Image Coding and Compression Method Based on Semantics of Pixels for Semantic Communication

Semantic image transmission is promising for both image reconstruction or downstream task execution. However, it is still challenging to simultaneously achieve pleasing visual clarity and superior downstream tasks. To tackle this issue, we propose a semantic-preserving image coding and compression method based on the semantics of pixels (SP). Specifically, we first project the semantic importance onto pixels, which is quantified by a gradient-based mechanism. Then, we design an SP-based loss function to train the image coding and compression network. This loss function not only preserves the semantic information to boost the performance of intelligent tasks but also preserves the pixel information to achieve pleasing visual effects. Experimental results confirm improvements in both aspects of performance. Especially for intelligent task performance, the proposed method achieves up to 22.22 % and 78.40 % performance gain ratio compared with the state of art deep learning-based and the traditional method, respectively.