YOLOv11-KW-TA-FP: Real-time Crack Detection

• The paper presents YOLOv11-KW-TA-FP, a multitask model that integrates dynamic KernelWarehouse convolution, triple attention, and FP-IoU loss to enhance crack detection and segmentation.
• It replaces standard convolutions with KWConv modules and applies a Triple Attention block in the feature pyramid, significantly improving small-target crack sensitivity in complex backgrounds.
• Experimental results on the Crack-Seg dataset show improved performance over YOLOv11n, with detection precision at 91.3% and segmentation mAP@50 rising from 65.4% to 76.3%.

YOLOv11-KW-TA-FP is a real-time, multi-task concrete crack detection and segmentation model built on the lightweight YOLOv11n backbone. It was proposed to address inefficient manual inspection and the suboptimal performance of existing deep learning models, particularly for small-target crack detection within complex backgrounds. The model combines three modifications—dynamic KernelWarehouse convolution in the backbone, Triple Attention in the feature pyramid, and an FP-IoU loss in the detection head—and adds a parallel mask-prediction branch for crack segmentation. On the Crack-Seg test set, it reports 91.3% precision, 76.6% recall, and 86.4% mAP@50 for detection, while segmentation mAP@50 rises from 65.4% in YOLOv11n to 76.3% (Huang et al., 15 Aug 2025).

1. Architectural definition and task scope

The model is defined as a real-time concrete crack detection and segmentation system derived from YOLOv11n. Relative to vanilla YOLOv11n, standard convolutions in the backbone are replaced by KWConv modules, each up-sampling operation in the neck is followed by a TA block, and the detection head uses FP-IoU instead of CIoU. A parallel mask-prediction branch is attached at the head for crack segmentation (Huang et al., 15 Aug 2025).

Its design target is explicitly multi-task: bounding-box detection and binary crack segmentation are performed within a single architecture. The modifications are placed so that feature extraction, cross-scale fusion, and localization loss are each altered by a dedicated mechanism. This suggests that the system is organized as a coordinated optimization of representation, attention, and regression rather than as a single-module patch over YOLOv11n.

All reported experiments use the Crack-Seg dataset with a 7:2:1 split, 200 training epochs, SGD, batch size 16, initial learning rate $lr_0=0.01$, momentum $=0.937$, and image size $640\times 640$ (Huang et al., 15 Aug 2025).

2. Backbone modification through dynamic KernelWarehouse convolution

The backbone extension embeds Intel’s KernelWarehouse dynamic convolution into each backbone layer. The mechanism is described in three steps: kernel partitioning, cross-layer warehouse sharing, and normalized attention fusion. A standard $k\times k$ kernel is uniformly split along the channel axis into $m$ equal units, and a shared warehouse stores $n\gg m$ such units without linear parameter growth. Layers within a stage reuse the same warehouse $\mathcal{E}=\{E_j\}_{j=1}^n$, and each layer dynamically assembles its effective kernel from a weighted combination of units (Huang et al., 15 Aug 2025).

The normalized attention fusion weights are defined as

$$a_j=\frac{z_j}{\sum_{i=1}^n |z_i|}, \qquad j=1,\dots,n,$$

where $\{z_j\}$ are attention scores produced by a lightweight selector network. The formulation allows negative weights for adversarial interactions and ensures $\sum_j a_j=1$. Early in training, a temperature $=0.937$0 linearly anneals from 1 to 0 over 20 epochs to stabilize mixing. The dynamic convolution output at location $=0.937$1 is

$=0.937$2

The reported functional effect is richer multi-scale feature representation with negligible parameter overhead. The paper attributes particular benefit to small-crack sensitivity under complex backgrounds, while maintaining a parameter cost similar to standard convolution. In the discussion, the module is further described as mixing hundreds of fine-grained kernel units under a fixed parameter budget and markedly improving small-crack recall (Huang et al., 15 Aug 2025).

3. Triple Attention in the neck and the segmentation branch

Within the feature pyramid network, each up-sample is followed by a Triple Attention block composed of three parallel branches: channel attention, spatial attention, and long-range attention. For an input feature map $=0.937$3, the channel branch computes

$=0.937$4

with an MLP bottleneck of size $=0.937$5. This branch re-weights channels to emphasize crack-relevant feature maps (Huang et al., 15 Aug 2025).

The spatial branch pools across channels, concatenates the average and max maps, and applies a $=0.937$6 convolution plus sigmoid:

$=0.937$7

$=0.937$8

The long-range branch views spatial dimensions as a sequence and applies an LSTM-inspired unit:

$=0.937$9

This branch is stated to capture spatial dependencies over large receptive fields up to 100 px, effectively linking distant crack segments. The three branches are fused by element-wise sum or gated combination, and the resulting attention map is multiplied back onto $640\times 640$0. The stated effect is stronger channel–spatial and long-range interaction modeling, suppression of background clutter, and highlighting of faint micro-cracks (Huang et al., 15 Aug 2025).

