10 7.
Destroy Model – The test image is deleted or is overwritten after the results have been obtained. This is important in reducing memory usage.
Region is one of the basic frameworks in determining object recognition. As proposed by Gu et al., 2009
in “Recognition using Regions” stated that each region created, it represents by a rich set of image cues shape, colour and texture that accounts
as a major target in image processing for recognition object. In this paper, the objective is to provide a region database of image by introducing Generalized Hough transform in
detection and segmentation to generate hypothesis of object locations, scales and support, followed by verification classifier. This approach building on a novel region segmentation
algorithm which produces robust overlaid regions. By proposed this approach shown that by creating a specific region able to recognize object robustly but proper region creation
must be applied. As a result, there are two conclusions obtained 1 cue combination significantly boosts recognition performance; 2 region-based voting scheme reduces the
number of candidate bounding boxes by orders of magnitude over standard sliding window scheme due to robust estimation of object scales from region matching.
1.3.2 Defect Inspection System
In industry application, there is not much research on recognizing defects in gluing application mostly on applying it automation industry for sealing the car windscreen. But
there are a lot of applications that have same features of inspection system such welding. A research entitled ‘A Method for Recognition of Defects in Welding Line’ Peng, 2009
uses the feature extraction method in auto-recognize the defects in welding line. First, the digital images of the welding line must be binarized and extracted all the necessary
features of all defects for training and modelling in the training phase to figure out each
11 kind of defects. The algorithm is tested with 500 welding defect consists of all six type of
defect to evaluate the defect detection capabilities. The results of this experiments represents defect detection of each defect’s type but overall this system able to generate
94.3 success rate in classifying defect’s type. 6 Types of defects are identified through this research:
1. Air Hole – Usually are black, round and smooth in boundary and their locations
and sizes are uncertain. 2.
Crack – Cracks are black, in slightness and needle shape at least at one end, and usually parallel or vertical with welding lines.
3. Strip Inclusion – Strip inclusion are black, in strip and complex in boundary.
4. Round Inclusion – Round inclusion are black, round and complex in boundary.
5. Lack of Penetration – They are black, linear, and straight in both sides.
6. Metal Inclusion – They are bright in defects in the welding line images and their
locations, forms and directions are uncertain. Welding defect technologies is one of the most famous applications. The paper
entitled ‘Weld Defect Detection Based on Gaussian Curve’ Li and Liao, 1996 developed based on the assumption that a line profile of defect less weld image can be approximated
by a Gaussian distribution curve. The line profile variations of weld image caused by defects are classified into three defect patterns, defect-peaks, defect-troughs and defect-
slant-concaves. Weld defects are characterized by a local variation in material density, which manifests itself as a local intensity variation in the resulting radiographic images.
According to the results, the proposed method capable in achieved successful defect detection rate is 95 percent where 24 images tested consists of 61 defects.
12 Nowadays, there is too many techniques involved in applying inspection system in
surface. According to research entitled as ‘A Review of Recent Advances in Surface Detection using Texture analysis Techniques’ Xie, 2008 described all the techniques used
and its defect application in surface. The surface defects are loosely separated into two types; first is local irregularities which are the main concern for most visual surface
inspection and the other is global deviation of colour and texture. The techniques being discussed in this paper used to inspect textured abnormalities are break to four categories,
structural approaches, model based approaches, statistical approaches and filter based method. Every method has its own benefits and disadvantages. The most important thing in
providing better services is to understand the formation and nature of the defects. It is important to be able to accurately localise the defective regions rather than classifying the
surface as a whole. This can provide possibilities of classifying the defects and further studies of the characteristics of the defects.
In developing an automatic detection, there are a lot of filtering techniques available to that. One of that is Texture Features. Title research of ‘Automatic Detection of
Welding Defect using Texture Featur es’ are proposed Mery and Berti, 2003 based on two
types texture features: 1 features based on co-occurrence matrix, which gives a measurement of how often one gray value will appear in a specified spatial relationship to
another gray value on the other image; 2 feature based on 2D Gabor functions. The proposed method follows the pattern recognition schema: image segmentation, feature
extraction and classification. Image segmentation process is oriented towards the detection of edges by employing the LoG filter. Feature extraction main contribution on the
measurement of properties of the region and image classification divides segmented region into special region according to the extracted features. According to the results obtained, it
can be concluded that best texture features based on the co-occurrence matrix are the mean
13 of the difference entropy and the mean of the difference variance, for a distance o
d
= 3. On the other hand, the best Gabor features are at
p
= 6 scale. Other researchers are using optimised filter in classified texture defect. As shown in
paper of ‘Optimised Filters for Texture Defect Detection’ founded by Sobral, 2005 present a new approach to texture defect detection based on a set of optimised filter. The
main purposed in introducing this method because it is applied to one wavelet sub-band and its size and shape are tuned for a defect type. The wavelet transform provide an
excellent framework to deal with multi-scale defects. The first level of the wavelet packet decomposition divides the power spectrum into four square parts and the second level
divides the same spectrum into sixteen parts and the third level in sixty four parts. Therefore the filter applied in higher wavelet those results in easier to detect defect by
analysing only a small part of the spectrum. The result showed that this masks capable to detect most of the defect and achieved an acceptable detection rate.
Inspection system are not just specific to welding and surface, texture of materials also one of the important subject in using defect inspection system. Textured material can
be classified as textile, web, paper and wood. One of the researchers done ‘Defect
Detection in Textured Materials Using Gabor Filters’ by Kumar and Pang, 2002 proposed Gabor Filter that provides features of decompose image into components
corresponding to different scales and orientations. Besides that, it also achieves optimal joint localization in spatial and spatial frequency domain. In the spatial domain, Gabor
function is a complex exponential modulated by Gaussian function. It forms a complete but nonorthogonal basis set and its impulse response in the two-dimensional plane. As a result,
the proposed of multichannel filtering scheme successfully segmented the defects of varying size, orientation and resolution that proved to be robust.
14 In textile industry, there are so many problems involving defect inspection. Many
fabric defects are very small and undistinguishable, which make the very difficult to detect. According to Chan and Pang, 2000
entitled as ‘Fabric Defect Detection by Fourier Series Analysis’ by implementing Fourier series analysis to monitor the spatial frequency
spectrum of a fabric. If defect occurs in fabric, the regular structure of fabric changed so that the corresponding intensity at some specific positions of the frequency structure would
change. In this paper, the researchers classify defect as four classes: 1 double yarn; 2 missing yarn; 3 webs or broken fabric; and 4 yarn densities variation. Central spatial
frequency spectrum approach is proposed that would reduce computational time for defect detection and provide more parameters for defect classification. According to the results,
the method proposed is able to extract seven significant parameters that capable for classifying defects in fabric.
In the area of Machine Vision inspection system, there are a lot of different approach has been studied intensively and plenty of machine vision software is available as
was stated. In all the above research, the machine vision system and software used has two common similarities, first is the three basic framework of the process involved; image
acquisition, preprocessing image enhancement and image analysis; second is the two phases required for shape matching, the training phase and the recognition phase. In the
above research, it seems that most model based vision programs are developed for a specific task and the environment is implicitly coded into the system. It is difficult to
modify the knowledge or extend the scope of such system, and it also requires long development time.
15
1.3.3 3D Recognition Camera Placement