Brain-computer interfaces (BCI) are systems that connect human minds and machines using brainwaves as control signals. Those systems use acquired brainwaves through ElectroEncephaloGraphy (EEG), which is more convenient for this scenario. Unfortunately, the produced signals are noisy and non-stationary; this decreases the quality of the signals and negatively impacts the performances of the BCI systems. Also, the inconvenience of most feature extraction algorithms is that they cannot extract all the characteristics of the signals. Therefore, we propose a novel convolutional neural network (ConvNet) that extracts data from different resolutions of the input signal. It will allow us to extract more relevant features that will increase the overall performance of our method. Also, it permits to increase the depth of the network without risking any gradient vanishing problem. The method was benchmarked with the dataset BCI competition IV-2a. The results prove that our method performs better than Filter Bank Common Spatial Pattern and Riemannian geometry techniques.