The segmentation pathway is lightweight. A mask-prediction branch mirrors the detection head’s feature maps and uses a $640\times 640$1 convolution to predict a binary crack segmentation mask. To address extreme background/foreground imbalance, it uses weighted cross-entropy:

$640\times 640$2

with higher weight on crack pixels (Huang et al., 15 Aug 2025).

4. FP-IoU as the bounding-box regression objective

FP-IoU is introduced because standard IoU-based losses such as CIoU apply a monotonic penalty as IoU decreases, which can under-penalize poor-overlap boxes and converge slowly on low-quality samples. The loss integrates two components: a piecewise Focaler IoU mapping and Powerful-IoU v2 geometric penalization (Huang et al., 15 Aug 2025).

For Focaler IoU, the base IoU loss $640\times 640$3 is remapped piecewise as

$640\times 640$4

For PIoUv2, letting $640\times 640$5 denote ground-truth width and height and $640\times 640$6 denote edge offsets, the scale-adaptive penalty is

$640\times 640$7

with

$640\times 640$8

$640\times 640$9

The combined FP-IoU loss is

$k\times k$0

The stated consequences relative to CIoU are threefold: non-monotonic attention $k\times k$1 emphasizes medium-quality boxes, scale-aware penalties $k\times k$2 are applied for slender crack shapes, and stronger gradient signals are maintained when IoU is low. The paper associates these changes with accelerated convergence and improved localization on blurred or tiny cracks (Huang et al., 15 Aug 2025).

5. Detection, segmentation, and ablation results

The principal test-set results compare YOLOv11-KW-TA-FP with YOLOv11n under the same Crack-Seg protocol.

Model	Detection $k\times k$3	Segmentation mAP@50
YOLOv11n	87.1 / 73.3 / 79.2	65.4
YOLOv11-KW-TA-FP	91.3 / 76.6 / 86.4	76.3

The paper also reports rapid convergence of box, class, DFL, and mask losses, with stable precision, recall, and mAP curves after approximately 25 epochs (Huang et al., 15 Aug 2025).

The ablation study isolates the contributions of KWConv, TA, and FP-IoU. Starting from the baseline $k\times k$4 precision, $k\times k$5 recall, and $k\times k$6 mAP@50, adding KWConv yields $k\times k$7, $k\times k$8, and $k\times k$9; adding TA yields $m$0, $m$1, and $m$2; adding FP-IoU yields $m$3, $m$4, and $m$5. Pairwise combinations report $m$6 mAP@50 for KW+TA, $m$7 for KW+FP-IoU, and $m$8 for TA+FP-IoU. The full combination reaches $m$9 precision, $n\gg m$0 recall, and $n\gg m$1 mAP@50. The report states that all three modules contribute and that their synergy yields a total gain of 7.2 points in mAP@50 over baseline (Huang et al., 15 Aug 2025).

6. Robustness, nomenclature, and relation to YOLOv11

Robustness is evaluated under both data scarcity and corrupted training conditions. When the Crack-Seg training set is subsampled to 30%, 50%, 70%, 90%, and 100% of its original size, the reported $n\gg m$2 values are $n\gg m$3, $n\gg m$4, $n\gg m$5, $n\gg m$6, and $n\gg m$7, respectively. The paper states that performance plateaus after approximately 70%, which it interprets as evidence that KWConv+TA provides strong generalization from limited data. Under data augmentation with random rotation, flips, scaling, and Gaussian noise $n\gg m$8, segmentation robustness improves from $n\gg m$9 to $\mathcal{E}=\{E_j\}_{j=1}^n$0 across Precision, Recall, IoU, Dice, MDR, and FDR, respectively; IoU and Dice increase by 3.9% and 5.4%, while MDR and FDR decrease by 0.9% and 2.5% (Huang et al., 15 Aug 2025).

A recurrent source of confusion is nomenclature. In the YOLOv11 overview, the labels “KW,” “TA,” and “FP” are used in a different sense, referring respectively to C3k2, C2PSA, and SPPF as core YOLOv11 modules within the backbone and neck (Khanam et al., 2024). In YOLOv11-KW-TA-FP, by contrast, the same letter groups denote KernelWarehouse convolution, Triple Attention, and FP-IoU loss (Huang et al., 15 Aug 2025). The overlap is terminological rather than architectural identity. A plausible implication is that the model should be read as a YOLOv11n-derived crack-inspection specialization, not as a renaming of the generic YOLOv11 module set.

In the report’s concluding characterization, the combined system is a 5.83 M-parameter model running at $\mathcal{E}=\{E_j\}_{j=1}^n$1 FPS on RTX 4070Ti while maintaining 91.3% precision, 76.6% recall, and 86.4% mAP@50 for detection, together with improved segmentation metrics. Within the scope of the reported experiments, it is positioned as an efficient solution for automated infrastructure health monitoring and resource-constrained on-site deployment (Huang et al., 15 Aug 2025).