Interpolation filter based on a neural network.

This work presents an alternative system of interpolation of several signals based on artificial neural networks using MATLAB. Different architectures were tested starting with a neural network with multilayer perceptron topology, seeking to obtain twice as many samples of the original signal, whose training algorithm was performed from a database built by acquiring ECG signals captured in a laboratory by the authors. When testing the performance of the network under more demanding circumstances, such as the generation of more samples and responses to a greater variety of signals, it was evidenced that the performance decreased significantly, so the same architecture was retrained using a new public database of larger size. Looking for alternatives, a model is made using the radial based network topology, then the results obtained in the two topologies are compared using as input signals a composite Sin, SinC, a component of the Lorenz attractor, a high order polynomial and different ECG signals. The results obtained show better results in the radial basis network topology compared to the multilayer perceptron topology, however both topologies generate interpolations quite close to the original signal. [ABSTRACT FROM AUTHOR]