Fraud investigation report: FE-YOLOv5 (Journal of Visual Communication and Image Representation, 2023)

Academic fraud report · Geng Detector

Summary

This report assesses "FE-YOLOv5: Feature enhancement network based on YOLOv5 for small object detection" by Wang et al. (DOI: 10.1016/j.jvcir.2023.103752, J. Vis. Commun. Image R., 2023). The verdict is "highly suspicious" (academic beautification suspected). The central issue is selective reporting in the abstract: the paper claims YOLOv5 improvements of 2.8% and 2.9% on VisDrone2019 and Tsinghua-Tencent100K respectively, but examination of Tables 1 and 2 shows the 2.8% corresponds to overall AP on VisDrone while the 2.9% actually corresponds to small-object APs on Tsinghua-Tencent (overall AP gain on that dataset is only 1.4%). The two metrics are presented as if they were the same measure. A second concern is that ablation values add perfectly (1.3 + 1.5 = 2.8). All tables report point estimates only, with no standard deviations or significance tests. Image-level analysis was not possible from the text. Confidence is moderate; findings rely on tabular numbers and could be confirmed by verifying the original PDF.

Verdict

Highly suspicious. The paper contains a clear case of selective/misleading reporting in the abstract and presents an implausibly clean ablation pattern. Not proven fabrication, but consistent with academic "beautification."

Key findings

Selective metric reporting in the abstract: The 2.8% and 2.9% gains attributed to VisDrone2019 and Tsinghua-Tencent100K mix two different metrics (overall AP vs. small-object APs).
Implausibly perfect ablation additivity: Individual module gains (FEM +1.3, SAM +1.5) sum exactly to the combined gain (2.8) over the YOLOv5 baseline (AP 18.2 → 21.0).
No uncertainty estimates: Tables 1–4 report only single point estimates; no standard deviation, confidence intervals, or significance tests are provided, despite gains as small as 1.4%.
Reference list anomalies: Authors' names are missing from the parsed reference entries (e.g., "[1] accurate object detection...", "[2] Computer Vision..."), possibly a PDF parsing artifact but worth verification.
Image figures not analyzable: Figures 4, 7, and 8 could not be inspected at pixel level from the available text.

Evidence highlights

Abstract: "Compared to YOLOv5, the … was improved by 2.8% and 2.9%, respectively."
Table 1 (VisDrone2019): FE-YOLOv5 AP = 21.0, YOLOv5 AP = 18.2 → 21.0 − 18.2 = 2.8% (overall AP).
Table 2 (Tsinghua-Tencent100K): FE-YOLOv5 AP = 63.6, YOLOv5 AP = 62.2 → 1.4% (overall AP, not 2.9%). APs for FE-YOLOv5 = 52.8, YOLOv5 = 49.9 → 52.8 − 49.9 = 2.9% (small-object APs).
Table 3 (ablation): Baseline 18.2, +FEM 19.5 (+1.3), +SAM 19.7 (+1.5), +FEM+SAM 21.0 (+2.8). 1.3 + 1.5 = 2.8 exactly.
DOI: 10.1016/j.jvcir.2023.103752.

Notes

The VisDrone AP arithmetic confirms the 2.8% claim; the Tsinghua-Tencent claim does not match overall AP and only matches APs.
Perfect additivity in ablations is statistically uncommon in deep learning but not impossible; it warrants request for raw multi-run logs.
Authors should be asked for variance estimates on both datasets and clarification of which metric is reported in the abstract.
Findings are limited to what is visible in the parsed text; original PDF should be inspected for confirmation.