c# - Recognize sub image -

i have quite simple task: goal detect image pattern inside image. in particular, have identify correct side of cassette drug-of-abuse test picture acquired scanner lowest error rate possible. i’m using aforge.net library c#.

doa test picture:

link http://www.atbsrl.com/media/so/doa-test-sample.png

test in progress picture:

link http://www.atbsrl.com/media/so/schema03.png

the cassette faces pretty similar, first (a side) has 3 adulteration strips (with colored areas) , other drug-of-abuse strips (b side).

to pretty sure recognize side, tried mix different approaches, e.g. color matching , template matching, @ moment tests aforge.net didn’t produce expected results because of variability of input image.

do have suggestions?

use convolutional neural network. there nothing can classical image processing have low error rate.

to apply this, collect many images possible (preferably thousand or more; hundred may not enough, ten thousand ideal) of each side, , make text file filenames , labels side (0 front or 1 back). then, take off-the-shelf package caffe or libccv , train on data set. can use network model caffenet in caffe work fine, follow this tutorial retrain trained caffenet application (which easier starting scratch, , can work less data). started, may easiest launch aws instance caffe preinstalled, this one. considering nature of data, think <1% error rate (or possibly lower, if have more data).

if can modify object being detected, use simpler algorithm: example color each face, or put fiducial marks or simple barcodes.