//pseudocode for matrix transformation estimation given pairs of input/output matrices 
//this is more accurate than taking averages but much much slower

//training

in[inst][i]
outputsample[inst][o]
trained[]


for each outputsample (o) // pixel
	for each inputsample (i) //pixel
		for each instance (inst)	//instance
			trained[o]+=out[inst][o]/in[inst][i]
	trained[o] = trained[o] / number_instances / number_outputs 

//to get result

newin[i]
result

for each outputsample(o)
	result[o] = newin[o] * trained[o]