Inquiry Regarding Threshold Calculation in Dead Pixel Correction

[SonalSheth107107]

I hope this message finds you well.

I am currently working on implementing a dead pixel correction module as part of an image signal processing (ISP) pipeline, and I came across your work and solutions in this domain. I am particularly interested in understanding the method or rationale used to determine the threshold values in your dead pixel correction algorithm.

Could you please provide any insights into how these thresholds are calculated or selected? For example:

• Are they empirically derived or based on sensor noise characteristics?
• Are they adaptive per image or fixed across conditions?
• Is there a published reference or guideline you could point me to?

Any information or references you could share would be greatly appreciated, as it would help me refine the accuracy and reliability of the correction algorithm in my implementation.

Thank you for your time and support. I look forward to hearing from you.

Best regards,

Sonal Sheth

[marianadeem-10xe]

Hi Sonal,

The dead-pixel correction (DPC) threshold is calibrated once per sensor for each ISO setting, and those values can then be applied to every image the sensor captures. In other words, the threshold must account for varying conditions (different ISO levels), but it stays constant for images taken with the same sensor at the same ISO.

There are essentially 2 ways to correct dead pixels:

• Static
• Dynamic.

In the static approach, a series of experiments are conducted to identify the permanently defective sites. Dark frames (captured with no light) are used to detect abnormally bright "hot" pixels, while flat-field frames (captured under uniform illumination) are used to reveal unresponsive "dead" pixels. The coordinates of the identified defective pixels are then saved stored in a fixed defect map, and during processing, the ISP applies correction exclusively on these sites.

A dynamic approach, such as the one implemented here, on the other hand first identifies the dead pixels and then applies correction. In our DPC block, the threshold is used for identification of the dead pixels. A pixel is flagged as defective if its value differs from all neighbors by more than the set threshold (|P-Q_i| > Threshold for all i=1,2...8). Correction replaces defective pixels with the average of two neighbors in the direction of minimum gradient to preserve edge details.

• Tuning:
  To tune this threshold, one would have to look for dead pixels in the pipeline output with DPC disabled, if any are found, enable DPC and progressively reduce the threshold until the dead pixel is corrected. But very low threshold would have a blurring effect due to the False positives.

• Lower Threshold → catches more defects but risks blurring.

• Higher Threshold → fewer corrections but may miss defects.

The identification of dead pixels is crucial for tuning the threshold in order to achieve a good balance between true positives and false positives. However, when dead pixel locations are not available, extreme pixel values may be artificially introduced to test the algorithm's functionality.

Kind regards,
Maria