Compression - Computerphile
HTML-код
- Опубликовано: 5 сен 2024
- Most of us deal with data compression on a daily basis, but what is it and how does it work? Professor David Brailsford introduces compression with regards to text and pictures.
/ computerphile
/ computer_phile
This video was filmed and edited by Sean Riley.
(some questions from Brady, in case you are confused!)
Computer Science at the University of Nottingham: bit.ly/nottscom...
Computerphile is a sister project to Brady Haran's Numberphile. See the full list of Brady's video projects at:periodicvideos.....
Without compression this video would be 1920pixel*1080pixel*24bit depth*25fps*(7*60+38)seconds = 569825280000bit = 71.23GB big.
I love any 45+ computer science professionals/lecturers, they have a very calm way of explaining ideas :)
yes, one video at a time! there's a lot to say about a lot of topics! this is just a RUclips channel!
>Brady
He sounds sooo passionate about this!!!
This is the best explanation I have ever watched regarding compression... Professor David Brailsford is awesome !!
I came up with the dumbest idea for my friend and I to develop, a lossy alternative to .zip and .rar. Just imagine all the affects you could get.
My lecturer asked our class to search about _data compression_ & _data encryption_ , and I stumbled upon this video. I was surprised to see that he's a professor at the University of Nottingham, which is also the same university that I'm attending right now.
I love the nonchalant and easy-to-understand way you explained encryption, professor! 👍🏼
Probably one of the best explanations about something technical I've ever heard. Bravo!
I hope Computerphile also discusses fractal compression (and fractal interpolation) because the concepts are so frikkin' cool.
BTW, the key signatures in standard music notation can also be thought of as a compression scheme.
Thanks for the great stuff, Brady!
I see a lot of people complaining about computerphile being too basic, too simple, too trivial. Well, for me, its just perfect. Im studying materials science so i allready know a whole lot of stuff discussed in periodic videos or sixty symbols. But i am completely ignorant about computers and these videos are a easy to digest sustenance for my brain :3
Send us your details and we'll give you your money back! ;)
Hope you stick with us and some of the videos to come will be more to your liking. We are never going to jump straight into the deep end... We're wading in gradually so our bits don't get too cold!
>Brady
Great explanation. It's basically like factoring out an equation. All about efficient use of space
well thats the way brady rolls if you check his other 10000 channels you will see that almost all his channels are for teaching people who dont know a lot about a topic or subject and making it more understandible and fun. And hes doing a damn good job at it
this interviewer gets to meet alot of awesome people .. and lots of awesome information from speaking with all these people
Wow, I really love Prof. Brailsford's style of explanation
As always Brady asks the best questions. As i expected this channel is getting better and better!
I find this video is very introductive to the topic, and gives a very brief overview of what compression is.(great for people like myself) It was made clear in the video that the next video will include more advanced explanations and discussions. On a side note, I have no idea why people marked your comment as spam. You gave criticism and an explanation yet people aren't happy with opposing views.
How kind and helpful of you. Good thing you're around, otherwise people might think it's okay to ask questions.
I can't help but think Brady must be one of the most interesting people ever to talk to on account of all the information and stories he hears during these videos.
This is to good to be free.
I'd say the hardest part of getting into computer science is the lack of simple and understandable explanations. It can be discouraging trying to learn about a topic and getting overwhelmed by advance information, which happens most of the time. So I think this channel is doing a great job introducing people so as long as they bring on the more advanced stuff later too.
I like Professor David Brailsford. He manage to describe one of the more complex computer subjects in an easy to understand way without oversimplifying it.
I was honoured to test out Fractal Image Compression to see if it was viable back at uni. It's what you think it is, you take a smaller bit of the pic, and then try to match that to bigger portions of the pic, and you would store the smaller block and the 'transform' needed. In the end, its not as good a current ones, it takes ages to scan every possible combination of blocks to find the best case, but you could get insane compression ratios on certain styles of images. Happy to see these videos
You've gotta go tip-to-tip for optimal efficiency - use the middle-out!
