An Underwater Image Enhancement Method Based on Balanced Adaption Compensation

Compensation for underwater optical image enhancement has achieved good performance by reducing the discrepancies among attenuated channels. Physical-based compensation approaches usually employ non-convex optimization, which results in ill-posed compensation in practice. However, the enhancement results of learning-based methods mostly tend to a certain pattern of synthesized data. In this paper, a balanced adaption compensation (BAC), leveraging the difference between learning-based and physical-based compensation applicabilities, is proposed to adapt the performance of data-driven semantic transfer and scene-relevant reconstruction. The adaptation of BAC breaks the limitation of scene diversity on compensation and reduces the dependence of the trained network on training patterns. Moreover, BAC is a perceptual enhancement method, that preserves the texture information of targets to ensure that enhanced results of BAC suit high-level visual tasks. Extensive qualitative and quantitative evaluations of underwater image enhancement with BAC show notable performance improvement against other compensation methods.