1 00:00:00,120 --> 00:00:05,850 So now let's move on to troubling binaries, Asian and adaptive threshold in which is chapter nine and 2 00:00:05,850 --> 00:00:06,930 nine to not book here. 3 00:00:07,380 --> 00:00:08,820 So let's go to open this up. 4 00:00:09,000 --> 00:00:14,880 I've already got it open here, so let's just disclose that and we can begin. 5 00:00:15,360 --> 00:00:17,880 So firstly, what is binaries issued? 6 00:00:18,360 --> 00:00:21,990 So you may think it's making something binary and you'd be right? 7 00:00:22,440 --> 00:00:24,330 Well, what are we making binary exactly? 8 00:00:24,360 --> 00:00:25,860 Well, we're making the colors. 9 00:00:25,860 --> 00:00:31,110 Or should I say the pixels in an image zero or one? 10 00:00:31,650 --> 00:00:32,720 So how do we do that? 11 00:00:32,730 --> 00:00:37,890 So before we, let's just run this code just to declare this stimulus function. 12 00:00:38,880 --> 00:00:43,890 And now we move on onto this slide, where I'm going to show you an example of what binaries station 13 00:00:43,980 --> 00:00:44,340 is. 14 00:00:44,790 --> 00:00:47,580 So take a look at the original image right here. 15 00:00:48,120 --> 00:00:53,280 It's a smooth gradient from white to black with all the shades of gray in between. 16 00:00:54,360 --> 00:00:57,660 And now you see half of it is black, half of it is white. 17 00:00:58,230 --> 00:01:00,410 So that's exactly what finalization does. 18 00:01:00,440 --> 00:01:07,830 It turned that image into basically everything above a certain color threshold or intensity, which 19 00:01:07,830 --> 00:01:09,810 is probably 127 here. 20 00:01:10,440 --> 00:01:13,020 Everything above it basically goes to white. 21 00:01:13,200 --> 00:01:15,870 Everything below 127 goes to black. 22 00:01:16,350 --> 00:01:22,500 That's how we get this type of photocopy type effects on using computer vision, using computer vision 23 00:01:22,500 --> 00:01:25,740 algorithms to create a binary vision type effect. 24 00:01:26,220 --> 00:01:29,430 So we're going to take a look at the other types of binary session right now. 25 00:01:30,000 --> 00:01:34,980 This inverted binary addition, which is basically the opposite instead of everything going to white, 26 00:01:34,980 --> 00:01:35,950 everything goes to black. 27 00:01:35,970 --> 00:01:40,320 And so if everything going to black on this end, it goes to white. 28 00:01:40,360 --> 00:01:42,570 So you just sort of flip that around. 29 00:01:43,470 --> 00:01:46,440 Now what is trunk now? 30 00:01:46,620 --> 00:01:49,330 Truncation may not be immediately obvious. 31 00:01:49,350 --> 00:01:55,530 What is what's happening here, but if you look carefully, you can see it move some black degree and 32 00:01:55,530 --> 00:01:57,990 then it steers degree right around here. 33 00:01:58,320 --> 00:02:04,740 But retrogressive image and trunk basically means instead of sending everything to white to 255, which 34 00:02:04,740 --> 00:02:10,110 is a value up here, it sends everything to the last highest value, which is a truncation threshold, 35 00:02:10,110 --> 00:02:11,460 which is 127 mm here. 36 00:02:11,910 --> 00:02:15,360 So anything above 127 becomes 127. 37 00:02:16,170 --> 00:02:22,350 Now let's look at two zero two zero sort of does something similar, but it's very opposite exactly. 38 00:02:22,350 --> 00:02:29,760 As you can see here, everything below 127 goes to black, and it leaves this nice, smooth gradient, 39 00:02:30,240 --> 00:02:33,650 which which was existing in the original image right here. 40 00:02:33,660 --> 00:02:36,300 So everything above the threshold remains untouched. 41 00:02:36,780 --> 00:02:39,030 Everything below the threshold goes to zero. 42 00:02:39,300 --> 00:02:41,910 That's why it's close to zero and to zero. 43 00:02:41,910 --> 00:02:45,450 Invert is usually is exactly the opposite of what this is. 44 00:02:45,840 --> 00:02:49,590 So everything above the treasure instead of going to white, goes to black. 45 00:02:50,130 --> 00:02:58,230 So you're making all the values above 127 zero zero here and leaving all the values below 127, as is. 46 00:02:58,410 --> 00:02:59,660 So we don't touch those. 47 00:03:00,330 --> 00:03:02,130 So let's move on to this. 48 00:03:02,130 --> 00:03:03,480 Let's get a scan image here. 49 00:03:03,490 --> 00:03:08,520 So I'm going to use would get which is a nice little command you can use to download files of the internet 50 00:03:08,520 --> 00:03:09,360 to your terminal. 