Hand gesture recognition is one of the increasingly advancing fields, wherein important applications range from human-computer interaction and sign language translation to assistive technologies. The CNNs were established to be highly effective in image classification and feature extraction tasks thus recommending their usage for hand gesture recognition. We designed a robust hand gesture recognition system based on CNN, where we drew on the accuracy obtained by deep learning for the classification of hand gestures from the image data. We propose a novel architecture combining several convolution layers with pooling and dropout techniques for enhancement of generalization ability. The system has been experimented and tested using the publicly available hand gesture dataset that contains diverse hand shapes and orientations. Experimental results are presented to show that a new CNN-based model approach outperforms purely machine learning approaches for both classification accuracy and robustness against varying lighting conditions as well as hand positioning. Finally, the sensitivity of different architectures of the network to the variation of hyperparameters is analyzed. As for these aspects, the paper further discusses possible applications of the proposed system in virtual reality (VR), augmented reality (AR), and sign language interpretation, along with their future prospects in multimodal human-computer interaction.