1
00:00:01,430 --> 00:00:06,440
So moving on to conclusions and bring that the first thing to notice that the convolution isn't and

2
00:00:06,490 --> 00:00:12,620
we can see the function Pizzi it's an actual mathematical operation once that's performed between two

3
00:00:12,620 --> 00:00:19,560
functions and it produces a function which is obviously an output coming out of the original operation.

4
00:00:19,580 --> 00:00:21,230
So this is sort of illustrated here.

5
00:00:21,380 --> 00:00:22,760
We have an image here.

6
00:00:23,070 --> 00:00:29,250
Then we run it by a function of convolution and our function has offices in kernel size.

7
00:00:29,540 --> 00:00:31,260
And what do we mean by kernel size.

8
00:00:31,310 --> 00:00:36,840
Well this really here represents a 10 by 10 image or pixel image here.

9
00:00:37,460 --> 00:00:39,340
And this is a 5 by 5 you know.

10
00:00:39,780 --> 00:00:44,560
So what this does here the convolution basically operates on one pixel at a time.

11
00:00:44,600 --> 00:00:53,690
So operating on this here of a pixel to do so deconvolution basically operates on this pixel but it

12
00:00:53,690 --> 00:00:59,230
takes into consideration all the pixels around it and its calculation you'll see why that's important.

13
00:00:59,240 --> 00:01:06,380
When we get introduced to blurring which is what we're about to do now so there is an operation which

14
00:01:06,380 --> 00:01:11,320
you obviously figured out that is an image so you can clearly see here.

15
00:01:11,360 --> 00:01:13,060
This is a shopping shift an elephant.

16
00:01:13,140 --> 00:01:19,570
I took a call just to do the zoo and this is the blurred image of the elephant.

17
00:01:19,970 --> 00:01:22,250
No actually she's not quite blurry.

18
00:01:22,610 --> 00:01:27,200
So this is an example of a matrix The Matrix is 25.

19
00:01:27,260 --> 00:01:29,090
Sorry five by five.

20
00:01:29,310 --> 00:01:35,190
Now we multiply meet the new matrix by one of the twenty five.

21
00:01:35,400 --> 00:01:40,740
And that's what we normalize that's called normalization of okeydoke Matrix because we want all the

22
00:01:40,740 --> 00:01:45,450
elements in this matrix to sum to one which I mentioned here.

23
00:01:45,480 --> 00:01:53,010
Now the first method I'm going to introduce you to is open see these filter Tuti takes an image and

24
00:01:53,010 --> 00:01:54,910
runs it past the kernel here.

25
00:01:54,930 --> 00:01:57,360
So we're about to introduce Saturno code.

26
00:01:57,360 --> 00:01:59,610
So let's go Tokuda and do so now.

27
00:02:00,840 --> 00:02:03,410
So let's look at this blurring of code here.

28
00:02:03,420 --> 00:02:08,460
We're going to use the open see the filter to see which is the most basic form of Pluton we can you

29
00:02:08,460 --> 00:02:10,080
do in if NCB.

30
00:02:10,220 --> 00:02:12,830
I'm going to do it by using a tree by treacle.

31
00:02:12,970 --> 00:02:18,060
Once you normalize by dividing by 9 so we're going to see the effects of blurring an image with a tree

32
00:02:18,060 --> 00:02:22,110
bettery you know and then beggaring it with a larger seven by seven canal.

33
00:02:22,200 --> 00:02:25,220
But you will also know all of these before dividing by 49.

34
00:02:26,190 --> 00:02:28,100
So let's run this one and take a look at it.

35
00:02:28,170 --> 00:02:29,060
In.

36
00:02:29,370 --> 00:02:30,970
This is our original image.

37
00:02:31,050 --> 00:02:33,550
You can see that elephant creases quite clear.

38
00:02:33,550 --> 00:02:37,750
So it's has a lot of detail in this image so that's run it by the tree by tree.

39
00:02:39,100 --> 00:02:40,580
And you can see it's definitely blue.

40
00:02:40,600 --> 00:02:46,690
It's not a substantial amount of Larry but it's definitely noticeable when you compare it to the original

41
00:02:46,690 --> 00:02:47,530
image.

42
00:02:47,980 --> 00:02:49,750
So let's take a look at the seven by seven.

43
00:02:49,750 --> 00:02:53,510
We expect a lot more heavy blurring and that's exactly what we see.

44
00:02:53,820 --> 00:02:58,160
This is probably what a person who needs high prescription glasses might see.

45
00:02:58,690 --> 00:03:01,800
So that's a fast learning algorithm.

46
00:03:01,840 --> 00:03:03,360
We're going to take a look at a couple of now

47
00:03:06,190 --> 00:03:12,260
writes If you scroll down over here you'll actually see some more blurring functions that are available

48
00:03:12,320 --> 00:03:13,810
in open C.v.

49
00:03:13,850 --> 00:03:16,840
So the first one we're going to use is actually a box filter.

50
00:03:16,890 --> 00:03:20,390
So it's called the CB2 to Blair function and it's an averaging filter.

51
00:03:20,390 --> 00:03:26,660
What it does it averages the pixels in a box that runs between each pixel.

52
00:03:26,780 --> 00:03:28,470
So it puts the input image here.

53
00:03:28,640 --> 00:03:30,240
Then it takes to really find here.

54
00:03:30,260 --> 00:03:33,730
This is a box size and averages on all the pixels here.

55
00:03:33,740 --> 00:03:36,770
So it sort of introduces a heavy blurring effect.

56
00:03:36,770 --> 00:03:44,630
It's actually quite quite a drastic learning effect which you will see shortly and that's why there's

57
00:03:44,660 --> 00:03:46,220
also the Gaussian blurring here.