51 00:03:10,770 --> 00:03:15,870 So we just don't have that file and that file is this file scandal, JPEG. 52 00:03:17,100 --> 00:03:18,270 So what we're going to do? 53 00:03:18,900 --> 00:03:23,460 We're going to open this up and display the original image here, which you can see below. 54 00:03:24,600 --> 00:03:30,600 This nice original image here, which is a scan of a photocopy of a book of I believe it's a book, 55 00:03:30,600 --> 00:03:34,680 is a picture I took, yes, of all of my data science textbooks. 56 00:03:35,490 --> 00:03:40,560 And what we're going to do, we're going to apply each of those different thresholds that you saw above. 57 00:03:41,130 --> 00:03:43,650 So let's take a look and see how they perform. 58 00:03:47,030 --> 00:03:49,730 All right, so let's take a look at our thresholds. 59 00:03:50,270 --> 00:03:54,890 So the first one we do was a binaries and threshold at 127. 60 00:03:55,790 --> 00:03:57,110 Now this looks quite bad. 61 00:03:57,410 --> 00:03:59,780 Isn't it quite different from the original image? 62 00:04:00,260 --> 00:04:08,570 But that's just because everything at 1:27 remember, let's go up to the top in binaries issue everything 63 00:04:08,570 --> 00:04:10,400 below a value becomes black. 64 00:04:10,910 --> 00:04:14,690 So what this is seeing is that everything that was below 127. 65 00:04:15,230 --> 00:04:18,380 And when I say 127, I hope, you know, I'm talking about the pixel value. 66 00:04:18,440 --> 00:04:22,700 I remember pixel values and grayscale images range from zero to 255. 67 00:04:23,540 --> 00:04:31,100 So everything below what the 127 intensity goes to black and everything above it remains the same. 68 00:04:31,610 --> 00:04:33,080 That's what actions are. 69 00:04:33,290 --> 00:04:38,420 Everything above it goes to 255, which is why you have these little patches of white in between. 70 00:04:39,410 --> 00:04:41,600 Now let's take a look at the other trestles. 71 00:04:44,140 --> 00:04:51,880 So the inverse binary threshold here at one threshold 127 gives us is actually a pretty nice image. 72 00:04:51,940 --> 00:04:55,960 So it actually looks a lot better than the original threshold that we did. 73 00:04:56,590 --> 00:05:01,190 So you can see all the text now becomes white and all the white spots become black. 74 00:05:01,210 --> 00:05:03,520 Well, the lighter colored spots, I should say. 75 00:05:04,150 --> 00:05:07,930 However, you can see there's some sort of clipping area going on here. 76 00:05:07,960 --> 00:05:09,670 So basically, these are overblown. 77 00:05:10,180 --> 00:05:10,990 So these are bad. 78 00:05:11,170 --> 00:05:12,760 These are literally bad thresholds. 79 00:05:12,760 --> 00:05:18,980 But these distress rules are set to I would say this is too low here that would give us this. 80 00:05:19,270 --> 00:05:22,900 So you need to increase this and it probably would get some of this text texture. 81 00:05:25,550 --> 00:05:26,330 So let's take a look. 82 00:05:26,510 --> 00:05:31,220 The trunk removed would trunk just keeps everything like water, moderate, medium, gray. 83 00:05:31,880 --> 00:05:36,670 So everything above that was supposed to be nice, bright and white kind of stays at gray. 84 00:05:37,010 --> 00:05:40,280 It's a lot of green smudge threshold to zero. 85 00:05:40,880 --> 00:05:44,000 Basically, everything that was below lower value goes to black. 86 00:05:44,000 --> 00:05:48,320 As you can see, the darker is an image here of everything goes to black, whereas everything above 87 00:05:48,320 --> 00:05:50,090 127 remains the same. 88 00:05:50,870 --> 00:05:54,200 And no threshold to zero in this is basically the opposite of that. 89 00:05:54,410 --> 00:05:55,850 So you can see the effect here. 90 00:05:57,050 --> 00:06:02,510 So remember, in those examples above, we had to set that threshold. 91 00:06:02,990 --> 00:06:08,360 But what if we wanted to calculate that this pressure automatically like, let's suppose we were taking 92 00:06:08,360 --> 00:06:13,460 scans of different documents that you have you need, like different ranges of thresholds to get the 93 00:06:13,460 --> 00:06:16,340 exact binaries so you don't. 94 00:06:17,120 --> 00:06:18,710 So, you know, it's nice and neat. 95 00:06:18,980 --> 00:06:22,010 So that's where we use adaptive threshold. 96 00:06:22,250 --> 00:06:28,100 These are little algorithms that actually run some calculations on the image and try to figure out what's 97 00:06:28,100 --> 00:06:29,570 the optimum threshold value. 