Abstract

In automotive industry there is an increasing need for a hundred percent automatic non- destructive X-ray inspection for quality assurance of castings. A new fully automated analysis and classification method for automatic defect recognition (ADR) in X-ray images is presented and compared to traditional methods. It is shown, that the new method – called “Trained Median Filter” (TMF) - overcomes all the weaknesses of established methods. Even flat defects in structured areas that tend to decompose are found reliably. In conclusion a performance estimation demonstrates the real-time capability in a batch testing system.

Introduction

Today aluminum castings are used in various fields of applications. Especially the automotive industry uses them to make their cars lighter and more energy efficient. Here safety critical parts have to be inspected hundred percent with X-rays to avoid defects in the delivered castings.

Because of their reliability and efficiency “Automatic Defect Recognition” (ADR) systems replace the human inspector at this job. These ADR systems use modern image processing methods to evaluate the X-ray images and classify the found defects. New technologies like powerful industrial PCs and flat panel detectors (1) lead to more performant algorithms. Due to the high sensitivity of these modern flat panel detectors, defect types became visible, which were not detectable before.

The biggest challenges for ADR systems are

• very small defects with very low contrast in the X-ray image,
• big low intensity defects that “hide“ behind the structure of the part.

Also these defects should be detected with a high reliability.

This study is divided into two parts. In the first part we will give a survey of ADR systems (i.e. human selected filters, neural network, and edge preserving median filter). To be concrete the PXV, the AI, and the ISAR (2) are well-known systems used today. Similarities as well as differences are shown and the dedicated methods are confronted with the above defect types. A short conclusion of this part lists the restrictions of the traditional methods. In particular the restriction to the kernel size is one major difficulty because of the local filters. A non-local filter approach (3) can be used to overcome this.

In the second part a new fully automated analysis and classification method for ADR is presented. It will be compared to the well-known methods mentioned above. It is shown that the new image-processing algorithm called the “Trained Median Filter” (TMF) is ideally suited to find both defect types reliably and fast. With the aid of the “unsupervised training” all parameter adjustments become automated, which makes the TMF algorithm easy to use even by non-experts.

In a subsection a performance estimation of the TMF (including test results) is given to demonstrate the real-time capability in a batch testing system. For illustration purposes some X-ray images of the segmented defects are presented.

The conclusion gives a summary of both parts and ends up with a short outlook. In the following part we start with the survey of ADR systems.

I. Survey of ADR systems

In general ADR systems for aluminum castings follow one main strategy for defect segmentation in X-ray images:

• process a flawless reference image from the actual captured original X-ray image of the test object,
• build a difference image by subtracting the processed reference image from the original image,
• create a binary image from the difference image by defining a threshold and verify that the so marked objects are real defects.

Though the known methods have these obvious similarities they differ in detail. For example one question is which pre-processing has to be done and how to process the reference image afterwards. Therefore diverse image processing routines have been developed.

The ‘Golden Image’ approach (i.e. having one fixed reference image) has shown to be insufficient: A difference image of a fixed reference image and original X-ray image will lead to undesirable artifacts due to the tolerances of positioning test objects, ridges of castings or special features of the different moulds for one casting.

System/ Function	Human selected filters (e.g. PXV)	Local image restoration (e.g. ISAR)	Neural Network (NN) filtering (e.g. AI)
Prior-knowledge	By human (expert)	None	Trained by NN (e.g. Hopfield-Tank)
Image enhancement (pre-process)	Included	Included	Included
Kind of training	Supervised	None	Supervised
Training period	1-2 months at runtime	None	Training at setup time
Mean setup time	4h per position	2-3h per position	2-3h per position
Online filtering	Selected special image processing filters for each of several ROIs	COMMED, an edge preserving median filter	Hopfield-Tank NN, with given parameter set for each ROI
Image segmentation	Difference of reference and original image	Difference of reference and original image	Difference of reference and original image
Pseudo / real defect	Very low	High	Medium
Defect verification	Negative list	Positive list	Negative list
Kind of classification	Good/Bad	Good/Bad	Good/Bad
Table 1: Comparison of today’s ADR systems.

To overcome these problems some systems use prior-knowledge of the test object or even try to restore the original X-ray image without prior knowledge. The advantages and disadvantages will be discussed below. In any case processing of the reference image has to be done in real-time, whereas the analysis of test objects in order to gain prior- knowledge is not a time critical procedure.

For all systems it is mandatory to mark all defects at least in the difference image. Afterwards the systems differ again regarding their way to minimize the number of pseudo defects. Today there are two methods in order to distinguish between mistaken and real objects of defects. The first method uses a so-called ‘positive list’. This means that all detected objects are compared with the templates of expected or even known defect types. In this case no more information about the test object is needed and only familiar defects will be detected. Therefore the remaining pseudo to real defect rate is naturally high. The second method uses a so-called ‘negative list’: Prior-knowledge of the test object is reused to decide whether the found defect is a real one or not. Hereby it is also possible to detect not-known defect types. This leads to a significant lower pseudo to real defect rate than for the first method. This rate can become very low in the case of well-adapted prior- knowledge (see Table 1).

