Underwater Image Restoration via Machine Learning Transmission Map of Atmospheric Scattering Model

Traditionally, image dehazing methods must rely on employing the Atmospheric Scattering Model (ASM) to perform image restoration. These methods are characterized in that two key parameters: atmospheric light A(x) and Transmission t(x) of the entire image are either assumed (but often inaccurately) or heuristically estimated, However, the estimation of A(x) and t(x) is non-trivial. In particular, since (x) varies with the scene depth, making ASM not attractive, especially for underwater applications wherein different wavelengths of light are absorb to different degrees. In this paper, we propose a novel approach which is different from traditional counterparts in that t(x) is learned directly using color features of hazed image as training input, and t(x) of its corresponding clear image as target output for training deep learning network. Our method comprises three steps: First, the input feature vectors are prepared by concatenating the global and local color features of the training and test input images. Second, the color features are standardized and subjected to training. Finally, prior to image restoration, the output $\hat t(x)$ of the trained network is blurred using the locality property. Using a six-layer neural network [32,128, 256,128, 32, 1] as an illustrative example, our method outperforms Fusion-based [1], Retinex-based [2], IBLA [3], AGC [4], GDCP [5], UWCNN [6], BL-TM [7], DehazeNet [8], in terms of contrast, sharpness, as well as the overall restoration appearance.