Fig. 1. Analysis of video frames of underwater structure detection with visual occlusion from suspended impurities

Fig. 2. Flow chart of suspended impurity occlusion removal method

Fig. 3. Flow chart of background displacement compensation and adjacent frame alignment

Fig. 4. Flow chart of detection process of underwater suspended impurity area based on dynamic visual perception

Fig. 5. Flow chart of suspended impurity occlusion area inpainting guided by the hybrid model

Fig. 6. Comparison between the proposed method and three typical marine snow removal algorithms. (a) Original images; (b) Banerjee method; (c) Cyganek method; (d) Koziarsk method; (e) proposed method

Fig. 7. Detection and inpainting effect of real underwater suspended impurity occlusion video

Fig. 8. Comparison of details before and after occlusion removal of suspended impurity

Fig. 9. Comparison of image quality indicators before and after occlusion removal of real underwater videos. (a) Comparison of mean values for Brenner function indicators; (b) comparison of mean values for EOG function indicators; (c) comparison of mean values for Laplace function indicators; (d) comparison of mean values for Tenengrad function indicators

Fig. 10. Occlusion elimination results of video frame 1 for synthetic underwater suspended impurities. (a) Synthetic images; (b) original images; (c) ground truths; (d) detection masks; (e) hybrid model maps; (f) restored images; (g) difference maps

Fig. 11. Occlusion elimination results of video frame 2 for synthetic underwater suspended impurities (a) Synthetic images; (b) original images; (c) ground truths; (d) detection masks; (e) hybrid model maps; (f) restored images; (g) difference maps

Fig. 12. Comparison of image quality evaluation indexes before and after inpainting of synthetic underwater suspended impurity videos. (a) Comparison of PSNR means; (b) comparison of SSIM means; (c) comparison of FSIM means; (d) comparison of GMSD means