Camera sensors are rarely perfect. Sensors will have various pixel defects such as dead pixels, hot pixels, etc. Some common sensor pixel defects:
- Dead pixels do not respond to light at all and they don't provide any information. Dead pixels can be black, white (or let's say the maximum output value) or any intermediate value (also called stuck pixels).
- Hot pixels respond do light normally but suffer from excessive dark current and can saturate at reasonable exposures even in the dark.
- RTS pixels respond to light and provide once in a while a sequence of correct values but they can randomly jump up and down with a well defined offset. RTS also can be in a dark current where the dark current value randomly jumps between few discrete values.
- Wide variance noise pixels have in average the right response to light but their noise is much larger than for the other pixels.
- Blinking pixels can be either dead blinking if they jump randomly between two dead states or blinking operating if they jump between the right value and a dead state.
- Clipping pixels behave normally up to a certain value (resp. from a certain value) where they will clip. They are only usable below (resp. above) their clipping value.
- The pixels that start at a too high value sometimes only have excessive offset (can be caused by an excessive FD leakage, or memory node leakage in global shutter pixels.)
While the list is quite exhaustive and covers most of the known effects, there are few more that could be added:
- Pixel with excessive image lag - can manifest itself as a non-linearity at low light. Can only appear at low temperatures. Depending on the array timing, can only appear in high frame rate modes.
- Large crosstalk pixels - some pixels can suffer from too much coupling between them
- Defective color response - can come from defects in color filter
- Anomalously high or low photoresponse (too high PRNU) - can be caused by contamination particles masking a part of the light, or defects in metals, light pipes, etc.
- Dark current that non-linearly depends on the integration time (for example, starts from high and getting lower after a couple of ms.)