In the following section we will compare three concrete ADR systems exemplarily.

Comparison

To compare their functionality three different but well-known ADR systems are listed in Table 1. They are used today in the automotive industry. The systems can be divided into methods without and with prior-knowledge.

The ISAR (Intelligent System for Automatic Roentgen-inspection) does not use prior- knowledge of the test objects. It processes the reference image with a so-called ‘COMMED’-filter (COMbined MEDian), an edge preserving median filter. This filter performs a local image restoration on the original X-ray image. After the standard segmentation strategy it verifies found defects with a user defined positive list of defect types. Due to the fact that there is no need for training there is not much effort to handle such system even for non-experts.

Fig 1: Small defects (upper row) and very big defects with structure (lower row).

The PXV (Philips X-ray Vision) and the AI (Automatic Inspector) are two systems, which are using prior-knowledge. Both methods include a supervised training period. During this period image processing experts have to stay in place at the batch testing system. A deep understanding of all chosen parameters for the special filters is assumed. The AI uses a Neural Network (NN) filtering method based on a Hopfield-Tank NN to process the reference image. It is trained during the setup time. In contrast to this the PXV is a more flexible approach. It allows choosing several to some extent complex image processing filters for specific ROIs (regions of interest) to process the reference image. The training is done at runtime to improve or just to verify the default values of the setup phase. In both methods also the standard segmentation strategy is realized. After this defect verification with a negative list follows respectively.

In any case all systems are adapted for real-time processing, doing standard image enhancement of the X-ray images and a good/bad classification of the test object in the end. They differ in their flexibility and adjustability. This also means that they require more or less time per position for training or setup (see Table 1 for an estimation).

The most important difference of the systems will be shown in the next section. Due to their variable approaches each method shows a specific behavior for non-trivial defect types.

Automatic Defect Recognition with non-trivial defect types

As mentioned before new technologies, especially the flat panel detectors that provide a gray value range up-to 16 bit, make it possible to find new defect types. They were not visible with the older systems, but represent a serious safety hazard. Small defects of less than 3 pixels in size or very big defects with more than 32 pixels in size with a low intensity should be noted as extreme examples.

For known ADR systems it is a big challenge to recognize such defects reliably. In Figure 1 we show two kinds of typical defects:

• very small defects with very low contrast in the X-ray image,
• big low intensity defects that “hide” behind the structure of the part.

The second type is a very difficult case, because in addition to low intensity near noise level the edges of the overlying structure are very steep. The possibility to detect small defects is restricted by the resolution of the X-ray image. Therefore a physical limit is given by the pixel size. Table 2 shows the restrictions on defect recognition quality for the described ADR systems, which arise from physical or artificial limits, respectively. All systems are bounded by the image resolution or pixel size. Moreover, its defect templates restrict the ISAR, because it uses a positive list for defect recognition.

Fig 2: Local filtering of an X-ray image with big low intensity defect.

Big defects cannot be detected if the filter kernel is smaller than the defect itself (see Figure 2). This is a more artificial limit of local filters. As also shown in Table 2 all described methods are limited by their kernel size. The methods, which use prior- knowledge about the test object like the PXV and AI, are able to distinguish between structure and defects, even if a lot of fine-tuning in the choice of filters or parameters has to be done. They use a negative list (see Table 1). The ISAR depends again on its positive list or the defect templates, respectively. Thus there is no possibility to differentiate between noise, structure, and defects, because this method works without any prior- knowledge.

System Defect type	Human selected filters (e.g. PXV)	Local image restoration (e.g. ISAR)	Neural Network (NN) filtering (e.g. AI)
Small, low contrast	Pixel size	Pixel size, defect templates	Pixel size
Big, low intensity, hidden behind structure	Knowledge of supervised training, kernel size	Intensity of structure, kernel size, defect templates	Knowledge of supervised training, kernel size
Table 2: Restrictions on defect recognition quality for today’s ADR systems.

Conclusion

All described methods still underlie some restrictions regarding defect recognition. A new method should overcome these characteristics even for the artificial limits. The following features should not be used:

• positive list for defect verification,
• limitation by the kernel size,
• time consuming supervised training.

The dedicated algorithm has to be reliable and fast especially for the new detectable defect types. In the second part we will introduce a new method for ADR that is called “Trained Median Filter” (TMF).

II. New method: Trained Median Filter

The following requirements for a new fully automated analysis and classification method arise from the survey of ADR systems:

• use prior-knowledge and a negative list for defect verification,
• use non-linear, non-local filtering to overcome restrictions of the kernel size,
• use an unsupervised training with automated parameter adjustment.

The so-called “Trained Median Filter” (TMF) meets these criteria. For comparison with the well-known systems of Table 1 an analogous overview is given in Table 3. The new approach obtains its prior-knowledge from an image database. It analyzes a training set of system characteristic, flawless X-ray images of the test object in a systematic manner. This operation is named ‘unsupervised training’. It is fully automated. As a result it stores a set of parameters in a knowledge base. The training can be performed offline or in parallel to runtime processing as well as in place on a local host or even remotely in a distributed system.

