Abstract
Steady-State Visual Evoked Potentials (SSVEP) have garnered significant
attention due to their promising applications in brain-computer
interfaces (BCI), medical diagnostics, and several other domains.
Enhancing the characteristics of SSVEP signals through intricate signal
processing has emerged as a pivotal research focus for more efficient
signal extraction. In this work, we introduce a novel layered
enhancement algorithm for SSVEP electroencephalogram (SSVEP-EEG) signals
based on fractional-order differentiation operators. This innovative
approach marries brain signal analysis with image processing
methodologies. By utilizing fractional-order differentiation operators
in tandem with the Laplace pyramid, the signal undergoes hierarchical
enhancement. This amplified signal is then reconstructed, which
facilitates an in-depth extraction of image intricacies and attributes,
ultimately accentuating the distinctiveness of SSVEP features. To
validate the efficacy of the proposed method, we applied it to three
recognized target identification algorithms: Canonical Correlation
Analysis (CCA), Filter Bank Canonical Correlation Analysis (FBCCA), and
Task-Related Component Analysis (TRCA) using a publicly available
dataset. Experimental outcomes underscore that, in contrast to
contemporary techniques, our proposed algorithm not only effectively
attenuates the trend components of SSVEP signals but also substantially
elevates the recognition precision of CCA, FBCCA, and TRCA.
| Introduction
Over the past three decades, Brain-Computer Interface (BCI) technology
has garnered significant scholarly attention. This advanced framework
facilitates direct interaction between the human brain’s activity and
computing systems, offering a novel avenue for encoding and decoding
information predicated solely on neural activity [1]. Among various
BCI modalities, the Steady-State Visual Evoked Potential (SSVEP) has
emerged as a particularly salient methodology, largely because
SSVEP-based BCI systems obviate the need for extensive user training
[2] and boast high information transfer rates [3]. As such, it’s
increasingly recognized as a promising modality for interactive
applications, often underpinning the development of sophisticated
systems [4]. One of the paramount applications of SSVEP is within
the realm of BCI, where the precise identification of a user’s intent
stands as a pivotal research direction [5, 6, 7, 8]. In SSVEP-based
BCIs, an array of flashing modules, each oscillating at distinct
frequencies, is used as stimuli, where each stimulus corresponds to a
specific operational command [9]. When a user fixates on a
particular stimulus, cortical neural activities get modulated, resulting
in the generation of periodic rhythms that resonate at the same
frequency as the stimulus. This activity is predominantly localized in
the occipital region of the cortex [10]. By pinpointing the peak
frequencies within the induced electroencephalographic signals, one can
discern the specific stimulus the user is focusing on, thereby inferring
the user’s intent [11]. Several algorithms have been developed to
enhance target recognition within this context, including Canonical
Correlation Analysis (CCA) [12], Filter Bank Canonical Correlation
Analysis (FBCCA) [13], Multivariate Synchronization Index (MSI)
[14], Individual Template-based Canonical Correlation Analysis
(IT-CCA) [15], Multiway Canonical Correlation Analysis (Mway CCA)
[16], and Task-Related Component Analysis (TRCA) [17], among
others.
Nevertheless, the inherent low signal-to-noise ratio of SSVEP signals
poses a formidable challenge, constraining its research, analysis, and
application. To enhance the extraction efficiency of SSVEP signals, it
becomes imperative to augment their characteristics. Yan et al.[18]
introduced an Image Filtering Denoising (IFD) method for SSVEP,
initially filtering multi-channel EEG signals to retrieve noise, which
is then subtracted from the original signal to procure the denoised
version. This approach essentially delineates blurred details from an
image and then excludes them to obtain finer details. However, this
method tends to inadvertently discard some essential information in the
process.
To overcome this impediment, we propose an Enhanced Feature Augmentation
Algorithm based on Fractional-Order Derivative Operator, termed Enhanced
Riemann-Liouville based on Laplacian Pyramid (ERLLP), to amplify SSVEP
Electroencephalogram (SSVEP-EEG) signals. The strategy first employs a
sharpening filter utilizing spatial differencing to extract image
intricacies, followed by an application of the fractional-order
derivative operator for optimal detail and contour refinement.
Contrasting the integer-order differential operators, like Prewitt,
Sobel, and Laplace, used in [5], fractional-order derivatives, as an
extension, not only offer image sharpening but, due to their weak
derivative properties, can also bolster the high-frequency components
while preserving the low-frequency portions in a nonlinear fashion. This
makes fractional-order derivatives more potent in analyzing and
enhancing image characteristics. Hence, we integrated the
fractional-order derivative operator and further refined the traditional
Riemann-Liouville fractional derivative operator, leading to the
inception of the Enhanced-RL operator. Subsequently, employing the
Laplacian pyramid facilitates layered signal processing, followed by
signal reconstruction to further delineate details and features.
Ultimately, using a public dataset, the study corroborates the augmented
efficacy of our ERLLP algorithm in enhancing the recognition
capabilities of user intent identification algorithms like CCA, FBCCA,
and TRCA. Experimental outcomes underscore that, in contrast to the IFD
and the methodologies detailed in [5], our proposed ERLLP algorithm
markedly escalates the accuracy of SSVEP signal recognition.
| Theoretical Framework
| Enhanced-RL Operator
Given the foundational work on the Riemann-Liouville (RL) differential,
Ni et al. [19] advanced the field by introducing a specific image
enhancement operator. Recognizing the computational challenges
associated with higher-order complexities, they strategically executed
differentiation operations, bifurcating them along both the positive and
negative directions on theandaxes. Additionally, differentiation was
executed across the four cardinal diagonal vectors. To facilitate this
refined operation, a 3×3 template was judiciously deployed.
Subsequently, the products of these eight distinctive fractional
differential templates, derived from varied orientations, were
amalgamated. The culmination of this intricate process resulted in a 5×5
convolution template, a depiction of which is presented in Figure 1: