Resource-Efficient Helmet Detection Using Structured L1 Channel Pruning of YOLOv11n

2.1 You Only Look Once variants for object detection

The YOLO family has undergone significant evolution since its introduction. YOLOv1 started the one-stage detection paradigm with direct prediction of bounding boxes and class probabilities [13]. Anchor-based mechanisms in YOLOv3 improved the accuracy of small object detection [14]. YOLOv4 added the CSPDarknet53 backbone together with various augmentation techniques [15]. The next generation of YOLO models focused on efficiency. YOLOv5 introduced nano to extra-large variants for different computational budgets [16]. YOLOv7 used efficient layer aggregation networks [17]. YOLOv8 introduced anchor-free detection with a decoupled head design [6].

YOLOv11 includes C2PSA modules and C3K2 blocks for more efficient multi-scale feature capture [6]. Lightweight variants have emerged for edge deployment. YOLOv5-nano achieves a 1.9 MB model size with 45% mAP on COCO [18]. YOLOv8-nano improves performance to 37.3% mAP with a 3.2 MB size [19]. YOLOv11n continues this trend, targeting devices with limited RAM and processing power.

2.2 Helmet detection systems

Helmet detection has been explored using various YOLO versions. YOLOv5-SN achieves 31 frames per second (FPS) on the Jetson Nano using a ShuffleNetV2 backbone with 90% parameter reduction [20]. SG-YOLOv5 maintains baseline mAP while achieving a 90.8% parameter reduction and a 2.7x FPS improvement [21]. YOLOv8 variants have also been applied to this task. The AI City Challenge Track 5 uses YOLOv8 for multi-class helmet violation detection [22], and YOLOv8-ADown-LSCD achieves 0.85 mAP@0.5 on a custom dataset [23]. The Kaggle Helmet Detection dataset, containing 764 images, is frequently used as a research benchmark [24]. Dataset differences limit direct comparison across these studies, and generalization across varying camera conditions remains an open challenge.

2.3 Neural network pruning techniques

Structured pruning removes entire channels or filters, producing hardware-friendly sparse models without requiring specialized sparse computation libraries. Filter pruning with an L1-norm criterion has proven effective for practical deployment [25]. Li et al. [7] introduced a magnitude-based approach in which filters with small L1-norm weights are removed.

Batch normalization scaling factors serve as channel importance indicators. Network Slimming adds an L1 penalty to batch normalization (BN) gamma parameters during training [26]. NS-YOLO uses BN-based pruning in YOLOv5, achieving an 88.19% parameter reduction with only a 1.6% mAP drop [12]. A-pruning uses adaptive BN evaluation to select optimal subnetworks, achieving a 75.8% parameter reduction with a 0.9% accuracy drop on YOLOv5 [27]. Target Capacity Pruning designs compression to meet specific inference time constraints; on YOLOv5, a 30% pruning ratio loses 0.6% mAP, and a 50% ratio loses 2.9% [28]. SAFP-YOLO uses spatial attention-based filter pruning on YOLOv7, achieving substantial inference speed gains on embedded hardware [29].

CM-YOLOv8 applies L1-norm channel pruning for coal mine detection, achieving a 40% model volume reduction with less than 1% accuracy drop [30]. YOLOv8-ALWP uses layer-adaptive magnitude-based pruning, achieving a 64.67% parameter reduction together with a 1.7% accuracy improvement [31]. Compression techniques beyond pruning include knowledge distillation [32], quantization [33], and hybrid approaches that combine multiple compression strategies for deployment on platforms such as the Jetson Xavier NX [34].

2.4 Research gap and positioning

Existing work concentrates on YOLOv5, YOLOv7, and YOLOv8 variants. YOLOv11n, despite its architectural advantages, has not been studied for helmet detection, and no prior work applies structured pruning to this architecture. The C3K2 and C2PSA modules introduced in YOLOv11 create dimensional constraints at Concat junctions that prevent the direct application of existing global pruning pipelines. This gap motivates a targeted inner bottleneck pruning strategy that operates within each Bottleneck block, preserving outer channel dimensions while reducing internal redundancy. The present work addresses this gap by providing the first structured L1 channel pruning study on YOLOv11n, validated through cross-validation and statistical testing, and benchmarked against YOLOv8n under identical experimental conditions.

4.1 Overall performance

Table 2 reports the 5-fold cross-validation results for all three model configurations, together with YOLOv8n evaluated on identical hardware and dataset as a same-hardware baseline. All metrics are reported as the mean plus or minus one standard deviation across five folds, except for YOLOv8n, which was evaluated on the original single split for direct comparison.

