When I was in college I wrote some software that did something sorta like this. My target was to post-process video content from TV shows with the channel logos in the corners.
My approach was to detect the logos based on rudimentary pattern-detection (mostly finding parts of the screen the didn't change over a period of time) then calculate the replacement pixel-color based on the other pixels around the edges of this logo, weighting their impact on the replacement pixels by their proximity to the pixel being calculated.
It worked reasonably well ... albeit much slower than the demo in this video ... and transparent logos were then introduced complicating the calculations. Also, sitcoms worked much better than sports. Their cameras were more stationary making detection and pixel-color replacement-calculation easier.
My approach was to detect the logos based on rudimentary pattern-detection (mostly finding parts of the screen the didn't change over a period of time) then calculate the replacement pixel-color based on the other pixels around the edges of this logo, weighting their impact on the replacement pixels by their proximity to the pixel being calculated.
It worked reasonably well ... albeit much slower than the demo in this video ... and transparent logos were then introduced complicating the calculations. Also, sitcoms worked much better than sports. Their cameras were more stationary making detection and pixel-color replacement-calculation easier.