1 00:00:00,510 --> 00:00:06,270 Hi and welcome to Chapter 10, where we take a look at dilation, erosion and edge detection. 2 00:00:06,840 --> 00:00:08,640 So let's go ahead and open this notebook. 3 00:00:08,890 --> 00:00:15,750 Already have it open here and you can see there's two more things added to the to this topic that we're 4 00:00:15,750 --> 00:00:16,260 learning here. 5 00:00:16,770 --> 00:00:22,200 It's going to be opening and closing, which are actually combinations of dilation and erosion, but 6 00:00:22,200 --> 00:00:24,450 you don't know what the dilation and erosion is yet. 7 00:00:24,450 --> 00:00:25,740 So let's take a look at this. 8 00:00:25,800 --> 00:00:33,540 So firstly, let's download our images and see all the major functions and all libraries, and that's 9 00:00:33,540 --> 00:00:33,840 done. 10 00:00:34,440 --> 00:00:36,510 So let's take a look at the definitions here. 11 00:00:36,660 --> 00:00:37,740 So let's zoom in a bit. 12 00:00:38,220 --> 00:00:42,630 Dilation adds pixels to the boundaries of an object in the image. 13 00:00:43,020 --> 00:00:50,610 Erosion removes pixels at the boundaries of an object, and opening is this erosion, followed by dilation 14 00:00:50,610 --> 00:00:52,680 and closing this dilation followed by erosion. 15 00:00:53,040 --> 00:00:54,930 So you can see this illustrated right here. 16 00:00:55,320 --> 00:00:56,540 This is the original image here. 17 00:00:56,550 --> 00:00:57,540 This is a number five. 18 00:00:58,020 --> 00:01:02,460 You can see the downward erosion issue, which we look at this. 19 00:01:03,090 --> 00:01:05,270 It removes the pixels from the boundaries. 20 00:01:05,280 --> 00:01:08,130 So all of these things that have boundaries here are gone. 21 00:01:08,730 --> 00:01:09,360 You can see it. 22 00:01:09,540 --> 00:01:11,250 This first column has gone. 23 00:01:11,550 --> 00:01:14,280 This top arrow here is gone as well. 24 00:01:14,790 --> 00:01:17,910 This one is gone, so everything basically thins out. 25 00:01:18,390 --> 00:01:19,560 So that's what erosion does. 26 00:01:19,560 --> 00:01:22,510 It erodes all the pixel around the objects. 27 00:01:22,530 --> 00:01:27,500 You only left with this in a line here, which luckily still looks like five just a very thin. 28 00:01:27,500 --> 00:01:33,270 The five dilation does the opposite dilation padgett of additional pixels around the boundary here, 29 00:01:33,840 --> 00:01:36,630 basically fattening it up to make it look like this. 30 00:01:37,440 --> 00:01:39,060 So that's pretty simple. 31 00:01:39,690 --> 00:01:44,820 But it's important to realize, though, that remember, I'm using zeros and ones here, but it's really 32 00:01:44,820 --> 00:01:48,060 zero and 255 for 255 wouldn't have fit here. 33 00:01:48,360 --> 00:01:55,740 So this relative binary scale, remember this is the white is what the algorithm sees as the image, 34 00:01:55,770 --> 00:01:56,790 as the object. 35 00:01:57,000 --> 00:02:05,940 I should say yes, because a lot, lot, lot of times people have people who frequently think if it's 36 00:02:05,940 --> 00:02:10,950 a black line, like if this was inverted, inverted and it was a black five on a white background, 37 00:02:11,490 --> 00:02:16,440 if you wrote it, you would get a thinner five when in fact you would get the opposite people, lots 38 00:02:16,440 --> 00:02:20,230 of people for a long time to open to the open to these algorithms. 39 00:02:20,730 --> 00:02:23,970 What's in reverse to what actually said it did, but it's not true. 40 00:02:23,970 --> 00:02:25,380 It's because it needs to. 41 00:02:25,380 --> 00:02:31,680 Actually white is what the object is, not the black counter-intuitive because we tend to think black 42 00:02:31,680 --> 00:02:34,980 is the object that we wrote like with a pair of black pen. 43 00:02:35,460 --> 00:02:36,450 But it's the opposite. 44 00:02:37,410 --> 00:02:44,100 So let's just run this code and you can see we've implemented our own erosion and deletions here together 45 00:02:44,100 --> 00:02:45,120 with opening and closing. 46 00:02:45,120 --> 00:02:50,910 And I'll explain opening closing before at first before I explain the rest of the code. 47 00:02:51,450 --> 00:02:53,250 So you can see this is the original here. 48 00:02:53,370 --> 00:03:00,870 Handwritten open CVI eroding it turns it out, significantly diluting it makes it nice, and Taco Bell 49 00:03:00,870 --> 00:03:05,190 basically molds both the font and opening you. 50 00:03:05,190 --> 00:03:11,820 Now opening the second definition of opening again, opening is erosion, followed by deletion. 