Underwater image enhancement via standard deviation reconstruction, dynamic sigmoid mapping and gradient-aware enhancement

Underwater images collected by optical electronic devices often suffer from color cast and low contrast due to the wavelength-dependent attenuation of light propagation. To deal with these challenging issues, this paper proposes an underwater image enhancement method based on an adaptive standard deviation reconstruction, a dynamic sigmoid mapping and a gradient-aware contrast enhancement, called SDGE. Concretely, we first study the influence of the standard deviation of Gaussian kernel on the color restoration results regarding different underwater images, and find it has a marked impact on colors. Meanwhile, we rebuild the standard deviation of Gaussian kernel based on the attenuation characteristics of each image, which enables us to adaptively adjust the smoothness degree of different images, thereby eliminating the color veil and preserving the significant structure and texture. Then, we utilize the maximum value of each channel to design diverse sigmoid functions, and dynamically simulate the nonlinear mapping of human color perception, achieving the correction of dynamic range and color distortion. Finally, we use an histogram equalization method to optimize the structure and design a gradient-aware coefficient to amplify the detail for producing the enhanced image. Broad experiments on four underwater image datasets demonstrate that the enhancement methods with our SDGE produce the considerable results in both qualitative and quantitative evaluations, i.e., our method achieves the best performance on the UCCS and UIQS datasets regarding all quantitative metrics, and our method at least increases by 95.67%, 76.74%, 1.28%, 9.94% and 3.75% compared with the second-best method in terms of the e, $\overline{r}$, Entropy, UIQM and UCIQE values, respectively. These competitive datas indicate our method has the excellent performance in enhancing the clear visibility, correcting the natural appearance and highlighting the details. Moreover, our method provides superior generalization ability for enhancing complex remote sensing images.