QTFD and DenseNet Based Motor Imagery Classification Method

‍ ‍Motor imagery （MI）-based brain-computer interfaces （BCIs） have gained immense popularity in a wide range of fields， including medical rehabilitation and entertainment. However， effectively processing non-stationary Electroencephalography （EEG） signals and extracting recognizable features from them remain major obstacles. In this study， our primary objective was to address this challenge by employing Quadratic Time-Frequency Distribution （QTFD） as an initial step to extract relevant EEG time-frequency features associated with MI tasks. By utilizing QTFD， we constructed high-resolution time-frequency representations of EEG fragments， enabling a more comprehensive analysis. Conventional linear time-frequency analysis methods often overlook nonlinear information and struggle to accurately depict the energy distribution of MI signals. In contrast， QTFD projects the signal from the time domain to the frequency domain using quadratic transformations， resulting in a more precise description of the signal’s energy distribution. Moreover， QTFD exhibits invariance to changes in time and frequency， offering relatively stable and accurate time-frequency characteristics. To further enhance feature extraction， we adopted the lightweight DenseNet network model. This model progressively extracts and integrates shallow and deep features from the time-frequency representations. DenseNet， with its small number of trainable parameters， is particularly well-suited for MI-BCI applications with limited available data. A notable advantage of DenseNet is its direct interconnection between each layer of the network， allowing feature graphs from all preceding layers to be accessed. Consequently， this facilitates the generation of a more distinguishable feature representation. Finally， we conducted experimental validation on the BCI Competition IV dataset， comparing the performance of our proposed classification model with various state-of-the-art algorithms. The results conclusively demonstrate that our method achieves superior classification performance， even when utilizing a reduced number of EEG channels. These findings underscore the efficacy of our approach in effectively processing non-stationary EEG signals and extracting informative features for MI-BCI applications.