Underwater image dehazing using global color features

Accurate underwater imaging is a challenging task. Unlike regular photography, visibility underwater is drastically hazed by water molecules and suspended particles distorting light rays and causing differential absorption of different wavelengths of color. Existing research produces well-adjusted but not perfect results. In this work, we present a novel method that dehazed underwater images by estimating global background light based on the optimal eigenvalues of a matrix constructed from three components: (1) gradient of the pairwise wavelength of color channels (blue-red (br), blue-green (bg), and green-red (gr)) (2) gradient of the wavelength of color channels, and (3) the color channels themselves. We estimate transmission maps via scene depth by exploiting the difference in absorption of the different color channel wavelengths. The proposed technique is executed by augmenting UWCNN (Under-Water CNN) with graph-cut theory. The resultant dehazed images successfully show greatly improved color. The proposed results outperform the existing methods in terms of entropy, UIQMn,m, UICM, UISM, and UCIQE. These improvements will provide stronger imaging tools to domains like submarine search and rescue, navigation, oceanography, and mapping. There is room for future research because our process slightly darkens images which leads to marginally lower UIConM values. Future study can also further evaluate the effect of color channels while assuming pixels with the highest intensity are used as the global ambient light.