Proof of concept
 

 

On the next example we would verify if the concept of face recognition can reliably and consistently be applied to buildings to detect physical changes. The following tests were performed under ideal conditions were lighting conditions remained the same throughout the test.

Given the following training set

 We obtained the mean image, which is used to find the difference between the input images and the ones in the training set.

Something that we noticed was that the system is extremely sensitive to movement. If you are interested in finding small errors you will have to get sure that the object you are testing is not moving at all. For this test, we fixed our model by placing it in the corner of a wood desk and fixing it by having an edge, on two of the sides, that went down as shown in the figure bellow. Later we would show you what happens when similar tests are not performed as rigorously as this one.

This side faced downwards to prevent any displacement

Here are the results that were obtained, followed by a brief discussion about them.

Test #	Description of the change made to model	Min distance	Max distance
1	training set image	3.5285	3.5357
2	training set image	3.5287	3.5362
3	training set image	3.5234	3.5356
4	training set image	3.5285	3.5358
5	training set image	3.5287	3.5362
6	training set image	3.5286	3.5360
7	training set image	3.5282	3.5356
8	training set image	3.5282	3.5356
9	training set image	3.5286	3.5359
10	star in the middle was rotated	3.5930	3.6005
11	star in the middle was removed	3.5747	3.5823
12	2 blue rods and 3 white rods were removed	3.5607	3.5684
13	1 white rod and 2 green ones were removed	3.5470	3.5548
14	1 green pin was removed	3.5418	3.5495
15	image has no defects	3.5418	3.5494
16	2 small fires were introduced	3.5457	3.5534
17	a piece of the left side was removed	3.5494	3.5571
18	1 white pin is missing	3.5453	3.5528
19	2 blue rods were removed	3.5393	3.5469
20	image has no defects	3.5352	3.5429
21	a star and its pins were removed	3.5555	3.5631
22	the star was put back but without the white pins	3.5610	3.5687
23	image has no defects	3.5280	3.5355
24	a piece at the bottom was removed	3.5321	3.5393

The first 9 test were performed using an image of the training set as the input image. As we can see, the min and max values are less than the ones that follow. We performed 12 test where we removed some of the pieces of the structure and 3 where it was left untouched.

When looking at the results, if we set our limit for known images to be bellow 3.5418, almost all defects are detected except the last one. We should note that in case 14 a conclusion can not be made based on the minimum only. To conclude, you should fix your structure very well if you want to be able to distinguish very minor changes.

Click here to download the training images
Click here to download input images

Next test