This paper proposes a deep learning framework based on convolutional neural networks aims at extracting and processing spatialtemporal features for an efﬁcient modulation recognition. In this architecture we integrate the strength of grouped and dilated convolutional layers to achieve the efﬁcient recognition in terms of recognition accuracy and less complexity. To allow multilevel feature learning and model generalization we deployed skip connections. Furthermore, to verify the performance of our architecture we performed experimental analysis on RadioML 2018.01A open-source datasets. According to the results, our model outperforms ResNet based model with regards to recognition accuracy and parameter utilization accuracy.