DEVELOPMENT OF OPTIMAL ALGORITHMS FOR EMOTION RECOGNITION

Abstract

Emotionrecognitionfromelectroencephalography(EEG)isvitaltoadvancinghuman–computer interaction, affective computing, and mental health assessment, yet its effectiveness is im peded by high dimensionality, noise contamination, and the nonlinear dynamics of neural signals. This thesis addresses these challenges through integrated research objectives: op timal EEG channel selection, optimal filter design for noise elimination, an optimal mode filtering framework, optimal wavelet design using nature-inspired algorithms, and optimally tuned deep learning models for EEG-based emotion recognition. The first objective is to develop a correlation-based channel selection method that identifies a compact set of electrodes most relevant to emotion recognition. This approach reduces redundancy while preserving discriminative neural activity. The selected channels undergo processing using volt–Kalman sub-band filtering, from which statistical features are extracted in the sub-bands. These features are then refined using the MRMR (Minimum Redundancy Maximum Relevance) algorithm, and the resulting features are input into a classification algorithm, which yields an accuracy of 88%. Building upon this framework, the second objective focuses on designing a COA-optimized IIR (Infinite Impulse Response) filter to achieve accurate rhythm extraction, thereby isolating clean 𝛿, 𝜃, 𝛼, 𝛽, and 𝛾 rhythms for subsequent feature analysis. From these refined rhythms, Hjorth parameters and entropy-based features are computed, which, then classified using a cubic and quadratic SVM (Support Vector Machine), deliver robust performance with accuracies of 94.4% and 93.0%, respectively. The second objective complements the framework by introducing a PSO-based adaptive IIR filtering approach for the extraction of clean 𝛿, 𝜃, 𝛼, 𝛽, and 𝛾 rhythms from EEG signals. The filter parameters are adaptively selected using Particle Swarm Optimization (PSO), where the optimization criterion is defined as the minimization of the mean squared error (MSE) between the original and reconstructed rhythms. Subsequently, the extracted rhythms are converted into time–frequency images using adaptive superlets (ASLT). Building upon these representations, a hybrid CNN (Convolutional Neural Network) and ResNet architecture (HCRNet)is proposed. Whentrained on ASLT-derived time–frequency images, the proposed HCRNet achieves a classification accuracy of 93%, thereby validating its effectiveness in advancing temporal-spectral modeling for emotion recognition. To further improve EEG quality, the thesis proposes an optimal mode filtering framework comprising two complementary strategies. In the first, Variational Mode Decomposition (VMD) is used to separate EEG signals into intrinsic modes, with Pearson’s correlation co efficient identifying modes that preserve meaningful neural dynamics. Further, the clean EEG is decomposed into the canonical 𝛿, 𝜃, 𝛼, 𝛽, and 𝛾 rhythms using the Wavelet Packet Transform (WPT),therebyensuring precise rhythm isolation. Joint time–frequency scattering i applied to the filtered rhythms produces robust feature representations that yield high clus tering separability, revealing the dominance of low-frequency rhythms in negative-emotion states and high-frequency rhythms in high-arousal emotions. JTFS-derived features are then evaluated across multiple classification algorithms, where 𝛿 and 𝜃 rhythms consistently out perform higher-frequency bands. Notably, the combination of 𝛿 and 𝜃 rhythms achieves the highest classification accuracy of 80.39% on Dataset 1 and 99.35% on Dataset 2, thereby showing the effectiveness of the proposed framework in emotion recognition. The second strategy employs Group Sparse Mode Decomposition (GSMD) with Bhattacharyya Distance to suppress modes that lack discriminative power. The refined signals are transformed using Superlet and Adaptive Superlet Transforms, and a dedicated super-resolution neural network (SRNET)designed for fine-grained time–frequency patterns achieves 99.63% accuracy while outperforming standard deep architectures. Complementing these contributions, the thesis develops nature-inspired optimized wavelet frameworks to enhance time–frequency feature extraction. An adaptive flexible analytic wavelet transform (AFAWT) optimized using Particle Swarm Optimization automatically tunes wavelet parameters to EEG characteristics, achieving 90.3% accuracy on video game emotion datasets. Similarly, a crayfish-optimized tunable-Q Wavelet Transform (TQWT) adapts decomposition resolution for emotion-specific structure, yielding state-of-the-art per formance on SEED-IV (91.47%) and DEAP (77.81%). A final extension introduces a PSO optimized Wavelet Scattering Transform, improving scattering feature quality and achieving accuracies of 97.6% and 99.4% across two benchmark datasets. The thesis culminates in the fourth objective, which focuses on developing optimally tuned deep learning models for emotion recognition. EEG signals are first converted into time–frequency images (spectrogram, scalogram, and SPWVD), which are then processed by an optimized CNN/GoogleNet architecture. Given the sensitivity of deep learning per formance to hyperparameter settings, this work incorporates Bayesian Optimization (BO) to automatically identify the optimal hyperparameters. BO efficiently explores the parameter space, achieving faster convergence and significantly higher accuracy compared to conven tional manually tuned models, thereby validating the robustness of the proposed framework. The optimized GoogleNet-derived features are classified using an SVM, achieving robust performance with accuracies exceeding 82% and AUC values above 0.95. Overall, the optimized filtering, decomposition, wavelet design, and BO-enhanced deep learning strategies collectively yield robust and high-accuracy emotion recognition. The pro posed framework demonstrates strong potential for deployment in brain–computer interfaces, affective computing, adaptive gaming environments, and mental health monitoring systems.