This dissertation addresses the challenges of EEG signal processing by proposing novel methods for signal preprocessing, feature extraction, classification, and hardware acceleration. First, a comprehensive review of traditional and state-of-the-art EEG techniques is presented, providing both theoretical foundations and practical guidance for future studies. We then analyze EEG signal characteristics and design MLP and SCNN models, demonstrating that beta waves offer more reliable indicators than alpha waves for eye state classification, with MLP achieving higher accuracy and efficiency while CNN ensures robustness through integrated preprocessing and feature extraction. Furthermore, EEG data collected via OpenBCI are used to introduce CWT-CNN and STFT-CNN networks, confirming that time–frequency analysis significantly improves classification accuracy, and highlighting the potential of branch learning strategies. Parallel CNN structures leveraging CWT, CSP, and STFT further enhance feature representation and surpass existing methods. To mitigate data scarcity, we propose a data enhancement technique specialized to EEG signal, achieving superior performance with our binary branch CWT-CNN. Finally, the study transitions from software development to hardware realization by implementing a lightweight CNN on the PYNQ-Z2 FPGA. Comparative evaluations with GPU and ARM platforms demonstrate that while GPUs achieve higher speed, FPGAs offer distinct advantages in energy efficiency, portability, and cost-effectiveness, making them highly suitable for real-time EEG-based applications in edge computing.