AUTOMATIC CRACK DETECTION IN THERMAL IMAGES

FOR METAL PARTS

S. Ghidoni, M. Minella, L. Nanni, C. Ferrari, M. Moro, E. Pagello, E. Menegatti

Department of Information Engineering

University of Padova

Via Gradenigo 6/A I-35131 Padova, Italy

ABSTRACT. In this paper a system for automatic crack detection is presented. The system is capable of analyzing metal parts by means of a laser excitation system and a thermographic camera. The laser creates thermal gradients inside the part under inspection, and the thermal camera observes how heat diffuses inside the part. Cracks are automatically detected thanks tocomputer vision algorithms specifically developed for this task, that are capable of measuring and classifying heat profiles. Different algorithms have been developed for rugged and smooth metal parts, since the reaction to laser excitation is rather different. The detection algorithms have been tested on several sequences and showed very good detection performance also with cracks of very small size, having a width of 120 µm.

INTRODUCTION

Quality inspection at the end of a production line is an important stage in industry, especially for high-performance components. Parts undergoing strong mechanical and thermal stress should be carefully checked, since small defects can affect performance and reliability of a component. Crack detection is one of the most common checks to be performed, because cracks are a common source of failure, and they affect a high number of different productions.

For metallic parts, crack detection is still performed exploiting a technique called “magnetic particle inspection” (MPI): the part to be analyzed is first washed, then put into a magnetic field and finally covered with magnetic particles, either in the form of a dry powder, or, more frequently, in a wet suspension. Cracks are easily detected because they cause leaks in the magnetic flux; such leaks are highlighted by the particles, which can be inspected by means of a UV light. The whole process is very complex and needs to be done manually; it is also extremely time-consuming, because parts need to be cleaned, magnetized, covered with particles, inspected, de-magnetized and cleaned again. Moreover, magnetic particles and their carrier are a source of pollution, and should be properly processed after use.

Given the complexity of MPI, a method for simplifying the process of crack detection and making it automatic is highly desirable: investigation on this topic is the aim of the ThermoBot project. The main idea is to exploit thermography instead of magnetic particles to detect cracks, and to apply this method to parts made of non-metallicmaterials, like carbon fiber. Inspection is performed by means of a laser and a far infrared (FIR) camera (also called thermal camera or thermocamera), that observes how the heat carried by the laser diffuses inside the part: since cracks cause alterations on the heat flux, these can be exploited to detect cracks.Inspection of parts with complex geometry requires the laser and thermocamera to frame the object from many different viewpoints:the demonstrators developed for the project are therefore built around a robot that is able to move the part under inspection in front of the acquisition system.

In this paper a systemfor detecting cracks in metal parts is described.It was developed as a part of the project ThermoBot ( funded by the European Commission in the Factory of the Future research program. The Thermobot inspection system is based on the analysis of the laser spot that is going over the crack. The system was tested on a number of metal sample parts worked with different finishing processes, namely smooth and rugged, in order to assess the performance obtained on different surfaces.

STATE OF THE ART

The topic of crack detection has been tackled in a number of different ways in the literature, given the strong importance of this type of quality check.A variety of approaches have been used, like the propagation of ultrasounds that is used in [1] to detect cracks and lamination defects in metallic pipes, or Eddy currents [2,3,4]. Other methods exploit magnetic cameras to detect crack in parts that are at high temperature [5], or magnetic flux leakage[6], while the method described in [7] studies the heat produced by the Joule effect.

Methods based on image analysis have also been exploited in the literature, ranging from detection of welding defects in pipelines [8] to concrete surface analysis [9] and theprotection of cultural heritage [10]. Thermographic image analysis systems have recently been proposed for performing in-situ non-destructive inspections during thermomechanical fatigue tests [11]; the system showed a high sensitivity, being able to detect cracks smaller than 500 µm. The system proposed in [12] is slightly different from the others discussed above as it is meant to inspect different types of materials during fatigue tests, and detect the cracks as soon as they appear.

Thermography-based crack detection is often coupled with excitation methods like eddy currents [13] or laser beams;in particular, lasers provide the inspection process with high flexibility, as it is possible to concentrate the heat on a small spot, and enabling and disabling the heat source can be done instantly, generating pulses at high frequency. This last characteristic is exploited in pulse thermography and techniques that are derived from it [14]. Another technique based on laser technology is the “flying spot active thermography” [15], that refers to a laser spot that causes a local excitation on the part under inspection. This is similar to the analysis method employed in the ThermoBot project, and was chosen in [15] to inspect high pressure turbine blades.

