Semantic-Constrained GAN-Based Joint Description-Feature Transmission Method for Image Semantic Communication

Traditional image coding standards struggle with high complexity and poor adaptability to low-bandwidth conditions, limiting their use in emerging applications like real-time remote sensing. While semantic communication offers a promising alternative, existing methods often suffer from semantic information loss. To address this limitation, we propose SCGDF, a novel semantic communication framework that jointly transmits textual descriptions and image features. Our encoder leverages SmolVLM2-0.5B for high-level semantic extraction and a CLIP-based model for feature vector generation, which are then jointly encoded and transmitted. The decoder employs a generative adversarial network (GAN), where the discriminator’s loss function is constrained by the cosine similarity between the transmitted semantic description and the features of the reconstructed image, thereby enforcing high perceptual consistency. Experiments on CIFAR-10 over an AWGN channel (SNR from -10 to 10 dB) demonstrate that SCGDF performs well in SSIM and FID, proving the effectiveness of the method and validating its efficiency and robustness for semantic communication.