Learning a Basis to Compress Underwater Media

A practical method of visually lossless image compression is proposed that has an estimated compression ratio better than 100:1. The method combines tested error-tolerant encodings with modern super resolution techniques that are known to have good rate-distortion performance. Current error-tolerant encodings are often designed for optical data, which have a spatial frequency distribution that varies from sonar and other modalities. This proposal outlines simple changes to the encodings that can cater to different spatial frequency distributions and focus on spatial frequencies that are informative to the super resolution method. These changes are learned using representative data, which makes them extensible to new modalities or operating conditions. The proposed method does not require extensive computation on the vehicle or edge device, making it a practical solution for robotic applications.