In the following, two techniques for detecting cracks in metal parts will be described. They were developed for addressing the task of automatic crack detection in parts with different characteristics and with different thermal excitation methods. Both smooth and rugged metal parts will be considered, excited using both pulsed and continuous heat source.

ANALYSIS OF SMOOTH METAL PARTS

Inspected parts

Parts that were considered for testing the inspection system belong to two categories:

  • Test part A: metal discs, composed of eight blades made of smooth and reflecting metal;
  • Test part B: portions of a crankshaft, made of rugged metal, and characterized by a more complex geometry.

Even though both test parts are made of metal and the physical principle on which the detection is based is the same for both, acquired images are very different, because of the different surface working.

Test part A is the simplest case: the flat and mirror-like surface provides very uniform images, that are easy to process. A laser beam that hits the part saturates the camera, and appears as a white spot surrounded by a region of decreasing gray levels, caused by heat diffusion from the hot spot. The crack detection algorithm for smooth metal parts like test part A is based on the analysis of the region where heat diffuses.

The analysis of the heat diffusion region is performed in three steps:

  • Hot spot detection,
  • Radial gradient analysis,
  • Tangential gradient analysis.

As a first step, the hot spot needs to be detected. This is rather easy as the laser spot is the portion of the image with highest temperature, and is always saturated in all working conditions we observed. It is also reasonable to assume that this is true in most acquisition setups, therefore hot spot detection is performed with a simple algorithm. A dynamic threshold, depending on the lowest and highest temperature values found in the image, is first applied to obtain a binary image in which the laser spot is the only white element. A dilation operator is then exploited to smooth the shape obtained by thresholding; the centroid is finally evaluated as the center of mass of the resulting shape.

Once the shape and location of the laser spot is available it is possible to analyze the surrounding area to perform crack detection.

Radial gradient

The first analysis involving the heat flux region is performed on its radial gradient. While a normal gradient operator considers pixel difference only along the horizontal and vertical axes, the radial version considers multiple directions intersecting in a central point, that corresponds to the hot spot centroid in our case.

Radial gradients are usually evaluated by comparing points that are aligned along a given direction intersecting the central point. However, when this is evaluated in the discrete domain of an image, an important side effect should be considered: the number of pixels at a given distance to the center is not constant, but depends on the distance itself. This is important in performing comparisons: for example, in [16] an approach is proposed that instead of comparing pixel values focuses on image areas of variable size, depending on the distance to the center.

Our approach to the problem is slightly different, and moves from outer regions towards the center. Consider the image portion of fig. 3 (left): the center is represented by the green spot at the bottom left corner. For each pixel, a vector connecting it with the center is shown, whose color is red if it crosses one pixel only, and in blue if the crossed pixels are two. It is important to separate these two cases as pixels with a red vector can be compared with one other pixel (indicated by the vector), as it happens in any other gradient evaluation algorithm. Pixels with a blue vector need to be compared to more than one other pixel, namely the two that are intersected by the vector pointing to the center, as highlighted in fig. 3 (center). As it can be seen from the figure, each pixel in the image is compared with no more than two other pixels, which solves the issue of considering image portions of increasing area described in [16].

Figure 1. Radial gradient schemes.

Using the method described above four different situations can be recognized, that are summarized in fig.3 (right).The pixels PA, PB and PC are connected to the center along a vertical, diagonal and horizontal direction respectively; for this reason, the radial gradient is evaluated by comparing them with one pixel, that is the one pointed to by the vector. In the case of PD, the line connecting to the image center has a stronger vertical component: PD should therefore be compared with the pixels placed along the down and down-left directions. The case of PE is specular. The values that the radial gradient will assume are therefore:

whereis the point at coordinates, and is a function that controls the different weight of the neighboring pixels.

Tangential gradient

The tangential gradient provides complementary information with respect to the radial version, and it is therefore important to fully describe the area surrounding the laser spot. The principle on which it is evaluated is similar to what was described for the radial gradient, with the only fundamental exception that in this case vectors do not lie on the direction connecting each pixel to the spot center, but are perpendicular to such direction. The scheme in fig. 5 has the same meaning of fig. 4 in this new context.

Figure 2. Four types of pixel comparison for the tangential gradient algorithm.