98 00:06:30,020 --> 00:06:35,960 So OpenCV has two implementations that are quite useful trust us to which the second one I want to show 99 00:06:35,960 --> 00:06:39,190 you which is the better one and adaptive trust means see? 100 00:06:39,950 --> 00:06:43,070 So you can see these are the parameters these threshold and algorithms stick. 101 00:06:43,840 --> 00:06:48,440 I'm not going to go through this in too much detail because you can read the breakdown of this here. 102 00:06:49,100 --> 00:06:50,960 But now let's just test this out. 103 00:06:51,230 --> 00:06:55,340 OK, so let's run this code and explain to cool it as we go along. 104 00:06:56,210 --> 00:07:00,110 So what we're doing here, we're firstly doing a regular threshold value. 105 00:07:00,290 --> 00:07:00,650 All right. 106 00:07:01,250 --> 00:07:05,330 So I should have probably set above what are we doing when you're using these thresholds? 107 00:07:05,750 --> 00:07:08,360 We set two parameters in the threshold here. 108 00:07:08,840 --> 00:07:14,630 You said 127, which is a threshold value and 255 which is a max value that you wanted to go to. 109 00:07:15,350 --> 00:07:17,300 So everything above that goes to 255. 110 00:07:17,360 --> 00:07:23,360 This is a way we can actually control the threshold so you can use these values to actually do threshold 111 00:07:23,530 --> 00:07:24,280 if you wanted to. 112 00:07:25,040 --> 00:07:31,210 So no, actually, I like to do it, but it's a good practice to actually do some Gaussian blur before 113 00:07:31,230 --> 00:07:33,740 threshold limit just to kind of smoothing the image how it looks. 114 00:07:33,740 --> 00:07:40,000 It has less noise in it, but we're not going to do it here, just left it in for offers as as it's 115 00:07:40,010 --> 00:07:40,760 good practice. 116 00:07:41,570 --> 00:07:45,180 So what we're going to do, we're going to take a look at the adaptive threshold. 117 00:07:45,680 --> 00:07:52,610 So if you use the adaptive threshold, then we use it to see V2 Dot Adaptive Threshold used it for the 118 00:07:52,610 --> 00:07:58,190 input image said the maximum value we wanted to go to, said the algorithm we want to use. 119 00:07:58,190 --> 00:08:03,860 This one is key to adaptive trust me and C and what type of threshold we're going to do. 120 00:08:03,870 --> 00:08:05,840 So we're going to do a normal threshold binary. 121 00:08:05,960 --> 00:08:08,750 You can do threshold inverse binary if you wanted to as well. 122 00:08:09,560 --> 00:08:15,470 And what we're going to do is take a look at a tree and five others a block size because those are calculations 123 00:08:16,460 --> 00:08:21,110 parameters to set the calculation how how does how would those two calculations both here? 124 00:08:21,680 --> 00:08:24,920 So you can use block size, which is usually the neighborhood of pixels. 125 00:08:24,920 --> 00:08:31,160 If you wanted to take into consideration a bigger area, you increase the block size and constancy, 126 00:08:31,550 --> 00:08:33,800 which is subtracted and a weighted mean no. 127 00:08:34,220 --> 00:08:38,120 This is something you can really play, which I haven't actually done much experimentation with this. 128 00:08:38,720 --> 00:08:43,740 However, these values with the default values use of open CVS documentation. 129 00:08:43,790 --> 00:08:47,750 So let's just go with them and you can feel free to experiment if you want. 130 00:08:48,410 --> 00:08:51,860 So another take a look at what suits threshold. 131 00:08:52,010 --> 00:08:54,620 So to run this usage threshold? 132 00:08:54,650 --> 00:09:03,140 However, what you do, we just put plus US-EU in here and you set the threshold to zero in the beginning 133 00:09:03,140 --> 00:09:03,410 here. 134 00:09:04,100 --> 00:09:10,340 So it's a bit of a it's a bit finicky, I should say, not finicky, but a little bit unintuitive. 135 00:09:10,340 --> 00:09:16,370 And how to use our to threshold is probably going to have to look at this, this statement in some documentation 136 00:09:16,370 --> 00:09:20,090 when my code to get it, but also works quite well. 137 00:09:20,960 --> 00:09:23,270 So we'll take a look at the results below. 138 00:09:23,450 --> 00:09:24,170 So let's see. 139 00:09:25,130 --> 00:09:26,850 This was our adaptive thresholding. 140 00:09:27,710 --> 00:09:29,600 Sorry, this is this is the original image. 141 00:09:30,290 --> 00:09:32,480 This is our regular thresholding binary here. 142 00:09:33,380 --> 00:09:35,480 This is our adaptive mean thresholding. 143 00:09:38,520 --> 00:09:43,080 And you can see that she looks quite nice, it looks very good, actually. 