Function/System	Non-local, non-linear filtering (e.g. TMF)
Prior-knowledge	Image database of good parts, characteristic for the batch testing system
Image enhancement (pre-process)	Included
Kind of training	Unsupervised analysis (offline, in place or remote)
Training period	1 minute per position
Mean setup time	2-3h per position (training inclusive)
Online filtering	TMF, using the trained knowledge base
Image segmentation	Difference of reference and original image
Pseudo / real defect	Very low
Defect verification	Negative list
Kind of classification	Good/Bad
Table 3: Functionality of the TMF.

After building up prior-knowledge in the way described above, the system is well suited to distinguish between defects and regular structures or to notice whether a part is incomplete. The last mentioned defect type cannot be found with methods not using any prior- knowledge, because they have no information about how a complete part looks like.

The online filtering of the TMF is a non-linear, non-local filter suitable for real-time processing. It is non-local in the sense that the filter kernel consists of the whole X-ray image (see Figure 2 and Figure 3). This also means that the TMF is not bounded to a special kernel size. The median function itself is a non-linear function, which belongs to the class of rank-order operations. With the aid of the processed knowledge base some pixels R_r with similar behavior are combined to each pixel X_i of the X-ray image.

Fig 3: A global filter kernel consists of the whole image.

The filter process comprises the following median operation

(1)

for each pixel X_i using its similar pixels R_r, r Î[0,K,R]

Besides the independence of the kernel size this operation does not have a certain sensitivity for a direction like a compass operator for example. Only the physical limits of pixel size and ROI specify the defect recognition quality. The segmentation process consists of the standard strategy explained in the first part and the defects are verified by a negative list. The algorithm ends up with the classification process.

As test results the following section demonstrates the real-time capability of the TMF in a batch testing system.

Results

The computation time of an ADR system under real-time constraints is predetermined by the part cycle of the batch testing system. Typical time constants for a modern batch testing system are listed in Table 4.

Process	Duration [s]
Position to position	0.8
Image capture per view	0.5
New image	»1.3
Whole part (incl.13 positions)	»20
Table 4: Time requests of a real-time environment..

To move from one position to another the mechanic needs about 0.8 s. Image capturing itself takes 0.5 s. Therefore approximately every 1.3 s a new, integrated X-ray image is ready to be processed. On the other side during this time the whole image processing for the last X-ray image has to be finished.

For example for one part with 13 positions about 20 s time is left for everything up to the good/bad-classification of the test object.

The time balancing for the training and the online filtering of the TMF is asymmetric. This means that more time is needed for training than for online filtering. But as already mentioned only the online filtering has to comply with the real-time constraints.

Fig 4: Segmented small (upper row) and big defects with structure (lower row).

The TMF is running on a general purpose PC (personal computer) with Microsoft’s Windows2000 environment. On a test system with an Intel Pentium III at 1 GHz the training for 51 test images takes 32.66 s for 512 x 512 pixels 16 bit gray value images. In this case the dedicated online filtering requires 0.113 s, which corresponds to 8.7 % of the available time in the sense of real-time computing. Both operations are proportionally bounded to the number N of pixels to be processed. Thus their complexities are O(N).

In Figure 4 the segmented defects are shown exemplarily that can be segmented after applying the TMF online filter onto the images of Figure 1. Like in Figure 1 the upper row shows small defects with very low contrast and the lower row big low intensity defects that were hidden behind a structure.

Conclusion

In this study we presented a new fully automated analysis and classification method for ADR. It is called “Trained Median Filter” (TMF). It overcomes the restrictions of the traditional ADR methods that are used today and have been described in the first part.

During an unsupervised training period the TMF automatically generates a knowledge base out of a database of flawless X-ray images of the test object. This method is not bounded to the filter kernel size, because it uses a non-local, non-linear filter approach. The prior- knowledge from the knowledge base can be reused for a negative list to distinguish between defects and structures for defect verification. As a result it is shown that this new approach is ideally suited to find also defects that are very small with very low contrast in the X-ray image or that are big with low intensity and are hidden behind the structure of the part. These defect types were not visible with older systems. Modern flat panel detectors made it possible to find them, but they are nearly undetectable for algorithms that have no information about the structure of the part. The measurement of the performance demonstrated the real-time capability of the TMF in a modern batch testing system.

Moreover the new system will make it possible to do a statistical evaluation of the quality of castings or even to do a finer classification of defect types than just a good/bad decision.

References

1. A. Krause, K. Bavendiek, A. Beyer, “Durchsatzerhöhung in der industriellen Röntgen- prüfung”, DGZfP Sep.1998, Bamberg, Germany, Vol. 63, No. 1, 1998, pp.301-306.
2. A. Stone, “What the Devil is Fully Automatic Real-Time X-Ray Inspection?”, CSNDT Journal,Vol.21, No.1, 2000, pp.6,8-10,12.
3. D. Filbert, H. Hecker, “Robuste Bildauswertung in der automatischen Röntgenserien- prüfung von Gußteilen“, DGZfP May 1994, Timmendorfer Strand, Germany, Vol.43, No. 1, 1994, pp.175-182.

© NDT.net - info@ndt.net

|Top|