Region-of-Interest Oriented Image Semantic Communication

Semantic communication has garnered significant attention for its potential to alleviate bandwidth constraints in traditional communication systems. Departing from the conventional transmit-then-understand paradigm, semantic communication extracts semantic features from data before transmitting them to the channel encoder. For image data, current methods primarily rely on autoencoder-based deep neural networks (DNNs) to extract semantic features, aiming to accurately reconstruct the entire image. However, since the receiver may be only interested in partial regions, current methods inevitably transmit a considerable amount of semantically irrelevant features. To overcome this limitation, we first propose a region-of-interest (RoI)-oriented, single-layer image semantic communication framework. This framework consists of two key components: 1) a semantic codec responsible for identifying and encoding/decoding the RoI, and 2) a channel codec that compresses/ decompresses and modulates/demodulates the extracted RoI semantic features. During training, a higher loss weight is applied to the RoI reconstruction loss to prioritize reconstruction accuracy in the RoI. Furthermore, to provide a broader range of interest options, we extend the single-layer framework to a multi-layer structure. By assigning different interests to each layer, we achieve progressive image recovery tailored to the RoI. Finally, to mitigate noise accumulation from multi-layer transmission, we design a signal-to-noise ratio (SNR) estimation module that adaptively adjusts channel coding based on the feedback symbols. Numerical results demonstrate the effectiveness of our approach in capturing the RoI while exhibiting enhanced resilience to channel noise.