144 00:09:43,620 --> 00:09:47,850 And Ozu, which I was saying was better, actually performed worse in this example. 145 00:09:48,660 --> 00:09:51,000 It probably depends on what you're doing, obviously. 146 00:09:51,450 --> 00:09:57,900 So maybe you just experiment with different with different methods if you want, depending on your use 147 00:09:57,900 --> 00:09:58,320 case. 148 00:09:59,250 --> 00:10:05,430 There's one thing I should know there was a bit of a mistake I made in the previous five minutes ago. 149 00:10:05,550 --> 00:10:08,760 Roughly, you can see the threshold binary how it looks here. 150 00:10:09,300 --> 00:10:10,950 That's the standard threshold binary. 151 00:10:10,980 --> 00:10:18,300 However, when asked to run threshold binary here, it actually looked like this big blob of mess. 152 00:10:18,900 --> 00:10:20,940 And it's not supposed to, actually. 153 00:10:21,360 --> 00:10:27,630 So we can take a look at the code and see and I already saw the mistake that was in that code and not 154 00:10:27,630 --> 00:10:31,770 sure why I left it in and why I didn't see it. 155 00:10:31,890 --> 00:10:34,920 But this line is supposed to be this here. 156 00:10:35,910 --> 00:10:40,710 So you can see if I can do this, you can see the differences here is that I've said the threshold to 157 00:10:40,710 --> 00:10:44,280 200 initially now rather than 127. 158 00:10:44,850 --> 00:10:50,740 So that's if you run this again, you would get the image we want below. 159 00:10:51,720 --> 00:10:52,180 All right. 160 00:10:52,320 --> 00:10:57,900 So that's it for finalization and adopted by finalization. 161 00:10:58,260 --> 00:11:01,250 We're now going to take a look at Esky images. 162 00:11:01,270 --> 00:11:07,170 Chushul Local Treasure Local is actually a very good algorithm I've used for actually in real life. 163 00:11:07,710 --> 00:11:13,620 For documents, getting it performs very, very well in terms of how it calculates the optimum threshold 164 00:11:13,620 --> 00:11:19,290 value, and we can see an example of it here, and you can see it looks quite good. 165 00:11:19,290 --> 00:11:22,590 It looks pretty much pretty similar to the adaptive mean thresholding. 166 00:11:23,220 --> 00:11:25,860 However, you can see it is better quality. 167 00:11:26,250 --> 00:11:34,490 There's some sort of noise around the words which make it a bit illegible, at least to me, but skeletons. 168 00:11:34,800 --> 00:11:35,340 Getty Images. 169 00:11:35,340 --> 00:11:35,940 I should say yes. 170 00:11:35,940 --> 00:11:41,070 I get images that show a local, which is done here, so let's take a look at the code. 171 00:11:41,550 --> 00:11:46,890 So you put Chushul local from ASCII image filters, so it's in the filter package. 172 00:11:47,370 --> 00:11:49,230 Probably a lot more the filters available there. 173 00:11:49,890 --> 00:11:54,060 And what we do, it actually converts image to HSV. 174 00:11:54,070 --> 00:11:59,850 We actually have to convert the image to HSV color space before implementing this, this type of threshold. 175 00:12:00,570 --> 00:12:05,340 And then what we do, we just take the HSV image here to call it V. 176 00:12:05,940 --> 00:12:09,810 We set this, we set these parameters here, which is block size and offset. 177 00:12:10,860 --> 00:12:16,710 These are probably configuration of parameters that you should probably play with so that you can actually 178 00:12:16,710 --> 00:12:19,050 get the most optimum threshold. 179 00:12:19,500 --> 00:12:22,560 However, just so you know, this is an adaptive threshold, and that's it. 180 00:12:22,890 --> 00:12:28,230 You don't actually have to set the threshold like you had previously admitted. 181 00:12:29,250 --> 00:12:36,090 So now we can run this and generate this nice image and this this is just a little note I put in here 182 00:12:36,090 --> 00:12:39,000 and white threshold by blurring before threshold is important. 183 00:12:39,570 --> 00:12:41,860 You can see this is a big reason for that. 184 00:12:41,880 --> 00:12:44,460 And actually, it's because it removes noise in the image. 185 00:12:44,850 --> 00:12:48,060 See this noise in the image here around the words that actually made it illegible. 186 00:12:48,480 --> 00:12:49,910 That's what removing. 187 00:12:49,920 --> 00:12:52,200 That's what we that's why we do some blur before. 188 00:12:53,010 --> 00:12:59,850 So that concludes, this lesson will go on to Chapter 10, which is dilation, erosion and edge detection. 189 00:13:00,120 --> 00:13:00,390 Thank you.