It is computationally exprensive, but you could

simulate the fovea. One sensor in the middle,

responsible for one pixel. In the neighborhood

other sensors, responsible for a bigger number

of pixels. The farther from the midpoint,

the greater the area one sensor is responsible

for.

A sensor measures the brightness of a pixel,

either from a grey value image or the

3 channels of RGB. And puts the measured

value into the input layer of a NN.

Learning: could be done supervised. You

click with the mouse pointer on a pixel and

classify it as interesting or noninteresting.

The computer stores a training pattern,

a pair of sensor data and desired output.

After the NN has learned all training

data, learned to classify input vectors,

it can be used to scan the whole image

and map each point via feed-forward.

Regards,

Joachim