Computational imaging through turbid, bubbly and flowing water near the telescope

• Prior restoration work generally assumes a static medium or studies a moving air-water interface; the paper reports no earlier measured-ground-truth study of bubbly, turbid water with flow.Source: §1 · p. 1
• The dataset records seven combinations of flow, bubbles, and turbidity with displayed-image and clear-water references, including recordings of physical objects.Source: §1 · p. 3
• Four- to five-fold contrast loss with lesser blur is measured on a Siemens star, consistent with scattering spread far beyond the field of view.The measurement applies with and without bubbles and flow.Source: §1 · pp. 2–3
• Eight-frame burst fusion adds little beyond a network trained on the frame mean across three architectures.Source: §1 · p. 3
• Synthetic training pairs leave displayed-image scores unchanged while raising Siemens-star correlation with its clear-water recording from below 0.5 to above 0.9.The approach extends the authors’ defogging work to dynamic water and supports transfer to physical objects without inventing structure.Source: §1 · p. 3
• One raw frame supports five-category recognition at 96.8%, matching recognition from restored images, while the chosen reference reverses raw-versus-restored rankings.Source: §1 · p. 3

2.1. Apparatus and data

• A 50 mm-aperture telescope records images through a 25 mm water path 45 mm from the objective across seven water conditions; paired recordings and physical objects support evaluation.Displayed-image identities stay within one train, tune, or evaluation split across conditions, while 24 physical-object recordings are evaluation-only.Source: §2.1 · p. 3

2.2. Image-formation model

• The image-formation model treats transmission, point-spread function, shift, veil, and noise as time-varying but shared across image positions.Near-lens geometry yields almost complete overlap of center and edge beam footprints; bubbles and flow change blur and shift, while turbidity lowers transmission and raises veil.Source: §2.2 · p. 4

2.3. Restoration, baselines and recognition

• A joint-RGB Restormer maps recorded frames to displayed targets, using measured pairs and synthetic pairs degraded with fitted attenuation, veil, blur, and noise.One update in four uses a synthetic pair; its generator does not explicitly model bubbles or refraction.Source: §2.3 · p. 4
• The study compares separately trained burst and repeated-mean inputs, architecture controls, a measured-only model, and a per-channel contrast stretch of the eight-frame mean.Recognition uses a pretrained ResNet-50 with five output categories and tests raw single frames, frame embeddings, and restored outputs.Source: §2.3 · pp. 4–5, Figure 2 · p. 5

3.1. Eight frames add little beyond their average

• Averaging eight frames before Restormer restoration adds 0.0322 SSIM over single-frame training; learned burst fusion changes trained-mean SSIM by −0.0003.Across conditions, trained-mean SSIM was 0.540 in C+B+M and 0.516 in T+B, with the three turbid conditions ranging from 0.516 to 0.586.Source: §3.1 · p. 6

3.2. Restoring images of real objects

• Turbidity cut Siemens-star outer modulation 4.4-fold from C to T and 4.8-fold from C+B+HM to T+B+HM; burst restoration reached SSIM 0.61–0.66 in turbid conditions.A contrast stretch revealed object structure, but in turbid conditions it amplified noise and lowered SSIM from 0.55–0.56 to 0.21–0.25.Source: §3.2 · p. 6

3.3. Synthetic training data lets restoration transfer to physical objects

• On physical objects, synthetic-trained networks stay within 0.05 of the recorded mean’s 0.92–0.94 Siemens-star correlation, while measured-only networks score 0.35–0.47.On displayed images, the mixed-minus-measured Restormer difference was −0.0006 SSIM; synthetic-trained networks also gave cleaner tone and less noise on physical objects.Source: §3.3 · pp. 6–7

3.4. Transfer to ordinary underwater photographs

• Without retraining, synthetic-trained networks raised UIEB SSIM from 0.760 to 0.765–0.846; measured-only Restormer scored 0.359, while contrast stretch scored 0.874.Figure 4 shows difficult dark-fish and reef scenes; UIEB references are human-selected enhancements, so the scores measure agreement with preferred enhancement rather than fidelity.Source: §3.4 · pp. 7–8, Figure 4 · p. 8

3.5. Recognition works on raw frames

• One raw frame classified 96.76% of evaluation images, versus 97.22% for burst-restored images; in C+B+M, accuracy rose from 80.7% to 91.3%.The C+B+M paired difference was +10.7 percentage points, with a 95% interval of +3.3 to +18.7 across 50 images in five categories.Source: §3.5 · p. 8

4. Discussion

• Near-lens beams overlap, making degradation approximately spatially uniform; wider-angle scattering removes light or adds a veil, while narrower lobes or wider fields move toward blur.The authors connect this regime to limited useful frame diversity and explain why contrast stretch and raw-frame classification can work well.Source: §4 · pp. 8–9
• The proposed restoration recipe combines measured pairs for realism with synthetic pairs from an attenuation-and-veil model to train networks to invert the medium.The discussion says restoration can improve tone and suppress stretch-amplified noise, while bubbles changing the point-spread function can raise recognition in clear water.Source: §4 · p. 9
• Against the digital source, the burst network beats the eight-frame mean by 0.294 SSIM and 11.25 dB PSNR; against clear water, the mean scores 0.842 SSIM and 27.24 dB versus 0.449 and 8.83 dB for burst restoration.The discussion cautions that neither reference alone measures fidelity: the digital-source target includes display and camera response, while a clear-water recording retains it.Source: §4 · p. 9
• The physical-object evidence is limited to static targets, one turbidity level, one bubble source, and 24 recordings of six objects; longer scattering paths or narrower lobes fall outside the stated model.Turbid and clear-water recordings came from different sessions, and evaluation images had been inspected during earlier analyses.Source: §4 · p. 9

5. Conclusion

• Contrast stretching reveals physical-object structure after recorded contrast falls more than fourfold, while averaging eight frames captures the multi-frame gain and learned fusion adds little.Synthetic training pairs also support transfer from displayed images to physical objects and ordinary underwater photographs, with less invented structure than training on measured pairs alone.Source: §5 · p. 9
• Frame averaging and contrast stretching are low-cost steps, and raw-frame classification could let underwater vehicles recognize objects without onboard restoration.A narrow-field optic keeps processing simple while the scattering lobe remains wider than the field of view.Source: §5 · p. 9
• Measured and synthetic pairs, evaluated against references that separate tone from structure, offer a recipe for learned imaging through rain on windows or fog on windshields where ground truth is scarce.Source: §5 · pp. 9–10