SIFT (Scale Invariant Feature Transform) is a very popular local feature descriptor, and is widely used in computer vision applications. Likewise Harris corner detector is also useful for feature extraction from images. I am looking for similar feature descriptors that can be used to match two 1D signals like speech, ECG etc. I am analyzing electromagnetic signals, and have noticed that some parts of the signal contain "interesting" features that is good for matching or pattern recognition (even visually), while some other parts don't contain so much discernible patterns. Is there any algorithm that can quantify or identify local features in 1D signal?
https://pdfs.semanticscholar.org/76fb/e5d76602a0658dfb32bd9ba561c3c24c6a7d.pdf
Its all about feature extraction out of a short window of a 1D signal and then matching that feature with a sliding window. However SIFT does not use exhaustive search like using sliding window but use DoG operator to detect some repeatable feature points. The approach is called as correspondence finding. Check the following paper. It will help you to find the other relevants.
"Solving correspondence problem with 1D signal matching" https://pdfs.semanticscholar.org/efa2/1037a0275c9ee9effb617202449d2829a820.pdf&ved=2ahUKEwifnJmBvN7aAhXFro8KHT96C4kQFjABegQIBxAB&usg=AOvVaw32_FxtuUkyWlVEgkGYWhGi
Lowe's SIFT Detector is based on ideas that were originally conceived in the context of 1D signals (time series) first presented by Witkin in 1983. It is called Scale Space Analysis, and it is based on detecting "interesting" features that are invariant to scale distortions. You can explore Tony Lindeberg extensive work on that. As far as I am aware of, the application of Scale Space Analysis directly to time-series is not widely extended. These techniques share a lot with Multiresolution Wavelets and I believe that is the reason why they have not been explored in applications so much (being Wavelets a very successful silver-bullet). On the other hand, if you are interested in the SIFT Descriptor, we have working a lot on that here: Conference Paper BCI classification based on signal plots and SIFT descriptors
. The idea is that you can use the SIFT Detector to classify those interesting features.- Badges
- Science topic