Perceptually Lossless Unmanned Underwater Vehicle (UUV) Sensor Data Compression

Data through water transfer rates are generally low due to limited bandwidth. The sonar images must be efficiently compressed to increase onboard storage and enable large amounts of data through water transfer. Mathematically, the sonar images can be viewed as a linear combination of some appropriate Green’s functions. Thus, they are highly coherent correlated and the associated matrices are rank deficient. This unique property can be exploited to achieve better sonar image compression. To this end, Corvid Technologies (Corvid) and Lawrence Berkeley National Laboratory (LBNL) propose to develop an efficient data-driven dictionary sparse coding-based sonar image compression algorithm. Firstly, the team will employ butterfly factorization to extract the salient features at different levels from the training data, then we will use a deep dictionary learning algorithm to build a more accurate dictionary by taking the butterfly factorization results as input. Once the dictionary is learned, Corvid and LBNL will use a Bayesian compressive sensing approach to extract the sparse coefficients of each image more efficiently. For the scenarios that multiple sonar images are associated with the same scene, we will use a multi-task Bayesian compressive sensing method to obtain more noise robust results by exploiting the joint information among those images. Due to sparse coding, the extracted coefficients can serve as features for better automatic target recognition performance.