Brain–computer interfaces (BCIs) have significant applications in neuroscience and clinical rehabilitation. However, due to the inherently low signal-to-noise ratio (SNR) of electroencephalogram (EEG) signals, they are highly susceptible to motion artifacts such as electrooculogram (EOG) and electromyogram (EMG), which severely compromises their reliability in real-world scenarios. Most existing denoising methods process the entire EEG recording directly, ignoring the coexistence of clean and contaminated segments within raw data. This often results in the unintended removal of valid neural information and degrades downstream decoding performance. Although some recent approaches attempt to segment clean and noisy regions, they still struggle with handling transition boundaries effectively. This paper proposes a Multi-Branch Generative Adversarial Denoising Network (MBGADNet) based on the WGAN-GP framework. The model comprises three core modules: (1) an improved multi-branch Inception encoder that extracts multi-scale frequency features to enhance the discrimination between EEG and artifact components; (2) a generator equipped with multi-head self-attention (MHSA) to model long-range dependencies within clean EEG segments and reconstruct artifact-free signals; and (3) a dual-branch discriminator that performs both adversarial classification and artifact region segmentation, guiding the generator to achieve artifact suppression while preserving semantic integrity. Experiments were conducted on a semi-synthetic EEG dataset with varying SNR levels. Results show that MBGADNet effectively retains neural components even under low-SNR conditions and achieves superior performance over state-of-the-art methods in terms of RRMSE (0.132 vs 0.146) and SNR (11.559 dB vs 8.327 dB), demonstrating its robustness and practical potential.