This is my favourite presenter so far. Don't get me wrong, I've liked everyone but he just seems so enthusiastic about what he says. :)
Hi, captions are quite a bit of extra work and we are super busy. I'd LOVE to have them... Numberphile has them done by an external organisation and I have nothing to do with it... Unfortunately computerphile does not have this same deal!
>Brady
Very clear and informative, and interesting. This channel is turning out just how I'd hoped it'd be.
pied piper does this well
Pennyinmouth middle out my friend, middle out...
In normal image files, stored data is list of pixels(BMP), or compressed list of pixels (JPEG, PNG).
In vector image files(SVG), stored data is a list of program intructions like this:
lineTo(x, y);
moveTo(x, y);
drawPolygon(x1, y1, x2, y2, x3, y3);
A vector graphic renderer program will do these intructions and render your vector graphic.
i like how he uses the word "predictability" instead of patterns (all the time).
if he's a teacher i'd like to attend his classes, he's kinda tranquilizing and explains well.
I appreciate these videos immensely. It does a pretty good job of explaining these concepts that I deal with every day. But at the same time, I have trouble explaining it to people who aren't as computer literate. I'll send them a link to an appropriate computerphile vid and generally they have been coming back with oh now I understand what you mean. Which helps me out professionally and is a credit to you Brady (and team).
Brings me back to my engineering classes where we had to compress sound (like mp3's), would have loved if he had gotten into the details on that topic... All I can remember was that you had to do a Fast Fourier Transform (FFT)... but that's all I can recall :)
I had no idea computerphile was this old. Even crazier how it was just as good then as it is now. Most channels back then still felt more "public access". Then again, there were some jewels on public access TV in the 90's.
That business document is hilarious. I especially liked the part about the meaningful and relevant scenarios at this point in time going forward.
JESUS, CAN'T WAIT FOR THE UPCOMING CONTINUATIONS!
One thing i would like to see explored here or in another video would be compression of several files vs data deduplication of several files. Data deduplication is a completly different technique but it is also quite fascinating i think.
I thought this was a great way to present the concept of compression. I'm a little surprised at the comments from people who complained there wasn't more in-depth information, even though they already knew it! Like many of your videos, this was (I think) for people that don't know much about compression, but are interested in knowing more.
the interpolated frames (P-Frames) are usually deltas from smeared around versions of the key-frames (I-Frames). (ignoring B-Frames for now...).
IOW: encode how the blocks in the I-Frame move around (via "motion vectors"), and subtract this from the current frame;
encode this subtracted image as a residual (still sort of like JPEG, just only encoding a residual image).
audio compression is more its own thing (often the video and audio encoding are handled separately, but share the container)
Well, I'm starting a computer science degree next year and I understood perfectly what he said. In fact, this is a topic of wich I have thought a lot and I really enjoyed the level of detail they gave. Will see about the entropy thing but the level of this video was perfect for me.
This just became one of my favorite channels.
Ow darn. Just when we get the the actual info you leave us with this cliffhanger.
Well played Mr. Brady. Well played.
Here's one method (called LZW): I'm going to give you a sequence of codes. These codes can be symbols from the message, or they can be references to the compressed file. The code "1" means to look back at the first and second items in the compressed message, use them, and keep expanding. "2" means the second and third, etc. E.g.:
this is the theme -> this 3 1e 8me [3 means "is", 1="th", 8="1e"="the"].
If there's a lot of repetition, the dictionary gets a lot of codes that expand to big chunks.
I really love Computerphile and Numberphile videos. Waiting for more, thanks!
Just suggestions for future videos :)
DCT in particular (or more general frequency based algorithms) is quite difficult to understand and would need its own video, but it is the basis for almost all lossy compression used in the real world right now (mp3, jpeg, mpeg, etc).
Some of them are quite easy to explain, such as RLE, but are still used practically.
