You do not take image compression and image content geometry into account. A good T would randomize the nearby color values, so the picture would look like a colorful random noise, with all the edges visible on it. The problems are:
1. Image compression algorithms rely on shades of different form to a great extent. These are replaced by noise, and since compression is lossy, inverse of T would fail.
2. Even if the decompression would not be a problem, I think it would be relatively easy to write an AI algorithm that could find out T using the image geometry info still present. If not the original colors, but a grayscale stream could easily be reconstructed.
This is like a dictionary-based text encription, which is easy to crack using statistics.
It is no wonder noone uses such systems, but rely on mathematical constructs and stream encryption systems based on these.
