Fig. 1. Flowchart of proposed method

Fig. 2. Depth map estimation. (a) Original image; (b) gradient map; (c) estimated depth map based on gradient; (d) estimated depth map based on color difference

Fig. 3. Results of background region segmentation. (a) Original image; (b1) gradient map; (b2) enhanced gradient map; (b3) edge information; (b4) initial detection only using gradient information; (c1) maximum value of G-B channel; (c2) R channel; (c3) difference between two channels; (c4) initial detection only using color difference; (d) candidate region of background light

Fig. 4. Comparison of different background light estimation strategies. (a) Original underwater images; (b)-(d) estimated background light and corresponding restored results using method of reference [7], method of reference [9] , and proposed method

Fig. 5. Comparison of restored results using different methods. (a) Original images; (b)-(f) restored results by RCP, IBLA, ULAP, UWCNN, and proposed method

Fig. 6. Failure cases of proposed method. (a) Scenes with non-uniform illumination; (b) scene depth and background light estimation; (c) restored results

Table 1. Comparison of different quality evaluation metrics

Table 2. Comparison of average values of different quality evaluation metrics