Deep Learning-Driven Joint Antenna Selection for MIMO Systems: Optimizing Capacity and Gain via Improved ResNet

Abstract

To address the limitations of conventional deep learning models in antenna selection for Multiple Input Multiple Output (MIMO) systems—particularly in terms of insufficient feature extraction and degraded channel state information (CSI) modeling—this paper proposes a structurally enhanced ResNet-based joint transmit–receive antenna selection algorithm. We introduce a novel Shadow block that improves the residual network through two key innovations: 1) the use of $1\times 1$ convolutions to compress redundant features and reduce computational overhead; and 2) a Channel-Split Activation Fusion (CSAF) module, which performs multi-branch nonlinear transformations using diverse activation functions (e.g., ReLU, GELU, SELU, Sigmoid) on partitioned channel groups. This design significantly enriches feature representation and channel discrimination. The antenna selection task is formulated as a multi-class classification problem, and the enhanced ResNet is trained on a large-scale dataset of channel matrices. Experimental results demonstrate that our proposed model achieves classification accuracies of 70.30% and 69.10% for channel capacity and gain optimization tasks, respectively—approaching the performance of exhaustive search-based optimal selection while maintaining efficient inference. Compared with a range of classical and modern deep learning baselines, Shadow-ResNet consistently delivers superior accuracy, lower performance loss, and improved robustness. The source code is available at https://github.com/super123chen/AntennaSelection, and the dataset is provided at https://pan.baidu.com/s/1gM5nXzpyQpC2lGdMsaOYcg (Password: 3kqu).