Defect detection

Abstract

A method of detecting defects in an object from an image of the object, comprises: estimating, in the image, the position of a feature of the object; finding parts of the feature which deviate from th

Description

The present invention relates to the detection of defects in an object from images of the object which have been obtained, for example, using a video probe, video borescope, or remotely operated cameras. It is commonplace to carry out visual inspections of components within a complex machine using remote visual inspection devices as part of their maintenance. It is particularly important to carry out such inspections on safety critical machines, such as aircraft engines. In such machines, inspections must be carried out at defined intervals, and those inspections must be made of particular components of the machine. As it currently stands, carrying out remote visual inspections requires the use of highly skilled and specialised inspectors who are trained to carry out inspections which identify any defect of the machine. An inspector orientates the optical sensor of a visual inspection device in order to obtain an image of any component to be inspected which enables a part of the component to be assessed. Most remote visual inspection devices will include a display screen that the inspector can use to make the assessment. Furthermore, it is commonplace to record an image of the component. If the remote visual inspection identifies a component with a defect, that component can undergo maintenance. While such inspections are effective, there is a need to improve the accuracy of the inspection process, particularly to reduce the number of costly false positive defects detected. According to a first aspect of the present invention, a method of detecting defects in an object from an image of the object comprises: estimating, in the image, the position of a feature of the object; finding parts of the feature which deviate from the estimated position; filtering the parts which deviate from the estimated position using filtering criteria to retain only those which match defect criteria, wherein the filtering criteria is forked into multiple paths which use different filtering criteria, and wherein the path selected is dependent on the deviation from the estimated position. Preferably, the path selected is based on the size of the defect determined by the deviation from the estimated position. For relatively small defects in which the deviation is relatively small, a different path may be selected in which the filtering criteria are different from the filtering criteria selected on a path for relatively larger defects. Advantageously, the minimum peak size is increased and the maximum peak size decreased in the path selected for very small defects. Preferably, the method is for detecting defects in a blade, and the feature of the object is an edge of the blade, such as a leading edge or a trailing edge of the blade. According to a second aspect of the present invention, a defect detection device for detecting defects in an object from an image of the object comprises a computing device programmed and configured to carry out the method of the first aspect of the invention. An example of the present invention will now be described by way of example only with reference to the following drawings: Figure 1 is an image of an object showing a defect in that object; Figure 2 is a pictorial flowchart showing the steps of a defect detection process according to the present invention; Figure 3 is an image representing the input step of the defect detection process of Figure 2; Figure 4 is an image showing the image of Figure 3 after it has been rotated and masked; Figure 5A and 5B are images showing the estimation step of the defect detection process of Figure 2 made up of a detected edge estimate and a refined detected edge estimate that only contains parts of the estimate where there is high contrast on that part of the image; Figure 6 are images showing the refinement step of the defect detection process of Figure 2; 21 03 24 Figure 7 is a graph showing the blade edge. Peaks are indicated where defects are found in the defect detection process of Figure 2; Figure 8 shows the final output of the defect detection process of Figure 2; and Figure 9 are photographs of blades with large and small edge defects. 5 1. Blade Edge Defect Detection Process: In this example, a defect in the edge of a blade is defined as having the following properties: 10 • A break in the edge profile of the foreground blade • With a non-negligible width In other examples, the object might be different, and the definition of the defect might be different. The edge of the blade might be the leading edge or the trailing edge. 15 2, Process Overview: The defect detection process is applied to the image of a part of an edge of a blade shown in Figure 3, and the pictorial flowchart of Figure 2 shows the steps of the process as follows: Inputs • Selected image featuring blade edge • Configuration given from blade counter tool Processes • Mask and rotate image • Estimate approximate blade edge position • Remove blade edge estimate where edge is not clear • Refine blade edge • Find parts of refined edge which deviate from estimate • Filter deviations to retain only ones which match defect criteria Outputs • Pixel coordinates of identified defects with height / width information 3, Process description Input Step: The inputs for the process are a single image of a blade (Figure 3) and a set of configuration data from a blade counter tool. The blade counter tool contains a store of the configuration data, and passes this data along with an image to the detector.. The configuration data includes information about the blade edge, such as its shape, its orientation, how it will appear in any image, the extent to which any image of the blade edge is to be masked and the definition of a defect. The blade counter tool may be set up by a user to define the edge detector by identifying the length of a blade edge, its shape and its orientation. In this example, the length of the edge detectors created during setup are used to determine the diameter of the circular image mask. This configuration data is used to calculate the rotation required to orient the image for analysis, and the image mask parameters to be applied to mask parts of the image so that the remaining image can be processed. The image is then presented in a way which is consistent with other images of blade edges so that the following steps are carried out in a consistent and repeatable way. Figure 3 shows an example input image that might be input into the detector. We are only interested in the blade edge, but the image includes lots of other features which are not relevant. The orientation of the image will also vary depending on the way in which the image is taken, leading to inconsistencies between images taken of the same or different blade edges. Masked &Rotated Image Step: The next stage for the detector is image masking and rotating. Figure 4 shows the