The Analysis of EEG Spectrogram Image for Brainwave Balancing Application Using ANN

Analysis of EEG Signal Processing Techniques based on Spectrograms

“…The aforementioned techniques are commonly used in both domains. The Time-Frequency domain allows extracting information in the two domains simultaneously; EEG analysis is based on the timefrequency image processing technique or spectrogram, a technique commonly used in Short Time Fourier Transform, which maps the signal into a two-dimensional function of frequency and time [2]. This section shows a review of the literature on extracting features of EEG signals using STFT.…”

“…Statistical measures can be extracted with this method, such as contrast, correlation, and energy, among others. Different authors use these features and others for classification; in an analysis done by Mustafa et al, [2] to classify mental stages through spectrograms, the authors extracted 80 statistical features for four orientations of the matrix, reducing the features vector by applying PCA, and used K nearest neighbors(KNN) to classify the stages. In [18], a comparison was made of two classifiers -Support Vector Machine and Artificial Neural Networks-following the same methodology, but the features vector had other statistical features, improving the accuracy for KNN (using Euclidian distance) in comparison with the Artificial Neural Network (ANN).…”

“…Regardless of the approaches for the data sets, the method used for the preprocessing stage is filtering, which is applied in few experiments. For the feature extraction, statistical values of the time-frequency representation are utilized; it was seen that feature selection methods were applied in only three experiments; some authors implemented this stage through PCA [2,3,27] [3], sum of distances [7], and random forest [26]. Finally, the classification was generally performed by KNN, SVM, and ANN; SVM was the best classifier in regard to accuracy.…”

“…For the purposes of the experiments in the literature, the resolution obtained with STFT is enough, implying less computational work and less execution time. However, it was shown in [2,4,5] that dimensionality is a problem to be solved; experiments where PCA was used do not report execution time and [22] do not apply reduction of characteristics, although they do note that it is necessary to reduce the vectors, as they have thousands of values. All works use different features, most of them based on texture; some works extract features from GLCM applied to spectrograms; although the features are similar, changing some of these affect accuracy.…”

“…The signal generated by this electrical activity is a complex random signal and is non-stationary [1]. Signals are normally composed of four brainwave groups: Delta, Theta, Alpha, and Beta [2]. Each band has different frequencies and amplitudes related to activities in the human body.…”