PLOS ONE Retracts Underwater Image Enhancement Paper Over Similarities to Prior Work

A 2024 study that promised a new way to sharpen the murky images captured beneath the ocean’s surface has been struck from the scientific record. On October 12, 2026, the editors of PLOS ONE issued a formal retraction of the paper, originally published in February 2024 under the title “Underwater image enhancement using multi-task fusion,” citing similarities between the article and previously published work by different authors that call into question the validity and provenance of the reported results. The retraction notice, bearing the DOI 10.1371/journal.pone.0360350, closes a chapter that began not with a dramatic scandal but with something quieter and, in modern science, increasingly common: a reader or automated system noticing that a published paper looked too much like someone else’s.

The retracted study, authored by K. Liao and X. Peng, appeared in volume 19, issue 2 of the journal and carried the PubMed identifier 38408101. Its subject was one of the most practically important problems in computer vision: restoring clarity to images degraded by water. Anyone who has seen footage from a consumer action camera a few meters below the surface knows the problem. Light is absorbed and scattered as it travels through seawater, and the absorption is not uniform across the spectrum. Red wavelengths vanish first, within the first few meters, followed by orange and yellow, leaving the characteristic blue-green cast that dominates underwater photography. Suspended particles scatter what light remains, washing out contrast and blurring fine detail. For marine biologists counting fish populations, archaeologists documenting wrecks, engineers inspecting pipelines, and autonomous underwater vehicles navigating by camera, that degradation is not merely an aesthetic nuisance; it is an obstacle to science and industry alike.

Multi-task fusion, the approach named in the retracted paper’s title, sits within a broader family of deep learning techniques that have transformed underwater image enhancement over the past decade. Rather than training a single neural network to perform one transformation, multi-task architectures ask a model to learn several related objectives at once, for example color correction, dehazing, and edge or detail restoration, and then to fuse the outputs of these parallel tasks into a single enhanced image. The rationale is that the tasks are complementary: a network that learns to estimate the scattering of light in turbid water may also learn representations that help it recover true colors, and combining the task-specific outputs can produce results that are more robust than any single-task model. Variants of this idea, including hybrid fusion methods that blend classical physics-based models with learned components, have proliferated in the computer vision literature, particularly in venues focused on engineering applications of artificial intelligence.

It was precisely in that crowded field that the problem emerged. According to the retraction notice, similarities were identified between the Liao and Peng article and a paper by different authors: “HFM: A hybrid fusion method for underwater image enhancement,” by S. An, L. Xu, Z. Deng, and H. Zhang, published in Engineering Applications of Artificial Intelligence in 2024. The notice states that these similarities call into question the validity and provenance of the reported results, language that in editorial practice typically signals concerns about overlap in text, figures, methods, or results that cannot be reconciled with independent discovery. The PLOS ONE Editors concluded that retraction was warranted, and the article now carries the retracted label in the journal’s own records.

What happened next is as revealing as the finding itself. The retraction notice reports that all authors either did not respond directly or could not be reached. That silence matters, because retraction is normally a process, not a verdict handed down unilaterally. Journals typically contact authors, invite a response, and in cases of honest error allow corrections or partial retractions. When authors cannot be reached or decline to engage, editors must weigh the evidence they have. In this case, the documented similarity to prior work by different authors was sufficient for the editors to act, and the notice was published with the authors’ non-response recorded for posterity. The episode illustrates a structural weakness of the modern publishing system: authorship confers credit and responsibility in equal measure, but when questions arise, authors who vanish leave editors to adjudicate alone.

The case also highlights how retractions are discovered in the first place. The notice does not describe the mechanism, but the broader ecosystem includes several. Software such as iThenticate screens submissions for textual overlap before publication. Post-publication peer review communities and independent watchdogs compare papers line by line. Image forensics specialists examine figures for duplicated panels or recycled data. And increasingly, readers themselves use search tools to check whether a paper’s methods section mirrors an earlier publication. Underwater image enhancement is a small enough subfield that researchers know each other’s work intimately, which makes unattributed reuse especially likely to be caught. A paper on a niche topic does not disappear into the noise; it lands on the desks of the few dozen people who read everything in that niche.

For the research community, the consequences of a retraction are concrete. The article remains online, as PLOS ONE and most publishers keep retracted papers visible with a prominent watermark and notice, but it is flagged in indexing databases, and the retraction is linked in both directions: the retracted paper points to the notice, and the notice points to the paper. Citation practices shift accordingly. Researchers who cited the work are expected to update their references where the retracted findings are load-bearing, and systematic reviews must screen for retracted studies. The authors of the prior work, An and colleagues, whose hybrid fusion method appeared in Engineering Applications of Artificial Intelligence, now have a documented priority claim in the record, which matters in a field where methodological novelty determines funding, careers, and attribution.

The scientific substance at stake deserves attention too. Underwater image enhancement is not an abstract exercise. Accurate color recovery underpins automated coral reef monitoring, where classification algorithms trained on degraded images can misidentify species. Clear imagery improves the performance of object detection models that track fish stocks and detect illegal fishing. In underwater robotics, visual odometry, the technique of navigating by tracking features across camera frames, degrades badly in turbid water, so enhancement pipelines feed directly into navigation reliability. The field has consequently attracted intense competition, with benchmark datasets such as UIEB and others driving rapid iteration. That competition is productive, but it also creates pressure: a crowded subfield with fast-moving benchmarks rewards authors who publish quickly, and occasionally tempts some to cut corners. Retractions in hot areas are not evidence that the science is rotten; they are evidence that the sociology of competition operates the same way in computer vision as it does in biomedicine.

PLOS ONE, as a mega-journal that publishes thousands of papers annually and emphasizes scientific soundness over novelty, has developed a substantial institutional apparatus for handling these cases. Its editorial policy treats retraction as the mechanism for correcting the literature when findings are unreliable, and its notices are deliberately factual, stating what was found and what the authors did or did not do in response. The notice in this case follows that template precisely: it names the concern, cites the prior work, records the authors’ silence, and states the editorial decision. Readers seeking the underlying evidence are directed to the two papers themselves, whose DOIs are provided in the notice’s reference list. That restraint, saying only what can be documented, is what separates a retraction notice from an accusation, and it is why retraction notices have become a genre of scientific literature in their own right.

For the wider public, the takeaway is less about one paper than about the immune system it reveals. Science is often imagined as a straight line from discovery to publication, but the record is continuously edited: errors are corrected, datasets are updated, and papers that cannot stand are removed. The retraction of the underwater image enhancement study is a small, well-documented example of that editing in action, caught by comparison with prior work, adjudicated by editors, and recorded permanently with a DOI of its own. The legitimate work in the field, including the hybrid fusion research that preceded it, continues to advance the goal that motivated the retracted paper in the first place: teaching machines to see clearly through water, so that the oceans can be mapped, monitored, and understood with the sharpness that their study demands.

Subject of Research: Retraction of a PLOS ONE paper on underwater image enhancement using multi-task fusion due to similarities with prior published work

Article Title: Retraction: Underwater image enhancement using multi-task fusion

Article References: The PLOS One Editors (2026). Retraction: Underwater image enhancement using multi-task fusion. PLOS One, 21(10), e0360350. https://doi.org/10.1371/journal.pone.0360350

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360350

Keywords: PLOS ONE, retraction, underwater image enhancement, multi-task fusion, computer vision, deep learning, research integrity, plagiarism, image processing, scientific publishing, hybrid fusion, color correction