4 ms·
Years ago, fresh out of training, I got a job that involved image recognition, and I spent a couple of weeks reading the code dealing with image processing. I w
by krylon 4y ago
Years ago, fresh out of training, I got a job that involved image recognition, and I spent a couple of weeks reading the code dealing with image processing. I was confused, until I realized that all the operations were basically the same, the only difference was the matrix of weights applied to the pixels.
That same week, I had a chat with a friend who at the time was dabbling in audio processing, and we tried to understand how lowpass and hipass filters worked. When we did understand it, I realized they worked basically the same as the image processing operations, just with a temporal aspect to it.
- tkanarsky 4y agoyep, I'm currently taking a systems/signals course and it was incredibly illuminating how a lot of common image and sound filters -- edge detection, low and high pass filtering, equalization, etc -- boil down to the convolution of the signal with the impulse response of the filter.
- bob1029 4y agoI would focus on the differences between IIR and FIR filters if you really want to get your hands dirty. Pros and cons of each. That kind of thing.
- tkanarsky 4y agoThanks for the pointers! I'm very much a noob, this is really the first formal class I've taken on the topic. Are there any books, papers, or websites that stand out as particularly useful for grasping these topics?
- bob1029 4y agoI've found most elucidation come out of labs and practical work. I learned a lot about FIR filters by trying to do room audio correction on my own. That was where the latency tradeoff made itself apparent to me in a very obvious way. Reading it in a textbook is one thing, seeing/hearing live signals pass through your own filters is another. I think audio lends itself better to experimentation. You can actually construct passive filters and validate your assumptions with live equipment and not spend a dime. You likely already have all of the required gear in your office, short of some spare inductors, capacitors, etc. Image DSP work can be done entirely on a MacBook on an airplane, but I have personally found the added dimensionality to be less intuitive than with a time-domain signal.
- lostlogin 4y agoThis is a decent portion of MRI image processing, and the progress being made by software engineers at the moment is nothing short of spectacular.
- dls2016 4y ago> boil down to the convolution of the signal with the impulse response of the filter Not to be "that guy" but any linear, translation-invariant operator can be written that way. This is the essence of harmonic analysis.
- bob1029 4y agoOnce you learn that everything can be looked at in the frequency domain, things can get pretty spicy. DSP was easily the most impactful coursework I ever took at university. It applies literally everywhere. Information theory is at the heart of it all, and once you start to see how it all fits together, entire realms of possibilities open their doors to you. In our course we did a lot of image and audio work to help demonstrate that this concept applies to any domain. Discrete/continuous time domain, Some array of bytes on disk, doesn't really matter. You just need some rational way to quantize & sample your data. Also, as you get into control theory... more matrices! You can use state space representation (a matrix) to represent the entire valid control space of an N-dimensional system (albeit one that is LTI and where N < infinity).
- krylon 4y agoI remember the moment my brain went "So that's what some people need a fast multiply-accumulate for!"
- shuckles 4y agoDo you have any textbooks, course materials, or other references for these topics you'd recommend?
- LeonM 4y agoWhen I was an EE student (~2010), the book "Digital Signal Processing" by Steven W. Smith was considered to be the bible of DSP. You can read it for free online on his website here: https://www.dspguide.com/ https://www.dspguide.com/ or get a physical copy here https://www.amazon.com/Digital-Signal-Processing-Practical-Scientists/dp/075067444X https://www.amazon.com/Digital-Signal-Processing-Practical-S...
- bob1029 4y agoThe textbook that we used in the course was this one: https://www.amazon.com/Signal-Processing-First-James-McClellan/dp/0130909998 https://www.amazon.com/Signal-Processing-First-James-McClell... I don't think this is considered a fantastic book. The professor was simply amazing.
- alhirzel 4y agoAnother cool thread here is the Linear Canonical Transform[1], a generalization of the Fourier and other transforms. [1]: https://en.wikipedia.org/wiki/Linear_canonical_transformation https://en.wikipedia.org/wiki/Linear_canonical_transformatio...