of the model is shown in Figure 1. First,
noise reduction is done in the preprocessing stage. Then, we
create the defect features extracted from HOG images. Finally, in
the classifier design stage, a multi-SVM linear combination
classifier (MSVMLC) is constructed for classification and
recognition.
2. Histogram of OrientedGradients
Directional gradient histogram (HOG) is an image descriptor for
object detection which is widely used in computer vision and
image processing. This method uses the histogram of gradient
orientation (HOG) feature to express the detected object, extracts
the shape information of the detected object, and forms a rich
feature set.
Compared with other descriptors, HOG descriptors have some key
advantages. Because it runs on a local element, it is invariant to
geometric and photometric transformations except for better
capturing local shape information, which only occurs in larger
spatial regions. Moreover, the HOG is obtained in a densely
sampled image block, and the spatial position relationship