Self-Supervised Underwater Intelligent Perception for Deep-Sea Cage Aquaculture

In deep-sea net-pen aquaculture, underwater intelligent sensing is performed by an underwater camera for information acquisition, but underwater scattering and absorption effects affect the acquisition of underwater information. It is challenging to use neural networks to process the net tank aquaculture scenarios because the underwater data sets of the net tank aquaculture scenarios are not accessible. In this article, we propose a self-supervised deep-sea scene recovery method utilizing a homology constraint and a fusion strategy. Specifically, the scene radiation maps are derived based on a neural network and a prior extraction architecture, respectively, and two scene radiation maps originate from two different computational regimes. Finally, the perceptual fusion strategy is used to blend two scene radiation maps to obtain better performing results and minimize the error using the homology constraint. Extensive experiments confirm that the approach using perceptual fusion has excellent recovery capabilities. It is demonstrated through extensive experiments that our method outperforms state-of-the-art methods in terms of both visual quality and quantitative metrics.