AdaJSCC: Instance-Adaptive Joint Source-Channel Coding for Wireless Image Transmission

Recent works in wireless image transmission based on deep joint source-channel coding (DeepJSCC) have exhibited superior performance against many traditional source and channel coding methods. While so far, research efforts mainly concentrate on architecture and model improvements toward a static target domain. In this work, we propose a novel lightweight instance-adaptive DeepJSCC approach to further improve the transmission performance for each image instance. We leverage the idea of a dynamic convolution kernel to build the DeepJSCC codec. In such way, apart from the general coded content stream, an additional model stream is extracted from each individual image instance and transmitted to the decoder side for generating the dynamic convolutional kernel parameters. The model stream enables the codec to update the off-the-shelf pretrained models after deployment in a lightweight online fashion. Experiments over various wireless channel models verify the effectiveness and robustness of our method.