Compression is such a beautiful subject. Keep up the good work.
I'm glad to see this channel taking off. Great content, Brady.
The channel is definitely getting closer to what I was expecting and it's nice to see after being so terrified of the first video! Very much looking forward to see a little more advanced topics. :)
Physics and Math are very interesting to watch, but as a web developer this channel is a bit more engaging.)
and as always - amazing work, Brady!
Morse code is a compression, because it uses shorter sequences for common letters like e, i and longer sequences for rare letters. This technique is the basis of all modern compression algorithms and it is easy to understand for ordinary people.
Doing computer science at university and loving the videos, Brady! You have a subscriber for life for this channel and all other subjects you wish to enlighten! Cheers :)
there's a fine line between explaining the core principle/idea behind something, and a specific technique derived from that core idea. explaining the core idea only is usually more certain to be interesting, explaining specific technique tends to be a) more complicated; b) less "generaly usable", and therefore less interesting, in many cases. finding that line is hard, and it's better/safer to keep on that more general side, at least on the beginning.
typically, there is a "sliding window", where if something is seen which has been seen previously, we encode a reference to it (in terms of how far back it was and how long it was). on the encoder end, it is typically a "clever" version of a brute-force search (over the preceding 32kB or so...).
usually, it is based on if at the current encoding position there is a matching string in the window, and encoding whichever match is longest, otherwise emit a character and try again, ...
It would be interesting to have two vids from the computerphile and numberphile channels, respectively, dealing with information theory. One could be on the math of it, and then another on how it is used in computing and in things like compression and encryption, etc. There are tons of opportunities for the channels to complement each other with uploads that are timed like that.
No, dithering is where you add dots/pixels of a color to make a colour change smoother. For example if you had a red and white beside each other, to blend them you would add red dots to the white edge, and white dots to the red edge, and from a distance it looks the white morphs to the red. With compression, you start with 'dithered' images, and throw away the dots to save space. Of course you would only do this if you wouldnt miss it, make it 'fuzzy' like the video says.
Really great 1st video for numerous follow-ups...!
@SEThatered: JPEG is not a lossless compression and MP3 removes sounds within human hearing range, as the uncompressed source audio already has very little sound outside said range. Most lossy compressions tries to remove information you're not going to miss, but take it too far, and you'll notice.
This is great. I understand how the scene group Black-Box go around compressing their releases.
meaningful and relevant scenarios at this point in time going forward
I remember our physics teacher explaining this just like him, love it!
if the people targets casual audience that is at least a little interested in these things, but thought they're not, because of bad teachers (which is the audience that ALL of brady's channels target, imo), the content is exactly as awesome as it's meant to be.
and: I'm a programmer and I love watching this stuff, because I'm always curious how will they explain these things in the language of mortals :-D
(and sometimes I learn a bit of something new too!)
Wow, I enjoyed every minute of him talking... He's a really smart guy can't wait to see the next video.
sort of...
typically, we feed the audio through the MDCT, which converts from time-based samples to a number of frequency bands (for a block of audio samples). we can then divide these numbers by other numbers (based on quality settings and the "psycho-acoustic model" and similar), then the output numbers can be encoded.
often left/right is converted into center/side as well (this compresses better), in what is known as "joint stereo".
(center is like mono / average, side is left - center).
straight to the point this is 100% clean video I've ever seen keep it up from 2020 : )
That is very true. Pretty much any image that is defined as paths and simple colors or gradients can be described as a SVG file. In that sense, which format is best depends on the content of the image, as well as the size of the image.
Direct your fury at Sean - he was the editor! I was just asking questions!
(FYI folks, we always say who edited the film, etc, in the video description)
>Brady
Yeah, that was exactly what he was talking about when he said "dividing the picture up into blocks and squares and saying 'well, this is almost the same green; I'll use this as an approximation'" (except it's black instead of green in your example)
note that despite their fancy-looking mathematical definitions, as-implemented, the DCT transform (and its inverse), typically amount mostly to matrix multiplies with fixed-point numbers.
