Integrity Assessment Report: FE-YOLOv5: Feature enhancement network based on YOLOv5 for small object detection (DOI: 10.1016/j.jvcir.2023.103752)

Academic fraud report · Geng Detector

Summary

This report presents a tiered integrity review of the paper 'FE-YOLOv5: Feature enhancement network based on YOLOv5 for small object detection' published in Journal of Visual Communication and Image Representation (2023). The overall verdict is SUSPICIOUS (amber). The principal concern lies in the abstract, which claims that mAP was improved by 2.8% and 2.9% over YOLOv5 on two datasets. Verification against the reported tables shows that the 2.8% figure corresponds to the overall mAP gain on VisDrone2019 (21.0 vs 18.2), but the 2.9% figure does not represent the overall mAP on Tsinghua-Tencent100K (which is only 1.4%, i.e., 63.6 vs 62.2); instead, it corresponds to the AP_Small metric (52.8 vs 49.9). By juxtaposing the overall mAP of one dataset with the small-object AP of another without clarification, the abstract is misleading. Additional minor issues include a missing citation bracket for GAM in Section 3.1. Computational overhead analysis and ablation results appear internally consistent. No image-based duplication or fabrication could be assessed due to lack of original figures.

Verdict

🟡 Suspicious (Queriable) — The paper's experimental data are largely internally consistent and the methodological description is sound, but the abstract contains a misleading juxtaposition of metrics between two datasets that warrants clarification or correction. No evidence of outright data fabrication was identified from the textual and tabular content reviewed.

Key findings

Abstract misrepresents metric scope across datasets: The 2.8% mAP improvement (VisDrone2019) is a full-dataset metric, whereas the 2.9% figure (Tsinghua-Tencent100K) is the AP_Small value, not the overall mAP. The true overall mAP gain on Tsinghua-Tencent100K is only 1.4% (63.6 vs 62.2 in Table 2).
Missing citation for GAM: In Section 3.1 (Feature enhancement module), the Global Attention Mechanism (GAM) is introduced with an empty citation bracket [], though the work is presumably listed as reference [26].
Parameter / FLOPs overhead is consistent: FE-YOLOv5 has 9.14M parameters and 31.0G FLOPs vs YOLOv5-S at 7.03M and 15.9G, roughly doubling the compute cost. The reported accuracy gains are proportionate to this overhead, with no implausible jumps.
Ablation tables (Tables 3–5) are internally consistent: incremental additions of the proposed components produce monotonic improvements in line with expectations.

Evidence highlights

Abstract: *"Compared to YOLOv5, the [mAP] was improved by 2.8% and 2.9%, respectively."*
Table 1 (VisDrone2019): YOLOv5 mAP = 18.2; FE-YOLOv5 mAP = 21.0 → Δ = 2.8% (overall mAP, consistent).
Table 2 (Tsinghua-Tencent100K): YOLOv5 mAP = 62.2; FE-YOLOv5 mAP = 63.6 → Δ = 1.4% (overall mAP, inconsistent with abstract's 2.9%).
Table 2 (Tsinghua-Tencent100K): YOLOv5 AP_Small = 49.9; FE-YOLOv5 AP_Small = 52.8 → Δ = 2.9% (matches abstract's second figure, but is the small-object metric, not overall mAP).
Section 3.1: ...global attention mechanism (GAM) [], followed by 1×1 convolution... — citation bracket empty.
Table 5: Parameters 9.14M vs 7.03M; FLOPs 31.0G vs 15.9G.

Notes

The misleading wording in the abstract is classified as a writing/ethics issue rather than outright data fabrication, since the supporting tables reconcile the figures when correctly interpreted.
Detection limitation: The review was conducted on text and OCR-extracted tables only. Original images (network diagrams, detection visualisations in Figs. 1–5, 7–8) were not available, so first- and third-form image reuse / splicing checks could not be performed. Any future image-based assessment would require access to the source PDF figures at native resolution.
A formal erratum or abstract revision clarifying that the 2.9% gain refers to AP_Small on Tsinghua-Tencent100K is recommended.
DOI: 10.1016/j.jvcir.2023.103752