Deep Learning-based Image Restoration for Low-Power and Lossy Networks

Multimedia Internet of Things (IoMT) is witnessing explosive growth due to its applications in multiple areas. To cope with limited resources of low-power and lossy networks (LLN), it is common that: (i) images are captured with a degraded quality due to limited camera capabilities, (ii) a low-cost lossy compression is applied to reduce the amount of data to deliver which introduces additional distortion and (iii) transmissions are prone to losses that induce holes in the images, further degrading their quality and making them difficult to use. In this work, we propose a complete efficient encoding-transmission-reconstruction chain. In addition to the use of a low complexity image compression method, an appropriate packetization scheme is proposed. At the destination, more powerful resources are leveraged to apply deep learning models to compensate for the distortion caused by the adopted lossy compression as well as to fill in the holes induced by packet losses. The obtained results show the effectiveness of our proposal.