Electroencephalography (EEG) is widely regarded as an effective non-invasive technique for measuring neuronal activity in the human brain, offering high temporal resolution. However, EEG signals are often contaminated by noise and artifacts, which significantly impact their analysis. To tackle this challenge, various advancements in deep learning-based denoising techniques have been developed. Despite these advancements, identifying the optimal network architecture across a broad range of initial parameters remain complex. Manually tuning these parameters to achieve optimal performance is time-intensive and demands substantial expertise. Along the same lines, we introduce a novel residual block-based neural network for automatic artifact removal and a convolution neural network for classification. The architecture unfolds in two distinct stages: Firstly, the noise elimination, where the raw EEG signals undergo the Hilbert transformation, converting them into complex-valued signals. These signals serve as the input for a Residual block-based Convolution Neural Network constructed using the innovative technique of Differential Evolution (HT-DEResNet), which optimizes both the architecture and initial parameters. This enables the effective removal of artifacts without manual intervention. The model is applied across three complex datasets: HaLT, eye artifact, and major depressive disorder. The secondary stage comprises the development of a Multi-Convolution Neural Network (MCNN) for classification. To validate the effectiveness, the BCI competition datasets 2a & 2b are subjected to the HT-DEResNet framework. Subsequently, the signals are classified. The results not only surpass state-of-the-art methods in accuracy but also achieve a 17.73% improvement in performance attributed to noise removal.