Image Transmission Based on Semantics Communication in Vehicle-Infrastructure Collaboration System

In the intelligent vehicle infrastructure collaboration system (i-VICS), the sensing task evolves from being realized by a single vehicle to being accomplished by multiple automatons such as roadside and multi-vehicle collaborations. Image is a key carrier for information transmission and is the most important information source to improve the efficiency of traffic operation driving experience with comfort and safety. Traditional image coding methods focus on the research of preserving the maximum amount of information with as little codes as possible. However, limited by Shannon’s separation theorem, image communication based on conventional coding can only infinitely approximate the optimal information transmission performance that usually decided by the channel with less resource. In particular, when the wireless channel conditions change worst, the coding could not cope with the change dynamically, and then bit error rate increases sharply. To overcome the above deficiencies, this paper proposes an image transmission system based on semantic communication, called ITSSC to transmit images for i-VICS. Deep learning-based joint source-channel coding (Deep JSCC) is deployed to extract and encode the semantics for transmission. Swin transformer-based attention mechanism is applied to expand the perceptual field of features and improve the quality of image recovery. Finally, the proposed ITS-SC is validated on the traffic dataset KITTI. The results show that our proposed method can meet the image transmission requirements for the i-VICS scenario with bandwidth resources saving.