Bad Pixel Management
20.12.2019, 19:58
As shown in underperfoming pixel may happen - similar to most of the digital cameras where these pixel are substituted by the neighbor pixel.
Here is a real example of a detector with low performance:
This sample was taken to define underperforming pixel in to the following reasons:
1. Dead Pixel—Pixels that have no response, or that give a constant response independent of radiation dose on the detector.
2. Over Responding Pixel—Pixels whose gray values are greater than 1.3 times the median gray value of an area of a minimum of 21×21 pixels.
3. Under Responding Pixel—Pixels whose gray values are less than 0.6 times the median gray value of an area of in a minimum of 21×21 pixels.
4. Noisy Pixel—Pixels whose standard deviation in a sequence of 30 to 100 images without radiation is more than 6 times the median pixel standard deviation for the complete DDA.
5. Non-Uniform Pixel—Pixel whose value exceeds a deviation of more than ±1 % of the median value of its 9×9 neighbor pixel.
6 Persistence/Lag Pixel—Pixel whose value exceeds a deviation of more than a factor of 2 of the median value of its 9×9 neighbors in the first image after X-ray shut down.
Additionally a single good pixel surrounded by only bad pixel is also a bad pixel:
7 Bad Neighborhood Pixel—Pixel, where all 8 neighboring pixels are bad pixels, is also considered a bad pixel.
As mentioned above a single bad pixel is substituted with his good neighbor pixel - 4 or mainly 8:
What happens, if a bad pixel has a bad neighbor?
The definition is given also in E2597:
1. Single Bad Pixel—A single bad pixel is a bad pixel with only good neighbor pixels.
2. Cluster of Bad Pixels—Two or more connected bad pixels are called a cluster. Pixels are called connected if they are connected by a side or a corner (8-neighbor possibilities). Pixels which do not have 5 or more good neighbor pixels are called cluster kernel pixel (CKP).
Hint: CKPs are very unwelcome as it is difficult to correct them in a way that you wouldn't see it in the image.
3. A cluster without any CKP is well correctable and is labeled an irrelevant cluster. The name of the cluster is the size of a rectangle around the cluster and number of bad pixels in the irrelevant cluster, for example, “2×3 cluster4”.
4. A cluster (excluding a bad line segment) with CKP is labeled a relevant cluster. A line cluster with CKP is classified differently. The name of the cluster is similar to the irrelevant cluster; with the exception that the prefix “rel” is added and the number of CKPs is provided as a suffix, for example, “rel3×4 cluster7-2”, where 7 is the total number of bad pixels and 2 are those in this group that are CKPs.
5. A bad line segment is a special cluster with ten or more bad pixels connected in a line (row or column) where no more than 10% of this line has adjacent bad pixels. If there are CKPs in the line segment at one end it may be separated to a bad line segment and relevant cluster. In the Figure above in the lowest line a relevant cluster is located at the end of a bad line segment. The bad line segment is then separated from the relevant cluster. In this example, the bad line segment is a 1×24 Line24 and attached with a relevant cluster Rel4×3 cluster 8-5.
Here is a real example of a detector with low performance:
This sample was taken to define underperforming pixel in to the following reasons:
1. Dead Pixel—Pixels that have no response, or that give a constant response independent of radiation dose on the detector.
2. Over Responding Pixel—Pixels whose gray values are greater than 1.3 times the median gray value of an area of a minimum of 21×21 pixels.
3. Under Responding Pixel—Pixels whose gray values are less than 0.6 times the median gray value of an area of in a minimum of 21×21 pixels.
4. Noisy Pixel—Pixels whose standard deviation in a sequence of 30 to 100 images without radiation is more than 6 times the median pixel standard deviation for the complete DDA.
5. Non-Uniform Pixel—Pixel whose value exceeds a deviation of more than ±1 % of the median value of its 9×9 neighbor pixel.
6 Persistence/Lag Pixel—Pixel whose value exceeds a deviation of more than a factor of 2 of the median value of its 9×9 neighbors in the first image after X-ray shut down.
Additionally a single good pixel surrounded by only bad pixel is also a bad pixel:
7 Bad Neighborhood Pixel—Pixel, where all 8 neighboring pixels are bad pixels, is also considered a bad pixel.
As mentioned above a single bad pixel is substituted with his good neighbor pixel - 4 or mainly 8:
What happens, if a bad pixel has a bad neighbor?
The definition is given also in E2597:
1. Single Bad Pixel—A single bad pixel is a bad pixel with only good neighbor pixels.
2. Cluster of Bad Pixels—Two or more connected bad pixels are called a cluster. Pixels are called connected if they are connected by a side or a corner (8-neighbor possibilities). Pixels which do not have 5 or more good neighbor pixels are called cluster kernel pixel (CKP).
Hint: CKPs are very unwelcome as it is difficult to correct them in a way that you wouldn't see it in the image.
3. A cluster without any CKP is well correctable and is labeled an irrelevant cluster. The name of the cluster is the size of a rectangle around the cluster and number of bad pixels in the irrelevant cluster, for example, “2×3 cluster4”.