58
00:03:46,460 --> 00:03:51,920
So what this does instead of having a uniform boxful to actually uses a Gaussian boxful to I'll show

59
00:03:51,920 --> 00:03:54,620
you what a Gaussian box is in a sleights.

60
00:03:54,650 --> 00:03:59,110
But for now just imagine it's not going to have such a harsh effect as this blue function.

61
00:03:59,360 --> 00:04:06,640
Now median blue medium blue also uses a box filter type I mentioned here.

62
00:04:06,640 --> 00:04:09,250
However it doesn't actually average out everything in the box.

63
00:04:09,250 --> 00:04:15,190
What it does it finds a median value of all that image all the pixels in that box and then it actually

64
00:04:15,690 --> 00:04:17,800
puts median value as a pixel.

65
00:04:17,800 --> 00:04:21,450
So it's sort of a nice balance between this and averaging.

66
00:04:21,450 --> 00:04:23,240
Sorry Dessen Gaussian.

67
00:04:24,010 --> 00:04:29,160
And lastly we have bilateral blurring bilateral blurring is actually quite useful.

68
00:04:29,410 --> 00:04:34,210
And what it does it actually preserves all strength it removes noise quite well and strengthens the

69
00:04:34,210 --> 00:04:35,830
edges in the image.

70
00:04:35,830 --> 00:04:39,510
So let's take a look at Rhonda squid and you'll see what I'm talking about.

71
00:04:39,550 --> 00:04:46,540
So this is an averaging box filter here you can see it has pretty much a uniform type effect.

72
00:04:46,540 --> 00:04:52,210
This is a Gaussian during here not to go blurring but actually the seams would actually be less blurry.

73
00:04:52,500 --> 00:04:55,250
But ever is a seven by seven in this example here.

74
00:04:58,220 --> 00:05:00,110
And this is a medium bearing here.

75
00:05:00,110 --> 00:05:04,300
Now if you notice medium bearing sort of has a kind of pinata type effect to it.

76
00:05:04,550 --> 00:05:10,400
And that's because it's finding the median value in those segments of the image.

77
00:05:10,420 --> 00:05:15,430
So if you wanted to ever introduce sort of a carpenter type effect maybe a medium medium blue would

78
00:05:15,430 --> 00:05:17,210
be quite useful.

79
00:05:17,270 --> 00:05:22,600
And let's take a look at our bilateral hearing here.

80
00:05:22,600 --> 00:05:29,600
Now as you can see an bilateral blurring it's actually sort of a anti losing type effect on this image.

81
00:05:29,770 --> 00:05:34,720
But what you can tell is that it definitely does preserve vertical and horizontal lines strongly here.

82
00:05:35,100 --> 00:05:40,150
But images of areas of the ears and different areas and the elephant is actually quite blue.

83
00:05:40,210 --> 00:05:45,690
However you can see vertical lines here are actually quite strong.

84
00:05:46,110 --> 00:05:49,840
So here's a slide summarizing to just those different types of blurring we just saw.

85
00:05:50,160 --> 00:05:52,300
So these are the same notes they just discussed here.

86
00:05:52,500 --> 00:05:58,920
On each of the four methods however I told you I'd show you what a gaussian kernel looks like and this

87
00:05:58,920 --> 00:06:00,870
is exactly what a Garcia-Pichel is here.

88
00:06:00,870 --> 00:06:05,520
So you remember previously before and other criminals everything was the same value here.

89
00:06:05,580 --> 00:06:11,170
However goshi and kernel's have a peak toward the center and is slowly peter off to the edges here.

90
00:06:12,040 --> 00:06:16,080
This is sort of a treaty representation of what a Gaussian looks like.

91
00:06:16,330 --> 00:06:18,170
And this is a matrix form here.

92
00:06:18,280 --> 00:06:23,980
And again we normalize by some of all the elements here which in this case is 273.

93
00:06:24,100 --> 00:06:29,360
So heading back to bite in the books is one last image Geering method to just wish to discuss.

94
00:06:29,530 --> 00:06:31,590
And that's one called imaged invoicing.

95
00:06:31,840 --> 00:06:36,630
This is open Zeevi function here first and one means the noise in code.

96
00:06:36,640 --> 00:06:38,480
It's quite a mouthful here.

97
00:06:38,560 --> 00:06:42,630
And what this does this actually originates from computational photography methods.

98
00:06:42,910 --> 00:06:48,820
And as you know using cell phone cameras and digital cameras there's no way those tiny camera sensors

99
00:06:48,820 --> 00:06:53,120
can match the noise levels of complex big DSL or cameras.

100
00:06:53,140 --> 00:06:58,810
So what they use to compensate is these complicated algorithms that sort of try to get the best optimized

101
00:06:58,810 --> 00:07:02,090
noise levels so you don't see all that green in the image.

102
00:07:02,110 --> 00:07:06,220
So when we run this function it actually takes some time of research on it in the background.

103
00:07:06,220 --> 00:07:09,070
You get this nice very clean looking image here.

104
00:07:10,120 --> 00:07:13,300
So that the effects of running this algorithm.

105
00:07:13,560 --> 00:07:15,090
Now there's quite a few variations here.

106
00:07:15,100 --> 00:07:20,240
I wouldn't get into the details exactly right now these are some notes on the premises here.

107
00:07:20,260 --> 00:07:21,410
This is filled the strings.

108
00:07:21,430 --> 00:07:23,750
And these are some of the components you can play with.

109
00:07:24,310 --> 00:07:26,070
And that's it for blurring.

110
00:07:26,160 --> 00:07:28,480
Let's move on to shopping which is the opposite of Barry.