51 00:03:12,330 --> 00:03:15,150 So before you go to the image again, hopefully you forgot what it looks like. 52 00:03:15,600 --> 00:03:22,530 If you want to do erosion before deletion, erosion turns it out and dilation puts it back up. 53 00:03:22,920 --> 00:03:27,970 So you can publish, you should get something that looks similar to the original, but it doesn't quite. 54 00:03:27,990 --> 00:03:33,030 And the reason for that full is because when you wrote this here, you remember we were erasing pieces 55 00:03:33,030 --> 00:03:33,510 of the image. 56 00:03:33,840 --> 00:03:36,480 If you raise piece of this image, you're never going to get it back in. 57 00:03:36,480 --> 00:03:42,060 No amount of deletion is going to add pixels to something that doesn't exist, which is why you end 58 00:03:42,060 --> 00:03:44,310 up with this here, which looks kind of patchy. 59 00:03:45,120 --> 00:03:46,500 No closing. 60 00:03:46,560 --> 00:03:47,580 Closing is the opposite. 61 00:03:47,970 --> 00:03:49,080 Closing, what do you do? 62 00:03:49,110 --> 00:03:53,490 You delete this, so you kind of flattened image first and then you turn it out. 63 00:03:53,520 --> 00:03:59,190 So closing, actually, it would look pretty decent, meaning mainly because you painted it first and 64 00:03:59,190 --> 00:04:00,030 then you turned it out. 65 00:04:00,300 --> 00:04:01,860 So that's how you get to closing here. 66 00:04:02,340 --> 00:04:04,620 So let's take a look at the code and see how this is done. 67 00:04:04,770 --> 00:04:05,670 So it's quite simple. 68 00:04:05,700 --> 00:04:11,970 Firstly, you define a kernel, and the kernel basically specifies how much of of the erosion of that 69 00:04:11,970 --> 00:04:13,650 relation we want to do. 70 00:04:14,220 --> 00:04:18,540 Iterations basically tells you how much time you want to do it, so you can remember you can do it in 71 00:04:18,540 --> 00:04:20,950 a region over and over and over again. 72 00:04:21,780 --> 00:04:26,880 You just basically eroding it and then eroding it again or diluting dilating again, vice versa. 73 00:04:27,720 --> 00:04:34,980 So all we need to do is you see v2.0 would use the input image, have reloaded, set the kernel using 74 00:04:34,980 --> 00:04:36,180 a five by five kernel? 75 00:04:36,210 --> 00:04:38,250 That's basically how it how it works. 76 00:04:39,030 --> 00:04:44,160 So the kernel is a matrix that we use for the 2D convolution that is done in these functions here. 77 00:04:44,640 --> 00:04:50,520 So the logic, you know, you'll actually notice it's going to be more dilation or more iteration being 78 00:04:50,520 --> 00:04:50,940 applied. 79 00:04:51,480 --> 00:04:55,380 Similarly, we can do dilate using almost the exact same function. 80 00:04:55,680 --> 00:04:59,070 Actually, it pretty much is the same arguments, except. 81 00:04:59,780 --> 00:05:05,530 Dilate, we replace the road with dilate and then to do opening and closing. 82 00:05:05,890 --> 00:05:08,470 It's a bit different, but quite simple as well. 83 00:05:09,010 --> 00:05:16,340 You use CV 2.0 Morphology X and then just put into the image you set when you were to open a close. 84 00:05:16,340 --> 00:05:21,040 So it's more when the score open and more open to score close, depending on what you're doing. 85 00:05:21,640 --> 00:05:22,560 And that's it. 86 00:05:22,600 --> 00:05:23,510 Quite simple to run. 87 00:05:24,280 --> 00:05:25,990 So let's take a look at edge detection. 88 00:05:26,530 --> 00:05:30,670 So before I took a look at Kanae edge detection, I want to explain to you what edge detection is. 89 00:05:31,090 --> 00:05:33,340 So imagine this is a line, you know, image. 90 00:05:34,300 --> 00:05:36,910 So imagine this is the blue line we're looking at here. 91 00:05:37,930 --> 00:05:40,750 This is the intensity across the blue line you can see in the graph here. 92 00:05:41,020 --> 00:05:46,990 Density 255 is a brightness, so you can see it's bright and it's dark and it's bright again. 93 00:05:47,740 --> 00:05:52,900 So I don't know if you remember your mats from high school university, but if you do the first or the 94 00:05:52,900 --> 00:05:57,950 derivatives, which is the rates of change these points here you can see this is this. 95 00:05:58,120 --> 00:06:01,060 This point here is it's consistent as well. 96 00:06:01,390 --> 00:06:06,520 Firstly, here, but then there's a high rate of change here, which is why you reach this minimum point 97 00:06:06,520 --> 00:06:06,760 here. 98 00:06:07,270 --> 00:06:10,330 And then similarly, at this point, it goes up again quite quickly. 99 00:06:10,750 --> 00:06:12,160 And that's where you get the maximum points. 