Authors: Moheb M. R. Henein, Mahmoud Ayyad, Marouane Temimi – Stevens Institute of Technology
Title: A Framework for Restoring and Processing Degraded Images from USGS Hydrologic Imagery Visualization and Information System
Abstract: Optical monitoring of river streamflow, ice formation, and other hydrological analysis using on-the-ground cameras requires preliminary enhancement of the collected images to augment their potential. Numerous USGS stations are equipped with ground-based cameras to continuously monitor flow conditions. The collected images from USGS cameras experience adverse weather conditions such as rain, fog, and snow. These phenomena have a demonstrably negative impact on the visual quality of images captured by outdoor vision systems, significantly hindering their performance and limiting their potential in the determination of flow conditions. A restoration framework is proposed to reconstruct a better image with minimal noise introduced by raindrops, haze, and snow. A framework is developed for a potential automated deployment and processing of newly received images. Firstly, the level and type of degradation are predicted, hence, the image can be used directly if it is clear, processed by a restoration deep learning model, or rejected as a corrupted image. The degradation level for each type of noise is determined by three machine learning classification models, one for each type of noise. Two performance metrics were used to assess the performance of the proposed framework, namely, PSNR (Peak signal-to-noise ratio) and SSIM (Structural similarity index). Results on synthetic and real degraded haze images show that our framework can correctly restore an image for further analysis. For example, our framework has restored images with an average of approximately 20 PSNR and 0.95 SSIM under low, moderate, and severe haze degradations.