Underwater Image Enhancement Based on Intrinsic Images

The complex and changeable underwater environment and light conditions lead to quality degradation problems of underwater image in different degrees. Therefore, underwater image enhancement is an important and challenging task in computer vision. Relying on a powerful visual system, human beings have the ability to restore the real scene from a degraded image, and estimate the object material and light situation of the scene in the image. Inspired by this, we believe that separating the underwater lighting effect from the degraded image to produce enhanced result, is an effective method for underwater image enhancement. We utilize the powerful fitting capabilities of CNN to estimate the illumination image, separate them from underwater image and obtain reflectance image, and we treat the reflectance image as the final enhanced underwater image. Our underwater image enhancement model was implemented by leveraging the light transfer method, and well results were achieved state-of-the-art performance on both public UIEB [15] and EUVP [11] datasets.