Figure 4. Mean mAP50 with standard deviation across five folds for baseline, pruned_30, and pruned_50

Figure 5. Mean frames per second (FPS) with standard deviation across five folds

Figure 6. Mean Floating Point Operations (FLOPs) with standard deviation across five folds

Figure 7. Mean Open Neural Network Exchange (ONNX) model size with standard deviation across five folds

Table 2. Five-fold cross-validation results for baseline and pruned YOLOv11n configurations, with same-hardware YOLOv8n comparison

mAP50 remained stable across both pruning ratios. The 30% pruned model produced a slightly higher mAP50 than the baseline (0.7962 versus 0.7889, a delta of +0.0073), which reflects a regularization effect from fine-tuning rather than a genuine accuracy gain. The 50% pruned model was nearly identical to the baseline (0.7882, a delta of -0.0006). Both differences are small in absolute terms and are discussed further in Section 4.6. FPS increased from 203.49 ± 40.23 for the baseline to 218.17 ± 46.27 for pruned_30 (+7.2%) and to 238.42 ± 11.09 for pruned_50 (+17.2%). The narrower FPS standard deviation for pruned_50 (11.09, compared with 40.23 for the baseline) suggests more consistent throughput after aggressive pruning. Figures 4 through 7 show the corresponding per-metric bar charts with error bars across folds.

4.2 Model compression analysis

Table 3 reports compression statistics relative to the baseline, using ONNX export size as the deployment-relevant measure.

Table 3. Compression statistics relative to baseline, based on ONNX export size

The modest compression ratios, 1.052× and 1.089×, reflect the inner bottleneck scope of pruning. Only the cv1-cv2 channel pairs inside each Bottleneck block are modified, while the outer projection channels and all non-Bottleneck layers remain unchanged. This conservative scope is intentional: it avoids dimensional conflicts at Concat operations in the Feature Pyramid Network-Path Aggregation Network (FPN-PAN) structure and maintains architectural stability. As a consequence, the parameter and size reductions are smaller than those reported by methods that prune global channels or entire layers. The FLOPs reduction, 4.00% and 6.49%, follows a similar pattern, since the remaining backbone and neck layers continue to compute at full channel width.

4.3 Layer sensitivity analysis

Layer sensitivity was measured by temporarily masking 30% of channels in each eligible layer independently and recording the resulting proxy mAP50 drop without fine-tuning. Figure 8 shows the top 20 most sensitive layers. Table 4 lists the five layers with the highest sensitivity.

Figure 8. Top 20 layer sensitivity scores, where a larger drop indicates higher sensitivity to 30% channel masking

Table 4. Top-5 most sensitive layers to 30% channel masking (no fine-tuning)

The five most sensitive layers are concentrated in the early backbone stages, specifically layer indices 2 through 10. Masking 30% of channels in model.4.cv1.conv (layer index 7) reduced proxy mAP50 from 0.8210 to 0.0176, a drop of 0.8034, making it the most critical layer by a substantial margin. model.3.conv (layer 6) and model.2.cv1.conv (layer 2) followed, with drops of 0.7180 and 0.5059, respectively. These early layers encode low-level texture and edge features that are fundamental to the subsequent feature pyramid. Their high sensitivity confirms that the inner bottleneck pruning strategy, which distributes channel removal across all Bottleneck blocks rather than concentrating it in any single sensitive layer, is an appropriate design choice for preserving post-pruning recoverability.

4.4 Pruning criteria comparison

To evaluate the choice of the L1 norm as the pruning criterion, a proxy ablation compared L1, L2, and random channel selection at a ratio of 0.30, without fine-tuning. All three criteria produced the same parameter count and FLOPs, since the number of removed channels was identical and only the selection policy differed. Table 5 and Figure 9 report the resulting proxy mAP50 drops.

Table 5. Pruning criteria comparison at ratio = 0.30, proxy evaluation without fine-tuning

Figure 9. Proxy mAP50 drop before fine-tuning for L1, L2, and random pruning criteria at 30% ratio

The proxy evaluation showed that L1 produced the largest pre-fine-tuning mAP50 drop (0.1115), followed by Random (0.1105) and L2 (0.0893). The L1 criterion selects the channels with the smallest absolute weight magnitudes, which are the channels that the network currently uses least. After fine-tuning, all three criteria converged to similar final accuracy, with a delta from baseline of less than 0.01 in each case, consistent with the interpretation that magnitude-based pruning produces models that are more amenable to recovery, since the removed channels were already contributing little to the representation. L1 was selected over L2 for two reasons: its proxy inference speed was the highest among the three criteria (277.00 FPS, versus 241.63 for L2 and 244.32 for Random), and it is computationally simpler to calculate. The post-fine-tuning accuracy difference across criteria was negligible.