100 00:06:12,160 --> 00:06:16,030 So that's the rates of change being given here by first or the derivatives. 101 00:06:16,750 --> 00:06:20,980 So that's one way I can show you illustrating what an edge is. 102 00:06:21,160 --> 00:06:22,090 So you know what an edge. 103 00:06:22,480 --> 00:06:27,190 And you know, algorithmic algorithmically how we can find the edge using the first or the derivatives. 104 00:06:28,150 --> 00:06:29,530 So what are we going to do? 105 00:06:29,560 --> 00:06:32,530 We're going to take a look at Cannae edge detection, which is an algorithm. 106 00:06:32,950 --> 00:06:34,570 Very good algorithm, actually. 107 00:06:34,720 --> 00:06:35,590 It's a bit complicated. 108 00:06:35,590 --> 00:06:40,290 It's actually entire people on it that can give us the edges here. 109 00:06:40,300 --> 00:06:42,250 So let's take a look at a county edge function. 110 00:06:42,820 --> 00:06:46,870 So the cutting edge function should have been augments, the better. 111 00:06:47,260 --> 00:06:53,200 The the first argument is the input image, the second and two documents that women and Maxwell, respectively. 112 00:06:53,530 --> 00:06:58,630 The fourth argument, which is optional, is the aperture size, which is the size of the symbol colonel 113 00:06:58,630 --> 00:07:02,860 used for the image gradients finding image gradients default history. 114 00:07:03,190 --> 00:07:05,710 But we don't use them, but you use the default value here. 115 00:07:06,430 --> 00:07:08,290 So what we can see is kind of edge. 116 00:07:08,290 --> 00:07:09,460 We can suggest for shields. 117 00:07:09,460 --> 00:07:16,660 Yeah, just a minute max values here, and let's run just changed thresholds to take a look so we can 118 00:07:16,660 --> 00:07:17,650 widen the threshold. 119 00:07:17,650 --> 00:07:20,050 We can narrow the threshold and use them. 120 00:07:20,050 --> 00:07:20,880 Ed Schultz Yeah. 121 00:07:20,890 --> 00:07:22,660 And let's see what the results are. 122 00:07:25,840 --> 00:07:33,370 So look, this is the canny edge of the first image here with those parameters and put sort of parameters 123 00:07:33,370 --> 00:07:38,200 we used and you can see it looks, it looks quite good actually, in my opinion. 124 00:07:39,100 --> 00:07:41,490 This one would have wide thresholds look similar. 125 00:07:41,490 --> 00:07:42,850 It's not much different. 126 00:07:43,510 --> 00:07:48,220 This one with narrow thresholds definitely has less lines, and it's it's a bit harder to make out that 127 00:07:48,220 --> 00:07:49,250 there's a person here. 128 00:07:49,300 --> 00:07:55,420 You can probably see the car here, but it's a bit more visible, at least sort of in the sort of images 129 00:07:56,020 --> 00:08:01,390 and forward sort of thresholds we use basically again gave it a lot more lines as well. 130 00:08:01,780 --> 00:08:07,300 So there's something called up to a canny, not a canny was just a quick method that someone developed 131 00:08:07,780 --> 00:08:08,350 online. 132 00:08:08,410 --> 00:08:13,120 I think this year on StackOverflow should give you a lot of credit, but I can't remember and I can't 133 00:08:13,120 --> 00:08:14,170 find a way to sources. 134 00:08:14,170 --> 00:08:17,980 I remember pasting it into some of my code and using it. 135 00:08:18,790 --> 00:08:25,180 So what it does, it basically calculate it does do some blurring first, then gets a median image value 136 00:08:25,630 --> 00:08:29,040 and then uses its max value times. 137 00:08:29,050 --> 00:08:34,060 0.66 then points the max value four and five 1.3 three. 138 00:08:34,840 --> 00:08:40,390 So what we do after that is basically use these to get the thresholds or more thresholds range from 139 00:08:40,390 --> 00:08:41,500 zero to 255. 140 00:08:42,100 --> 00:08:45,660 So we're going it's going to get the values from that or to a function here. 141 00:08:46,300 --> 00:08:50,650 And let's see how it works because setting these strategies for kind of isn't always easy. 142 00:08:51,160 --> 00:08:55,630 So this looks quite nice, in my opinion, but you probably have a different opinion. 143 00:08:55,900 --> 00:09:00,100 Feel free to explore, even tweak this auto kidney function function. 144 00:09:00,580 --> 00:09:01,750 Tweak these thresholds. 145 00:09:01,990 --> 00:09:06,900 It's always good to experiment with of these functions on your own and with different images together. 146 00:09:07,150 --> 00:09:08,800 Get a few of them and what they do. 147 00:09:09,340 --> 00:09:09,850 Thank you. 148 00:09:09,850 --> 00:09:13,730 And let's move on to the next lesson, which is contours. 149 00:09:13,780 --> 00:09:14,770 Very important lesson. 150 00:09:15,160 --> 00:09:15,730 Stay tuned.