Automated inspection system pdf

According to metallography, struc- tures which are coarse enough to be discernible by the naked eye or under low magni- fications are termed macrostructures, while those which require high magnification to be visible are termed microstructures. Even if useful information can often be gained by examination with the naked eye of the surface of metal objects, microscopes are required for the examination of the microstructure of the metals.

Optical microscopes are used for resolutions down to roughly the wavelength of light about half a micron and electron microscopes are used for details below this level, down to atomic resolu- tion. Particular features of interest are: grain size, phases content, distribution of R.

Wilson et al. The examination of mate- rials by optical microscopy is essential in order to understand the relationship between properties and microstructure. While the traditional approach involved the direct ob- servation of the acquired images by human experts, resulting in a qualitative analysis of the results, with the diffusion of digital image processing techniques the analysis process became faster, simpler and more precise.

Computer Vision techniques have been used in Metallography Image Analysis[2,3] to study the properties of sintered- steel and the nickel-based superalloy [4], for the automated classification of heat re- sistant steel structures[5], for segmenting the phases of high strength low alloy steel [6], to study the pit formation on a titanium alloy [7], and for the segmentation of a of a two-phase Ti—6Al—2Mo—2Cr—Fe titanium alloy[8].

Tejrzanowski et al. In our work we proposed an automated inspection system to study the properties of a titanium alloy, and in particular of its microstructures, in order to classify the parts of the inspected material into different mechanical-physical phases.

The paper is or- ganized as follow: section 2 describes the features of the titanium alloy and one of its most important welding techniques; section 3 presents our classification system; sec- tion 4 discuss our experimental results; a conclusive section ends the paper.

Today, titanium and its alloys are extensively used for: aircraft engines and airframes, spacecraft, chemical and petrochemical production, power generation, nuclear waste storage, navy ship components, automotive components, food and pharmaceutical processing, medical implants and surgical devices. An im- portant aspect of this kind of alloy is its microstructural evolution as function of the thermo-mechanical history, depending on the applied manufacturing process.

The formation and the behavior of each phase is linked to the addition of alloying elements called stabilizers to titanium, which enables physical-chemical effects on the creation of the single microstructural type. The most commonly used titanium alloy is Ti-6Al-4V, that is the object of study of our work. It is significantly stronger than commercially pure titanium while having similar thermal properties.

Among its many advantages, it is heat treatable[11]. Ducato et al. Ti-6Al-4V phase diagram 2. The tool is rotated and plunged into the material so that the shoulder works on the plate surface and the probe is buried in the workpiece.

The friction between the rotating tool and the plate material generates heat, and the high normal pressure from the tool causes a plasticized zone to form around the probe.

The tool is then traversed, frictionally heating and plasticizing new material as it moves along the joint line [13]. Although the majority of common titanium alloys are gener- ally weldable by conventional means, problems with workpiece distortion, and poor weld quality, can occur. The development of FSW offers the possibility of a new method of producing high quality, low distortion, welds in Ti sheet and plates. The parent material was found to consist of a rolled microstructure of elongated gains of alpha light in a matrix of alpha and beta dark.

In the deformed weld zone, the microstructure shows evidence of Alpha-to-Beta phase transformation. The beta phase reverts on cooling, and the resultant weld microstructure consists of large alpha grains with a smaller amount of retained beta. The weld root zone microstructure in this case shows that only partial transfor- mation has occurred in this region.

After FSW, the preparation of a specimen to re- veal the microstructure of the welded material involves the following steps: sawing the section to be examined, mounting in resins, coarse grinding, grinding on progres- sively finer emery paper, polishing using alumina powder or diamond paste on rotat- ing wheel, etching in dilute acid, washing in Alcohol and drying. The specimen is then ready to be inspected by microscope.

In this paper we compare the results obtained with several low level features descriptors, using a common testing frame- work. The scheme of the overall system is shown in fig. The value of B will be further de- scribed in the experimental section. In our work we analyzed a set of texture descriptors, which are briefly described in the next sec- tion.

Color information, in this case, is useless as the images in the dataset are, in practice, monochrome. SVM is the most used and the simplest solution whenever a binary clas- sification problem has to be solved, therefore is well suited for our goals.

Informa- tion about the SVM setup will be given in the experimental section. Each feature is used to train separately a classifier.