4.5 Precision-recall curve and F1-score analysis

Precision-recall curves and F1-score curves are reported for fold 0 across all three model configurations, which is representative of the cross-validation distribution. Figures 10 through 15 show the PR and F1 curves for the baseline, pruned_30, and pruned_50 models.

Figure 10. Precision-recall curve, baseline model (fold 0)

Figure 11. Precision-recall curve, pruned_30 model (fold 0)

Figure 12. Precision-recall curve, pruned_50 model (fold 0)

Figure 13. F1-score curve, baseline model (fold 0)

Figure 14. F1-score curve, pruned_30 model (fold 0)

Figure 15. F1-score curve, pruned_50 model (fold 0)

The precision-recall curves show that the pruned models maintained an area under the curve close to that of the baseline. The curve shapes were consistent across the three configurations, with no visible collapse in the high-recall region. Precision at recall = 1.0 was slightly lower for the pruned models, indicating a marginal increase in false positives at the extreme of the recall range. The F1-score curves confirmed that the optimal confidence threshold shifted only slightly after pruning, and the peak F1 value remained comparable across configurations. For helmet detection, where missing a violation is more costly than generating a false alarm, this slight precision-recall trade-off introduced by pruning is acceptable.

4.6 Statistical significance

Table 6 reports the results of the paired t-test and Wilcoxon signed-rank test for mAP50 and FPS comparisons between the baseline and each pruned model, using the 5 folds results as matched pairs.

Table 6. Statistical significance tests for mAP50 and frames per second (FPS) (n = 5 folds, alpha = 0.05)

No comparison reached statistical significance at alpha = 0.05. For mAP50, p = 0.42 for pruned_30 and p = 0.95 for pruned_50, confirming that pruning did not significantly degrade accuracy at either ratio. The FPS differences were also not statistically significant (p = 0.53 for pruned_30 and p = 0.10 for pruned_50). This indicates that the observed throughput increases should be interpreted with caution given the variance across folds; the n = 5 sample size limits statistical power, and the p = 0.10 result for pruned_50 is a borderline finding that warrants evaluation on additional data. The absence of significance for mAP50 is the primary result of interest, as it provides statistical evidence that the pruned models performed equivalently to the baseline on this dataset.

4.7 Comparison with YOLOv8n on identical hardware

To provide a fair same-hardware baseline, YOLOv8n was trained and evaluated on the same dataset and hardware configuration, NVIDIA T4, used for the revision experiments. Table 7 compares YOLOv8n against the three YOLOv11n configurations.

The pruned_50 YOLOv11n achieved an mAP50 of 0.7882, exceeding YOLOv8n at 0.7733, while using 21% fewer parameters (2.378M versus 3.011M), 20.5% less ONNX storage (9.30 MB versus 11.70 MB), and 26.6% fewer FLOPs (5.97 G versus 8.14 G). FPS for pruned_50 (238.42) also exceeded YOLOv8n (211.67). This comparison establishes that pruning YOLOv11n produces a model that is simultaneously more accurate, faster, and smaller than the previous-generation architecture trained under equivalent conditions.

Table 7. Direct comparison of YOLOv11n configurations with YOLOv8n on NVIDIA T4 hardware and identical dataset

4.8 Detection quality analysis

The confusion matrices for fold 0 across the three configurations are shown in Figures 16 through 18. The pattern across configurations is consistent with the quantitative mAP50 results reported in Table 2. The baseline model correctly classified the majority of with_helmet and without_helmet instances, with a small proportion of false negatives concentrated in occluded or small-scale instances. The pruned_30 model showed a comparable classification pattern, with no systematic shift in the false positive or false negative distribution relative to the baseline. The pruned_50 model similarly maintained a confusion structure close to the baseline, consistent with the statistically negligible mAP50 delta of -0.0006 (p = 0.95) reported in Table 6. These observations confirm that channel removal at either ratio did not introduce a systematic bias toward either class.

Figure 16. Confusion matrix model baseline (fold 0)

Figure 17. Confusion matrix model pruned 30% (fold 0)

Figure 18. Confusion matrix model pruned 50% (fold 0)

Figure 19. Detection results sample for baseline (fold 0)