Academic Integrity Review Report: 'Towards Maximizing Semantic Coverage for Image-Text Retrieval' (ICASSP 2025)

Academic fraud report · Geng Detector

Summary

This review assesses the ICASSP 2025 paper 'Towards Maximizing Semantic Coverage for Image-Text Retrieval' (DOI: 10.1109/ICASSP49660.2025.10888503). The overall verdict is 'Questionable' (medium-severity concerns). The central finding is an apparent numerical inconsistency between the paper's claimed improvements and its own reported table values. The authors state an average rSum gain of 6.8% on Flickr30K and 4.9% on MS-COCO(1K), but direct calculation from Table I yields approximately 1.14% over the strongest baseline IMEB (530.1 vs. 524.1) and roughly 1% on MS-COCO—far below the stated figures. A secondary concern involves Table II (ablation study), where rows two and three appear to use identical module configurations but report different results (519.7 vs. 523.6), suggesting possible copy-paste errors. No image-level manipulation analysis was performed (insufficient pixel data). Citation timeline and baseline references appear legitimate. The confidence level is moderate; the numerical mismatch is reproducible from the paper, while the ablation issue is flagged as a possible reporting error rather than proven fraud.

Verdict

Questionable / Medium-severity concerns. The paper contains a clear and reproducible discrepancy between text claims and table data regarding reported performance gains, plus a structurally inconsistent ablation table. These may reflect careless reporting rather than intentional fabrication, but they fall below the rigor expected for a peer-reviewed venue.

Key findings

Inconsistent improvement claims vs. table values (medium severity): The text states a 6.8% average rSum improvement on Flickr30K and 4.9% on MS-COCO(1K). Direct calculation from Table I does not reproduce these figures.
Ablation table inconsistency (low–medium severity): Table II rows 2 and 3 appear to share the same module configuration (✓ ✓ ✓) yet produce different scores (519.7 vs. 523.6), which breaks standard ablation logic.
No image-level analysis possible: Figures 1–3 could not be examined at the pixel level (insufficient raw image data); no duplication or splicing evidence observed from text alone.
Reference timeline is consistent: 2023–2024 baselines cited appropriately for a 2025 venue; no anachronistic citations detected.

Evidence highlights

Flickr30K (BERT backbone), Table I:
Strongest baseline IMEB [29]: rSum = 524.1
SCA-Net (proposed): rSum = 530.1
Absolute gain: 530.1 − 524.1 = 6.0
Relative gain vs. strongest baseline: 6.0 / 524.1 = ≈1.14%
Relative gain vs. weakest baseline PFAN++ (477.6): (530.1 − 477.6) / 477.6 = ≈10.99%
Neither calculation yields the stated 6.8%.
MS-COCO(1K): Highest actual relative gain reported is roughly 1%, not the stated 4.9%.
Table II ablation rows 2 vs. 3: Identical checkmarks (1⃝, 2⃝, 3⃝) but different results (519.7 vs. 523.6).
DOI: 10.1109/ICASSP49660.2025.10888503 (ICASSP 2025).

Notes

The numerical mismatch is a reproducible, quantitative observation from the paper itself and not dependent on image forensics.
The ablation table issue could stem from a typographical/copy-paste error (e.g., a missing module checkmark in row 3) rather than data fabrication; clarification from the authors is warranted.
Authors and venue information (Junhao Xu, Zheng Liu, Zhumin Chen, Fei Shen; ICASSP 2025) appear legitimate; no authorship anomalies were investigated.
Recommended follow-up: post on PubPeer requesting (a) the calculation basis for the 6.8% / 4.9% improvement figures, (b) clarification of the Table II row configurations, and (c) release of code for independent verification.
Final determination of academic misconduct requires institutional investigation; this report is an evidence summary, not a verdict of misconduct.