Wireless Image Transmission Using Deep Source Channel Coding With Attention Modules

Recent research on joint source channel coding (JSCC) for wireless\ncommunications has achieved great success owing to the employment of deep\nlearning (DL). However, the existing work on DL based JSCC usually trains the\ndesigned network to operate under a specific signal-to-noise ratio (SNR)\nregime, without taking into account that the SNR level during the deployment\nstage may differ from that during the training stage. A number of networks are\nrequired to cover the scenario with a broad range of SNRs, which is\ncomputational inefficiency (in the training stage) and requires large storage.\nTo overcome these drawbacks our paper proposes a novel method called Attention\nDL based JSCC (ADJSCC) that can successfully operate with different SNR levels\nduring transmission. This design is inspired by the resource assignment\nstrategy in traditional JSCC, which dynamically adjusts the compression ratio\nin source coding and the channel coding rate according to the channel SNR. This\nis achieved by resorting to attention mechanisms because these are able to\nallocate computing resources to more critical tasks. Instead of applying the\nresource allocation strategy in traditional JSCC, the ADJSCC uses the\nchannel-wise soft attention to scaling features according to SNR conditions. We\ncompare the ADJSCC method with the state-of-the-art DL based JSCC method\nthrough extensive experiments to demonstrate its adaptability, robustness and\nversatility. Compared with the existing methods, the proposed method takes less\nstorage and is more robust in the presence of channel mismatch.\n