Radial gradient equalization

The gradient evaluated on the area surrounding the laser spot is not uniform, and it is higher in the locations closer to the spot, as it is easily understood by considering the heat transmission inside metal parts. From the image processing point of view, this breaks the homogeneity of the region being inspected, and should therefore be contrasted. To balance this effect, we propose a correction called radial gradient equalization (RGE), that is an amplification of the gradient that depends on the distance to the centroid of the laser spot. The equalization is applied in a region surrounding the laser spot until a maximum length defined by the parameter. The amplification is defined as a function of the distance to the centroid and increases linearly from the valueuntil a certain valuethat is reached at the distance, then the gain saturates:

Results obtained applying the RGE can be seen in fig. 6: the cracks at the bottom edge are highlighted after applying the equalization (right) with respect to the original image (left).

Figure 3. Comparison of the gradient image before (left) and after (right) the application of the RGE.

Edge-based crack detection

The proposed algorithm for detecting cracks is meant to analyze the images of the radial and tangential gradient, and has better performance when equalization is adopted. To detect cracks, the gradient image is divided into smaller parts (patches) that are analyzed separately. Each patch is binarized by means of an adaptive threshold that depends on the average pixel value. Some image enhancing functions are then applied, and finally cracks are detected selecting the contours in the image having a significant size, discarding the one containing the laser spot. The final algorithm for crack detection is rather simple, since it operates on images that are strongly enhanced at lower level.

An example of crack detection can be seen in fig. 7, where the binarized image is shown (left) together with the final result (right).

Figure 4. Example of image binarization (left) and final result (right) of the system

ANALYSIS OF RUGGED METAL PARTS

The algorithms described so far do not show good performance on rugged parts, because the detailed analysis of radial and tangential gradients suffer from the noise generated by rugged surfaces. A different approach has been developed to tackle this case, which is based on radial density profile (RDP) [17], a method that has been used to characterize medical images containing viruses. Such medical images are rather similar to the case of laser spot analysis as in both cases it is important to study the circular shape of an object. The RDP features are used to train a support vector machine [18], which is state-of-the-art among the machine learning classifiers.

While the algorithm previously described can be seen as a “direct method”, since it is aimed at detecting image elements that have the shape of a crack, the method based on RDP is “indirect”, because it detects cracks analyzing the evolution of the shape of the laser spot while it is heating a crack. The heat source was not pulsated in this case, as it would be impossible to study the shape of a pulsating laser spot.

DISCUSSION

The approaches presented in this paper have been tested on several sequences. Tests involved crack detection both on single images and in whole sequences.

In the case of smooth metal parts, only few images showing a crack were present in the dataset, that was made of a limited number of pictures. Tests showed a good performance of the algorithm, that was able to detect cracks in the region around the laser spot. However, cracks cannot be detected when they are too close to the laser spot itself, because this is masked by the algorithm, making any detection not possible. This means that in 38% of the images the crack is not detected, but the same crack was previously detected in other images, when it was located further from the laser spot. Considering the whole sequence, the algorithm showed optimal performance, because it detected all cracks in the sequences, but the dataset is still too small to thoroughly assess the system.

Performance of the crack detector for rugged metal parts was studied in more detail, thanks to a much larger dataset. The system was tested on 31 sequences, each one framing the laser going twice on a crack; all sequences were taken with the same sample part. The crack had very small dimension, having a length of 8.36 mm length and a width of 120 µm only.

The sequences are divided into two sets:

  • In set A, the laser power is kept at a given value of 7.5 W and the laser speed changes in the range [60-200] mm/s, with a step of 10 mm/s between two consecutive sequences;
  • In set B the speed is kept at the value of 60 mm/s while the laser power takes values in the range [5-20] W with a step of 1W.

As a testing protocol we chose a leave one out set protocol, that is, we trained the classifier on the features extracted from set A and tested it on set B, and vice-versa. Tests resulted in an average area under the ROC curve of 0.9337. The resulting DET-curve [19] is reported in fig. 7, which shows the good performance achieved by the system, considering that only one sample part acquired changing laser power and scan speed was employed for training and testing.

Figure 5. DET-curve resulting from tests performed on crack detection using RDP.

CONCLUSIONS

In this paper a system for automatic detection of cracks in metal parts was presented. The system is based on thermographic analysis of the part under inspection and exploits a laser excitation, either pulsed or continuous. The system exploits different algorithms depending on the metal parts to be inspected, that can be either smooth or rugged, and is able to detect cracks that are extremely small: the smallest detected crack has a width of 120 µm only. Even though the work is still in progress, a first performance assessment showed good results, indicating the approach is promising. Future works include the development of algorithms for thermographic image processing for detection of defects on thermographic images obtained with flash thermography.