In this case a simple adaptive threshold method is applied. For each block all the pixels whose grey values are above the average value of the block are labeled as Alpha lightest areas , and the other ones as Beta darkest areas. Two constraints a mini- mum and maximum value are imposed to this threshold value, to treat also the rare case in which all the pixels of a block are of the same class i. We preferred to use an adaptive threshold approach, rather than a global one, as it works also in case of not uniform illumination during the acquisition of the image.

It can be considered as a measure of the perceived image surface variations. Output is a 4-dimensional vector. Output is a 7-dimensional feature vector. Our machines are:. Depending on what you manufacture, defects can be expensive to detect. Complexity adds increased costs in inspection time, specialized equipment and training specialized personnel.

The systems developers at Coleman-Sciotex are at the forefront of industrial software development. This is an area that will make us stand out from off the shelf systems. Knowing how to evaluate your complex custom manufacturing process and create the custom software to actually make an automatic inspection system work requires expertise in the following subjects.

The intellectual pursuit of top quality systems development is always an amazing experience. We really enjoy the collaboration that comes from building long-term trusted partnerships with our clients. We create the significant competitive advantages that help our clients to be market leaders.

However the segmentation suffers from inaccuracy because paper-like colours exist in the cigarette and they are segmented as parts of background. The other obstacle to identification of individual objects is due to contact between objects. These connection inaccuracies should be reduced before further processing. We use morphological operations to reduce them by improving the binary images. Morphological operations are methods for processing binary images based on shapes.

As binary images frequently result from segmentation processes on grey level images, the morphological processing of the binary result permits the improvement of the segmentation result. These improved images therefore increase the accuracy of object identification.

The morphological operation is performed in two steps: i The shrink operation is used to remove pixels so that objects without holes shrink to a point, and objects with holes shrink to a connected ring halfway between each hole and the outer boundary. Step2: Label the binary image Contiguous regions are labeled to be identified as an individual object.

The kth region includes all elements in the labeled image L that have value k. For example, the pixels labeled 1 make up one object, the pixels labeled 2 make up a second object, and so on. Therefore, the number of objects in L is equal to max L. Therefore, the joined objects should be separated to be an individual object.

We use K-mean classification to separate them. If K is greater than 1, the object is re-clustered into K clusters see Figure 4. The final image includes only the cigarette objects, and the number of objects in the image and the centre points of each object are obtained from this final image. This information is very useful for the classification of defective cases. It is made of paper and forms part of the packaging inside the cigarette tin.

It is located on the circumference of the cigarette package and when viewed from above in an image, it appears as a semicircular object see Figure 6-d. If the spoon is missing then the cigarette package should be defected. It is therefore important to also recognize from an image the presence of a complete paper spoon component in the package.

Normally the spoon is on the cigarette package circumference on the extreme right of the image. However, due to movement in the manufacturing process, while it will always be on the circumference of the cigarette package, it may be displaced slightly in the image.

If the location of the spoon is always the same in images, then simply applying a semi- circular template of the spoon as a window will identify its presence. Therefore, we need to develop a new method to detect images that do not display a correct paper spoon so they can be classified as defective packages. Create a mask a Grayscale image, I b threshold, level 0. Segment a paper spoon Figure 7.

Create radar beams Step 1: create a mask image First, we segment the boundary of a tin by the threshold of the maximum intensity Figure 5. The Hough transform space of this threshold image is able to indicate the center cx, cy of a tin.

An average distance between the initial tin boundary and the centre is the approximation of a radius of the tin. A circle with this centre and radius is taken as a mask see Figure 5. Step 2: segment a paper spoon We segment a paper spoon from the grayscale image by the threshold Figure 6.

The initial area including paper spoon in the image can be segmented with luminance greater than 0. The paper spoon is segmented by masking with the image mask created in Step 1 see Figure 6. The radar beam is a linear scan radiating from the estimated center of the mask. As Figure 7 shows, when the red beam cross the paper spoon image, it obtains the width of the paper spoon handle from the image.

If the width is less than 20 pixels then it is rejected as noise and the radar beam is not included in the valid set of lines. If there are more than 20 valid lines, the image has a paper handle. Experimental Results We tested our method on good cases and defective cases. The defective cases include 11 images without the paper spoons defective cigarette packages that should be rejected.