4. A cluster (excluding a bad line segment) with CKP is labeled a relevant cluster. A line cluster with CKP is classified differently. The name of the cluster is similar to the irrelevant cluster; with the exception that the prefix “rel” is added and the number of CKPs is provided as a suffix, for example, “rel3×4 cluster7-2”, where 7 is the total number of bad pixels and 2 are those in this group that are CKPs.
5. A bad line segment is a special cluster with ten or more bad pixels connected in a line (row or column) where no more than 10% of this line has adjacent bad pixels. If there are CKPs in the line segment at one end it may be separated to a bad line segment and relevant cluster. In the Figure above in the lowest line a relevant cluster is located at the end of a bad line segment. The bad line segment is then separated from the relevant cluster. In this example, the bad line segment is a 1×24 Line24 and attached with a relevant cluster Rel4×3 cluster 8-5.
Re: Bad Pixel Management
20.12.2019, 20:37
The interpolation of bad pixels can reduce the contrast of very small indications.
The interpolation of the defect pixel reduces the contrast by 15% and the adequate reduction of noise is only achieved indirectly by an increase of the SNR – for instance, an increase of the SNR to 265 will be sufficient to visualise the interpolated indication.
The worst case is when the flaw is centered on a bad pixel:
Here an increased SNR of 385 would be required to see the flaw on the substituted bad pixel.
But in both cases you additinally have the good pixel who show the flaw (even without the substituted bad pixel)!
The required CNR for the digital radiographs depends on the ratio of the detectable flaw size to the cluster size of bad pixels. The next picture shows a recommended practice for the management of bad pixels and required increase of SNR and corresponding CNR. As mentioned in the section above, increasing CNR and SNR may enable smaller defects to be detected. The picture provides guidance on this dependence, as well as the dependence of these metrics on bad pixel management. Three zones are distinguished: Best practice zone, zone with risky detection, and zone of negative detection. It shall be noted that there are no exact CNR values agreed on for the limiting lines in Fig. 7 of ASTM E2597 yet.
The required SNR for the best practice can be recommended as follows: For example, if film and digital detector are applied at the same kV and the specific contrast (µeff) is comparable, it is sufficient to require a minimum SNR. Typical SNR values for best practice in film radiography are 150 for fine grained films (at net density =2), and 75 for coarse grained films. Since DDAs have typically higher unsharpness than films, an adequate increase of SNR is recommended. Here, the concept of normalized SNRN can be applied, and the recommended SNR shall be increased as follows:
SNRrecommended = SNRN * SRb/88.6um
with SRb – basic spatial resolution of the system
If the film application (coarse grained film) provides SNRN > 75, the adequate DDA radiograph taken by a detector with and magnification 1 and 200µm effective pixel size (SRb=200 µm), should exhibit a SNR > 170. This should apply for the last two areas in the figure above (right side of the figure beginning with: isolated bad pixels and irrelevant clusters); best practice line. The SNR selected will depend on the available time allotted to the inspection, and can so be improved using DDAs if time is available to do so. The required minimum SNR depends on the X-ray energy and may vary if compensation principle I and II are applied. The final proof of image quality requires always the visibility of the agreed IQIs as in classical radiography.
The interpolation of the defect pixel reduces the contrast by 15% and the adequate reduction of noise is only achieved indirectly by an increase of the SNR – for instance, an increase of the SNR to 265 will be sufficient to visualise the interpolated indication.
The worst case is when the flaw is centered on a bad pixel:
Here an increased SNR of 385 would be required to see the flaw on the substituted bad pixel.
But in both cases you additinally have the good pixel who show the flaw (even without the substituted bad pixel)!
The required CNR for the digital radiographs depends on the ratio of the detectable flaw size to the cluster size of bad pixels. The next picture shows a recommended practice for the management of bad pixels and required increase of SNR and corresponding CNR. As mentioned in the section above, increasing CNR and SNR may enable smaller defects to be detected. The picture provides guidance on this dependence, as well as the dependence of these metrics on bad pixel management. Three zones are distinguished: Best practice zone, zone with risky detection, and zone of negative detection. It shall be noted that there are no exact CNR values agreed on for the limiting lines in Fig. 7 of ASTM E2597 yet.
The required SNR for the best practice can be recommended as follows: For example, if film and digital detector are applied at the same kV and the specific contrast (µeff) is comparable, it is sufficient to require a minimum SNR. Typical SNR values for best practice in film radiography are 150 for fine grained films (at net density =2), and 75 for coarse grained films. Since DDAs have typically higher unsharpness than films, an adequate increase of SNR is recommended. Here, the concept of normalized SNRN can be applied, and the recommended SNR shall be increased as follows:
SNRrecommended = SNRN * SRb/88.6um
with SRb – basic spatial resolution of the system
If the film application (coarse grained film) provides SNRN > 75, the adequate DDA radiograph taken by a detector with and magnification 1 and 200µm effective pixel size (SRb=200 µm), should exhibit a SNR > 170. This should apply for the last two areas in the figure above (right side of the figure beginning with: isolated bad pixels and irrelevant clusters); best practice line. The SNR selected will depend on the available time allotted to the inspection, and can so be improved using DDAs if time is available to do so. The required minimum SNR depends on the X-ray energy and may vary if compensation principle I and II are applied. The final proof of image quality requires always the visibility of the agreed IQIs as in classical radiography.
