CTRCNet: a CNN–transformer semantic communication system for image transmission with adaptive rate control

Semantic communication research commonly relies on convolutional neural networks (CNNs) as the backbone. However, the limited receptive fields of CNNs hinder their ability to capture global features effectively. To overcome this limitation, we propose CTRCNet, a parallel semantic communication framework that combines the strengths of CNNs and transformers. Specifically, we utilize CNNs to extract local details from image features while Transformers capture global dependencies within these features. These features are then integrated and output through a feature cross-fusion module, which effectively combines the features from each branch. To optimize channel bandwidth utilization, we designed an adaptive semantic feature mask module. This module dynamically adjusts the size of the output features based on image content and channel conditions, thereby optimizing the model’s transmission rate. Experimental results demonstrate that under additive white Gaussian noise channel conditions, when CR=1/4 and SNR=3 dB, CTRCNet outperforms ARC-DJSCC by 2.62 dB in the transmission of CIFAR100 data. For Kodak24 data transmission, at CR=1/8 and SNR=6 dB, CTRCNet achieves a 3.8 dB performance gain compared with the “BPG + LDPC” method. Moreover, CTRCNet adaptively adapts its transmission rate to varying channel conditions.