Integrity Review Report: YOLOv5 Lightweight Traffic Sign Detection Method Based on SWimAM Design

Academic fraud report · Geng Detector

Summary

This report evaluates a 2024 paper published in Science Technology and Engineering (DOI: 10.12404/j.issn.1671-1815.2400927) that proposes a lightweight YOLOv5 variant with a custom SWimAM attention module for traffic sign detection. The overall verdict is highly suspicious. The most serious finding is that the proposed model's average precision (AP) for the 'indication sign' class is reported as 0.902 on both the CCTSDB (China) and GTSDB (Germany) datasets, an implausible coincidence given that the two datasets have completely different backgrounds, lighting, and sample distributions. Additional concerns include (1) a naming inconsistency where the final model is called C32SWimAM-SE-SIoU in the text but reverts to C32SimAM-SE-SIoU in the generalization tables, (2) mismatched FPS values for the same model across tables (43.3 vs. 43.63) and incorrect percentage claims in the text, (3) a misidentified 'parameter-free' learnable weight tensor in the SWimAM module, and (4) a misattributed SIoU loss formula that lacks the defining angle, distance, and shape cost components. Limitations: no image-based forensic analysis was possible, and the conclusions rely solely on textual and numerical evidence.

Verdict

Highly suspicious. Multiple independent numerical and conceptual inconsistencies were verified, including an implausibly identical AP value (0.902) for the 'indication sign' class across two unrelated national datasets, model-name substitution errors in the generalization experiments, conflicting FPS figures for the same model, and a formula misattributed to SIoU. These patterns are consistent with data fabrication, careless reuse of prior experimental tables, and superficial concept handling.

Key findings

Cross-dataset AP duplication (critical): In Table 4 (CCTSDB) and Table 5 (GTSDB), the 'indication sign' AP for the model C32SimAM-SE-SIoU is reported as 0.902 in both tables, identical to three decimal places across datasets with different imaging conditions.
Model-name swap in generalization tables: The text states the C32SWimAM-SE-SIoU model is used for transfer learning on CCTSDB and GTSDB, but Tables 4 and 5 are labelled with C32SimAM-SE-SIoU, indicating an incomplete replacement of an earlier model designation.
Conflicting FPS and incorrect percentage claims: The same final model reports FPS of 43.3 in Table 2 and 43.63 in Table 3. The text claim of a 7.33% speed improvement does not reconcile with Table 2 (computed value 6.51%) and only matches when Table 3's value is used.
Incorrect arithmetic on cross-model comparison: The paper claims an 11.34% speed improvement over YOLOv8s, but using YOLOv8s=38.89 and the model=43.63 gives 12.18%.
Conceptual misuse of 'parameter-free': The SWimAM module is described as both adding learnable weights (self.weight1) and retaining a 'parameter-free' character, a contradiction under standard PyTorch semantics where such a tensor contributes trainable parameters.
Misattributed SIoU formula: Equation 8 reportedly represents SIoU but lacks the angle, distance, and shape-cost components that define SIoU (Gevorgyan, 2022), instead resembling a generic IoU/Smooth L1 composition.

Evidence highlights

Table 4 (CCTSDB), 'indication sign' AP = 0.902; Table 5 (GTSDB), 'indication sign' AP = 0.902.
Table 2: YOLOv5s FPS = 40.65; final model FPS = 43.3. Computed improvement: (43.3 − 40.65) / 40.65 = 6.51%, not 7.33%.
Table 3: final model FPS = 43.63 (same model as Table 2, different value).
Table 3 vs. Table 2: YOLOv8s FPS = 38.89; final model FPS = 43.63. Computed improvement: (43.63 − 38.89) / 38.89 = 12.18%, not 11.34%.
References [19] and [20] are cited for SIoU, but Equation 8 does not contain the SIoU angle-cost term.
DOI: 10.12404/j.issn.1671-1815.2400927.

Notes

Verification is limited to textual, tabular, and mathematical content; no image-based forensic analysis (e.g., Western blots, microscopy) was applicable to this study.
The AP duplication finding carries the highest weight because three-decimal agreement across independent datasets is statistically very unlikely under genuine experimental conditions.
Findings 2–6 are individually indicative of carelessness or superficial drafting, and collectively strengthen the plausibility of Finding 1.
Final determination of misconduct requires investigation by the journal editorial office and the authors' institution, with access to raw data, training logs, and code.