We can see that it is a very small image it has only got 55 rows of pixels and 41 columns of pixels. That is, they lie between 0 and 1 and then let’s display that.Īnd there we see the crowd scene, next I am going to load an image of the robot Bender and that is in a PNG file also going to convert that to double precision and I have loaded that into the workspace variable Bender. I want to convert the colour image to grey scale and I want to convert all the pixels to double precision values.
I am going to load it from the file wheres_walle.png, it is a PNG format file. The first thing I am going to do is load the crowd scene into a workspace variable which I am going to call crowd. So this is the template, this is the pattern that I am going to go looking for in the image. I am going to zoom in as much as I can on his head and blacken out all of the other robots that I am not interested in. So we first of all need to create the template, so here I have chopped Bender out of the scene, so it is mostly Bender and a few of his neighbours as well.Īnd then what I am going to do is to use a Photoshop-like program to remove all of the robots in that scene that are not Bender. Here we revisit the original puzzle and there is Bender highlighted. Typically, a value above 0.8 is considered to be a reasonable match. 0 means that the two images are not very well correlated. +1 means that the images are identical and -1 means that one image is the negative of the other one. It looks much more complicated, but it has some advantages over the two simple measures above. A quite similar approach is the operator called ssd(), sum of square differences, and instead of using the absolute value operation we are taking the sum of the squares.Īnother measure is the zero - mean normalised cross correlation, commonly abbreviated to zncc(). So 0 means perfect match a big value means a less perfect match. If they are dissimilar, S will have a value greater than 0. So if the images are identical, the similarity measure will have a value of 0. Of course, they must be the same size, take their absolute value and sum it up. To do this we need an image similarity measure and the simplest one is called the sum of absolute differences, sometimes abbreviated to SAD or sad() and quite simply we take the difference between the corresponding pixels in the two images that we are trying to compare. So how do we compare two images, how do we tell that a particular input window W looks like Bender or doesn’t look like Bender? So the template will be the face of Bender and we are going to compare the face of Bender at every single location across the scene. Every input window, we are going to compare it with our template and the similarity score is going to be placed into the output image. We are going to look at the similarity between the particular input window which marches across the image left to right, top to bottom. What we are going to do is for every input window we are going to compare it with a template image, the template being the image, the pattern we are looking for and our function now is a similarity function. We are going to use the spatial operator approach in order to solve this problem.
So rather than us having to scan the crowd to find Bender, let’s see if we can automate the task. Here he is: he is hiding behind a Dalek and standing next to Eve. Let’s consider the problem of trying to find a face in a crowd and, in particular, I want to find the face of the robot Bender.
0 kommentar(er)