Re: Bad Pixel Management
20.12.2019, 21:27
Here are the results of a test with bad pixel correction done in the year 2003 and presented on the ASTM January 2004 meeting in Tampa.
The main question was if a detector with bad pixel can be allowed to be used for critical inspections. The post before gives the answer but the test results may be interesting, too.
The test was done to show the influence of plaque hole penetrameter visibility with different bad pixel correction maps. The correction of a single bad pixel is already discussed above, now the procedure of Multipe Pixel Correction should be faced:
The blue marked pixel are used two times for correction. What happens in a 2x3 cluster?
Here we see two CLKs and the blue marked pixel with white letter are used three times for correction.
The correction of two line defects with two good pixel between two bad pixel shows a lot of pixel which are used three times for correction:
but the worst case is where there is only one good pixel between two bad pixel in the bad lines:
Here the middle pixel are used six times for correction
The next test is a bad pixel grid; single pixel are used four times for correction:
and the worst case is where all good pixel are surrounded by bad pixel and each pixel is used eight times for correction:
For the test we created a bad pixel list with rows/colums, two pixel distance:
then we “and” and “or” both lists
We did the test with 5 different spaces of the bad lines - from one pixel distance (worst case) up to 5 pixel distance between bad pixel lines:
We used a Thales FS30 detector with the X-Ray dose of 120kV, 1mA from a 0.6mm focal spot at 0.6m distance and no magnification. 16 frames were integrated. The test object is a 30mm thick Aluminum block with two penetrameter 2% and 1% (0.62 & 1.2). The block is moved from “bad pixel area” to “bad pixel area” for each image:
When we used the second worst case - two good lines between two bad lines - we really did not see big differences in the visibility of the penetrameter holes:
[klick for 1:1 size]
We skipped the other distances and moved to the worst case with only one good pixel between two bad pixels and additionally the bad pixel grids:
[klick for 1:1 size]
even in the 1:1 view it is hard to see the differences.
Therefore we applied a high pass filter 9x9 and still it is hard to see big differences.
[klick for 1:1 size]
Finally we magnified to 2:1 and now you will see that the worst case 2-2 shows higher unsharpness - but it is funny that the 1T hole is not influenced by this unsharpness.
We had long discussions about this test. It may not be reprentative anymore as the detector technology has improved and this combination of scintillator and pixelsize (160µm thick active scintillator and 127µm pixel size with a SRb of 160µm) may compensate due to the higher unsharpness of the scintillator the information ==> see posting above.
Please feel free to post your opinion here
The main question was if a detector with bad pixel can be allowed to be used for critical inspections. The post before gives the answer but the test results may be interesting, too.
The test was done to show the influence of plaque hole penetrameter visibility with different bad pixel correction maps. The correction of a single bad pixel is already discussed above, now the procedure of Multipe Pixel Correction should be faced:
The blue marked pixel are used two times for correction. What happens in a 2x3 cluster?
Here we see two CLKs and the blue marked pixel with white letter are used three times for correction.
The correction of two line defects with two good pixel between two bad pixel shows a lot of pixel which are used three times for correction:
but the worst case is where there is only one good pixel between two bad pixel in the bad lines:
Here the middle pixel are used six times for correction
The next test is a bad pixel grid; single pixel are used four times for correction:
and the worst case is where all good pixel are surrounded by bad pixel and each pixel is used eight times for correction:
For the test we created a bad pixel list with rows/colums, two pixel distance:
then we “and” and “or” both lists
We did the test with 5 different spaces of the bad lines - from one pixel distance (worst case) up to 5 pixel distance between bad pixel lines:
We used a Thales FS30 detector with the X-Ray dose of 120kV, 1mA from a 0.6mm focal spot at 0.6m distance and no magnification. 16 frames were integrated. The test object is a 30mm thick Aluminum block with two penetrameter 2% and 1% (0.62 & 1.2). The block is moved from “bad pixel area” to “bad pixel area” for each image:
When we used the second worst case - two good lines between two bad lines - we really did not see big differences in the visibility of the penetrameter holes:
[klick for 1:1 size]
We skipped the other distances and moved to the worst case with only one good pixel between two bad pixels and additionally the bad pixel grids:
[klick for 1:1 size]
even in the 1:1 view it is hard to see the differences.
Therefore we applied a high pass filter 9x9 and still it is hard to see big differences.
[klick for 1:1 size]
Finally we magnified to 2:1 and now you will see that the worst case 2-2 shows higher unsharpness - but it is funny that the 1T hole is not influenced by this unsharpness.
We had long discussions about this test. It may not be reprentative anymore as the detector technology has improved and this combination of scintillator and pixelsize (160µm thick active scintillator and 127µm pixel size with a SRb of 160µm) may compensate due to the higher unsharpness of the scintillator the information ==> see posting above.
Please feel free to post your opinion here
Hosted by iphpbb3.com
Impressum | Datenschutz