IOW: typically no actual "cos" is used, as this would make things too slow. matrix multiplies are reasonably faster. likewise for integers vs floats.
luma is essentially a weighted average of red/green/blue (IOW: brightness).
these are both so we can abuse the numbers more before people will really notice.
On the first thing the prof said, I might be nit picking here but as I understood it (non-lossy) compression is about reducing the amount of data, whilst keeping the same amount of information, i.e. reducing redundancy rather than reducing the information (which would be lossy compression). :)
A computer uses 8 bits per character, so the message "morse code" takes up 80 bits. In morse code that message has 24 bits (not sure if I counted correctly), because the e takes up only 1 bit and o takes 3 bit. You always have to count in bits. If you search for Huffman coding, you will find, that it's basically the same thing and used in ZIP and JPEG.
sound compression uses sampling lengths, if you've ever seen a thing on an mp3 file when you right click on it it will say something like 192bit or 320bit. That's the number of times in one second that it samples the original sound file. files like .wav and .flac are lossless files, but take up about 10x more space but can sound much better if the sound file is very busy for instance an intricate orchestral piece.
Morse code is definitely a compression technique. It is one of the first techniques ever used to transmit text.
I agree morse code will tend to increase the amount of data present, but that is only a general case. Morse code can have a compression effect, especially if the person transmitting is adept at doing so.
Well JPEG is indistinguishable from its own compression format. The h.264 is just a part of the MPEG4 documentation and refers specifically to video compression. (Part 10, while the file type as defined .MP4 is part 14 I think.) MP4 is a container format that holds a lot more than just video, so its a bit missleading.
What it comes down to is that the way we handle video and it's standards are different from how we do images. The most you can say I guess is that these are all lossy compression.
Sean let me come along for this one! ;)
>Brady
The smaller file has a lower bitrate. It is more compressed and it can be noticable if you compress it too much. Blu-ray movies are actually generally 25-40gb of video stream data. Compressing that to 1gb would be noticable, particularly if you project it onto a very large screen. Just watch high contrast edges in motion to see it more easily. It's best to keep the source as high as possible and let people compress to their liking, as you generally can't go the other way with lossy.
1:38 "Whilst it is highly agreeable to make things like this up on the spot, one realises as one is doing it how much work it will be to animate"
I was just thinking how nice that graphic looks, especially with the extra bit of effort to put in some interesting text!
This is exactly the same baseline as Brady's other channels have been laying down for other fields, the computer savvy crowd is angry for nothing. It's all been elementary for me too so far but I'm still enjoying these just as I enjoy (e.g.) Numberphile videos about elementary topics I already know about. If anything, this video perhaps could have benefited from more visualizations.
yep. using a dictionary and Huffman coding works pretty well...
slight compression improvements are possible by using a bigger window / dictionary and arithmetic or range coding (like 7zip / LZMA), but the gains are fairly modest relative to the slowdown.
No, the number repressentation of the file needs to be evenly divisible by that constant so you would have to calculate the divisor for each file and store it with it. I cannot think if any example where that approach would not end up taking mroe space than it would have had otherwise. Ie, 10 and 100 are 5 characters while 1000 is 4. In binary, the difference gets more extreme.
WebP and VP9 are two new compression algorithms from Google which sound amazing. VP9 is smaller than JPEG at the same quality, supports animations, and it can be lossless. So basically, it can replace JPEG, PNG and GIF.
VP9 is a video codec which offers H.264 level quality, at half the size.
If anyone is interested about new compression algorithms, check those two out because they are pretty damn amazing.
The first really interesting video for me !
Looking forward to the next compression video.
excited about the next vid, i got really dissapointed when he got cut off.
That depends on the compression scheme, according to the Wikipedia article about the format. However, the most common schemes seem to be lossless.
