There's a bit of a mischaracterization of the way 2 dimensional Fourier transforms operate on images in this post. The 2D DFT (or DCT rather) doesn't deal with anything as complex as tracing shapes on a 2D plane like seen in the video. What it does is treat each 1 dimensional line of the image as a signal/waveform, with the pixel intensity as amplitude. Fourier-family transforms are separable, so the 2D (and ND) case is equivalent to the transform of each line followed by a transform along the resulting columns. So the actual representation that arises from this looks like so, with each image representing the corresponding cosine component: http://0x09.net/img/dct32.png (here grey is 0)Visually most of the sinusoidal components here are zero or nearly so. However if we scale them logarithmically, we'll see that it's actually not so: http://0x09.net/img/dct32log.pngWhat transform coders like JPEG do is reduce the precision of these components, causing many of them to become zero. Which is good for the entropy coder, and mostly imperceptible to us. Of course JPEG operates on 8x8 blocks only * rather than a whole image like here.It's hard to imagine this as an image, so here's a progressive sum starting from the second term, which essentially demonstrates an inverse DCT: http://0x09.net/img/idct32.pngmind that 0 is adjusted to grey in this rendering, and the brightness of the result is not an artifact of the transform.It's easier to understand what goes on with these transforms if you can visualize things in terms of the basis functions. Which in the case of a 32x32 image like above would be http://0x09.net/img/basis.png (warning: eye strain).All the examples above pertain to the DCT, partly because of JPEG and partly so I could avoid getting phase involved, but the principles apply equally to the other transforms in the family.* although recent versions of libjpeg can use other sizes

Applications are open for YC Summer 2019

Search: