Abstract:

Soft image delivery uses pseudo-analog modulation for wireless image transmissions to prevent cliff and leveling effects subject to channel quality fluctuation and to realize graceful quality improvement according to wireless channel quality. Despite its attractive feature of graceful performance, the conventional soft image delivery suffers from low image quality when the analog-modulated symbols are severely impaired by fading and strong channel noise. In this paper, we propose a novel scheme of soft image delivery to reconstruct highquality images from its low-quality observations. Specifically, the proposed scheme integrates deep convolutional neural network (DCNN)-based image restoration, i.e., deep image prior, into soft image delivery. The deep image prior learns a mapping function from the noisy image to the clean image based on user’s perception-aware loss function using multiple training images in prior to soft delivery. The mapping function can restore a clean image even when the received image is distorted by strong fading and additive noise. From the evaluation results, the proposed scheme can remove fading and noise effects from the received images by using DCNN-based image restoration. For example, the proposed scheme achieves up to 0.44 improvement compared with the conventional soft image delivery in terms of structural similarity (SSIM) index at a deep fading channel.