7:18
"Let's get you a piece of paper . . ."
Don't let him stop, Brady! If he goes below 60 MPH, he'll explode!
If all the numbers you were trying to compress were divisible by that constant (or if most were and you had a method of recording and restoring the lost data for the exceptions), then yes.
Computer memory uses this method actually. Instead of having an address for every bit or byte, it has an address for every x bytes. So memory address 5 actually will access, for example, bytes 20 through 23 as a group (for four-byte groupings, or 'words').
answer:
entropy = h = sum from 1 to n of -p(x)*log(2,p(x)) where p(x) is the probability of a given symbol occurring in the file (log is base 2)
I can listen to him all day!
Code would imply choosing a particular programming language. I think it's more interesting to talk about the underlying algorithms and techniques, and not the implementation details (probably what Brady thought).
Though truth to be told, sometimes bits pseudo-code help understand stuff a lot. And indeed, it's not bad to have some bits of code every now and then on a computer-related video :)
@SChil
True, for a 'smaller' file sizes such as the cd you mentioned, the time saved by compression then transmission would be negligable.
But then imagine the same scenario where, for whatever reason, you are wanting to send a file of several hundred gigabytes; then the time saved may be the difference between it taking 12 hours or 2 (if the file is easily compressed).
Then again, as you said, speeds are increasing so much that eventually even files of this size might become negligable!
Would be nice if you guys subtitle the videos (like in numberphile), i'm not fluent on english, and subtitles help me alot. But, nice job, keep doing it :)
I'm really starting to like this channel. Your videos have one big flaw though: they're much too short. And I don't only mean this as a compliment :) No really, there is so much more that could be said about this topic, you could fill hours and it wouldn't stop being interesting.
PNG compression is really interesting. I recommend checking it out in Wikipedia if it isn't featured in the upcoming video.
Vector graphics are a set of math plots for shapes. It is completely different from raster graphics which are sets of information about each pixel or area. This is why you couldn't have an svg of a photo for instance (well you could theoretically) and also why vector graphics don't lose anything when zooming way in.
it is more like:
converts RGB (red, green, blue) to YCbCr (luma, relative blue and red);
feeds pixel blocks through the DCT transform (turns it into frequency bands);
divides the DCT output by numbers (quantization factors) to make the values smaller;
encodes these numbers in the file (using Huffman coding and variable-width numbers).
say, 819/63 -> 13, but also 849/63 -> 13, ...
dividing by bigger numbers makes a smaller output image.
decoding reverses the above process.
(more or less...)
I did some work in the late 90s with compression algorithms. Huffman coding is interesting to play with and easy to understand.
but, yes, it is the loss due to those integer divisions which is the cause of both the blurring and also the various artifacts. this is the main thing that both adds loss, and also what makes the file smaller.
otherwise, for lossless encoding, images tend to be around PNG sized.
other lossy codecs mostly introduce other sorts of artifacts, depending on the specifics of how their math is done. DCT and division remain popular though due to being reasonably fast and compressing fairly well.
Me and my friends loved your channel. Good job, buddy.
Good point. I was being somewhat generic due to the lack of character space, and the vagueness of the term "publication" (which I typically associate with "logo making", and logos tend to be drawings).
I usually go lossless for images, but once it gets to things like music and video I tend to go for high quality lossy. As long as that JPEG is set to 85% or above in Paint.net or that mp3 has a bitrate of at least 192 kbps, I probably won't notice. It has it's huge advantages as well, as lossless compression can make videos that usually follow a ratio around 1 GB per hour. Even if you are at "100%" lossy, you still get a much more efficient file size while it is still indistinguishable from the lossless.
thanks!
I know this sounds crazy, but it is possible to get substantially higher compression rates than achieved today by predicting the future. I came up with the principle and demo over 30 years ago.
I HATE YOU BRADY! Why did you have to stop!!!
This is by far my favorite video on computerphile! Now I know something I've always been wondering about :D