result of the masking and rotation process for the example input image. The mask parameters are calculated from the configuration data taken from the blade counter tool. This masked area is the only region the process will process and look for defects. The image is then rotated, based on the angle of the edge detectors such that edge is vertical in image and the blade surface is on the right side of the image and the background on the left. Estimate Approximate Edge Step: Since edge defects are defined to be a break in the profile of the blade edge, the blade edge must be located in the image. The first step towards estimating the approximate position of the blade edge is to run a standard vertical edge detector, such as a Sobel edge detector, across the whole image. This detector outputs all vertical edges present in the image no matter how faint. Each edge is then scored based on size and location. The blade edge is assumed to be the longest in the image so the longer an edge, the higher the size score. It is also known that the blade is near the mask centre, so the closer to the mask centre the edge is, the higher the location score. The edge with the highest overall score is selected as the blade edge. The detected edge from the standard vertical edge detector might be incomplete, not covering the entire edge, perhaps due to the presence of a defect causing part of the edge to no longer be vertical, or there might be noise in the detected edge. A 2D polynomial is fit to the y, x pixel coordinates of the edge which extends to the image boundaries. This results in the creation of a smooth line or curve roughly covering the detected edge as well as other parts of the edge that were undetected. This is the estimated approximate blade edge shown overlaid in Figure 5 A. Edge Estimate Only High Contrast Step: Once the approximate blade edge is estimated, this estimation is analysed and any parts of the estimated edge that are deemed to have poor contrast (i.e. not a strong enough edge present) are discarded and not used in future steps. Figure 5B shows this stage of the detector, with an approximate estimate of the blade edge being found and the top and bottom parts of the estimated edge which extend into the masked area being removed. Refined Edge / Estimated Edge Difference Step: Blade edge refinement takes place, with the result shown in Figure 6. For each pixel point of the blade edge estimate, the pixels to its right along the same row (over the blade surface, not the background) are analysed. The closest pixel to be significantly brighter is deemed to be where the true edge lies. Repeating this process down the entire estimated edge results in a more accurate model of the true blade edge in the form of a refined edge. This process is necessary since some defects can be very subtle, and the approximate edge found by the original approximate blade edge estimation step may not provide enough detail to identify the defect. This stage also provides an estimate for the error or deviation between the refined blade edge and the approximate blade edge (i.e. what the edge might look like without a defect). Unfiltered Peaks Step: The differences between the approximate blade edge estimated above, and refined edge are analysed. The locations where there are significant deviations (peaks) of the refined edge compared to the approximate blade edge estimated above are recorded. Figure 7 shows the process of extracting potential defect positions from this detected edge into a graph. By detecting peaks in the refined blade edge and associated error the potential defect positions can be extracted. In this case, three locations are found with peaks suggesting a defect, each indicated with a dot in Figure 7. The peaks are then further analysed against a set criteria to determine is the peak is really a defect. There are a number of parameters that can be altered to improve the likelihood of detection for defects of a specific size. One example of this is to alter the resolution at which processing takes place, which helps in the detection of very small defects, but also introduces additional false positives that must be filtered out. Filtered Peaks Step: Furthermore, when potential defects are filtered out to remove false positives, the filtering criteria is forked into multiple paths, where various constraints are modified to mitigate the impact of the changes made earlier in the process. The peaks are forked based on their size - smaller defects can therefore use a different filtering criteria to larger defects. This is helpful because smaller defects generally have a different shape compared to larger defects, as can be seen in Figure 9. On smaller edge defects for example, the edge moves into the blade at a sharper angle than on larger defects and this characteristic can be incorporated into the filtering criteria. Also the minimum peak size is increased and the maximum peak size decreased for very small defects. In this example, two of the possible defects indicated in Figure 7 are filtered out leaving a single large defect indicated by the single dot in Figure 8. Useful information about any defects are stored, such as location, height and extent, and then returned in the form of a list. Figure 8 shows the final results of defect detection, a set of pixel positions describing the locations of any large defects in the image.

Claims

1. A method of detecting defects in an object from an image of the object, comprising:estimating, in the image, an approximate position of a feature of the object;finding parts of the feature which deviate from the estimated approximate position; andfiltering the parts which deviate from the estimated position using filtering criteria to retain only those which match defect criteria, wherein the filtering criteria is forked into multiple paths which use different filtering criteria, and wherein the path selected is dependent on the deviation from the estimated position.

2. A method according to claim 1, wherein the path selected is based on the size of the defect determined by the deviation from the estimated position.

3. A method according to claim 1 or 2, further comprising estimating the deviation of the feature from the estimated approximate position.

4. A method according to claim 2 or 3, wherein, for relatively small defects in which the deviation is relatively small, a different path is selected in which the filtering criteria are different from the filtering criteria selected on a path for relatively larger defects.

5. A method according to any one of the preceding claims, wherein the minimum peak size is increased and the maximum peak size decreased in the path selected for very small defects.

6. A method according to any one of the preceding claims wherein the method is for detecting defects in a blade, and the feature of the object is an edge of the blade.

7. A defect detection device for detecting defects in an object from an image of the object, comprising a computing device programmed and configured to carry out the method of any one of claims 1 to 6.

Images