We investigate the image super-resolution problem by considering the power savings and performance improvement in image acquisition devices. Toward this end, we develop a deep learning based reconstruction network for images compressed using hardware-based downsampling, bit truncation and JPEG compression, which to our best knowledge, is the first work proposed in the literature. This is motivated by the fact that binning and bit truncation can be performed on the commercially available image sensor itself and results in a huge reduction in raw data generated by the sensor. The combination of these steps will lead to high compression ratios and significant power saving with further advantages of image acquisition simplification. Bearing these concerns in mind, we propose LISR-net (Lossy Image Super-Resolution network) which provides better image restoration results than state-of-the-art super resolution networks under hardware constraints.
View on arXiv