A Novel Lightweight and Robust Source-Channel Coding Solution for MIoT Communication Based on DL Denoising/Super Resolution Model

Abstract

Multimedia Internet of Things (MIoT) network is prone to a variety of challenging constraints, especially in terms of performance and security. Employed MIoT devices can be limited in terms of power, computation, and memory, which make them suffer with the high volume of collected multimedia data. In this context, data compression is one solution to reduce the size of communicated data. However, existing lossy multimedia compression algorithms, such as JPEG and BPG, impose a practical challenge for several MIoT devices since they require high computation and memory resources. Another challenge is the errors that can occur during transmission due to channel errors, which require re-transmitting the erroneous data. In this case, channel coding is one solution. However, channel coding solutions impose overhead in terms of computation and communication resources. To reduce this overhead, in this paper, we propose a lightweight source and channel coding solution that should be applied only on the application server. This solution consists of down-scaling each input image with a factor α ≥ 2 at the MIoT device. This reduces the computation, communicated data size, and memory consumption, which would consequently reduce both energy consumption and latency. However, the down-scaled image might be corrupted by channel errors due to the reliance on the wireless connection. In addition, if the down-scaled image was encrypted, an error in one data block will propagate to other data blocks after the decryption process at the application server. Thus, to avoid costly data re-transmission or redundancy, our solution proposes to apply a Deep Learning (DL) denoising/super-resolution model at the server-side to recuperate high-quality images. This model plays the role of source and channel coding algorithm. The obtained results show the effectiveness of the proposed solution, especially in terms of enhancing the visual quality of the reconstructed high-resolution images from downscaled and erroneous ones with low latency, communication, and resource overhead.