Sensor Modality Translation through Contrastive Deep Learning

Physical Sciences Inc. (PSI), in collaboration with the University of Rhode Island, proposes to develop an advanced algorithm suite for data translation across sensing modalities to support the development of automated target recognition and classification algorithms for Unmanned Underwater Vehicles. The proposed Deep Diffusion Sensor Translation (DDST) leverages recent advancements in generative artificial intelligence, and latent diffusion models in particular, to enable highly realistic data synthesis to supplement these underwater ATR datasets. The DDST tool will be capable of translating between sidescan sonar, forward looking sonar, synthetic aperture sonar, imaging magnetometry, and visible sensing modalities. The DDST incorporates advancements in underwater image enhancement and three-dimensional scene reconstruction to normalize variability across instruments, environments, and sensing conditions. The DDST will also leverage PSIÆs computer vision and image fusion expertise developed under multiple DoD programs, through customization of an in-house super-resolution technique to the task of enhancing DDST inputs and outputs. The DDST technology will produce synthetic sensor outputs with quantifiable accuracy, achieving acoustical and optical reflectivity accuracies with PSNRs of 30dB and 25dB respectively.