Signal Reconstruction for Secure Imagery Communications
Samantha S. Carley, Stanton R. Price · ASCEND 2022 · 2022
View Video Presentation: https://doi.org/10.2514/6.2022-4220.vid We propose an image reconstruction pipeline that can intelligently reproduce an original image from a fraction of its original pixels. Specifically, we propose using the convolutional neural network U-Net to reconstruct “degraded” imagery to its raw, original form. We define degraded imagery as imagery that has been reduced in context by a defined percentage; essentially, a percentage of pixels are zeroed out of or taken out of the image. We investigated (1) the performance of U-Net for image reconstruction and (2) the effect of increasing levels of image degradation for various trained models and the observation of its effects on the reconstructed images. In this work, all research is simulated, meaning we are not analyzing an application of this method on real systems in real time, but are merely studying the method digitally for its reconstruction ability. The experiments include five trained models, using the U-Net architecture and similar training parameters, which are individually trained on different levels of pixel reduction. The investigation includes both a qualitative and quantitative analysis to display the comparison between the original imagery and the reconstructed imagery. Results show that each trained model can handle different levels of reduction within a range and are able to reconstruct context-limited images (images with greater than 75% of the pixels removed). We have been exploring this technology from the context of communication between collaborating systems. Areas that may be of interest for this technology is additional encryption for cybersecurity of an incoming/outgoing signal and efficient communication-linking between systems. The publicly available xView dataset is used in our experiments and analysis.