MCGA-Net: Multi-modal Chain & Global Attention

• The paper introduces MCGA-Net, a framework that integrates DCGAN-based augmentation, multi-modal chain feature fusion, and global attention for robust GPR defect detection.
• The architecture features a five-stage YOLOv8-style pipeline with COCO transfer learning that enhances precision, recall, and mAP in complex subsurface environments.
• MCGA-Net effectively addresses challenges like small-target detection and noise interference, outperforming traditional methods in GPR imaging.

The Multi-modal Chain and Global Attention Network (MCGA-Net) is a neural network framework designed for automated recognition of ground-penetrating radar (GPR) road defect images via hierarchical multi-scale feature fusion and global context-aware attention. MCGA-Net integrates DCGAN-based adversarial data augmentation, a novel Multi-modal Chain Feature Fusion (MCFF) module, a Global Attention Mechanism (GAM), and transfer learning from MS COCO, all within a YOLOv8-style detection backbone. The architecture achieves state-of-the-art performance in precision, recall, and mean average precision (mAP) for hidden defect detection in complex subsurface environments, demonstrating robustness to noise and small-target scenarios (Lv et al., 25 Dec 2025).

1. Architectural Composition and Signal Flow

MCGA-Net’s architecture consists of a five-stage pipeline designed to address key data and feature representation limitations in GPR defect detection:

1. Input: Single-channel B-scan GPR image of shape $256\times256\times1$.
2. Stage I (DCGAN-based Augmentation): High-fidelity synthetic GPR images are generated to augment the training set, mitigating data scarcity and maintaining morphological integrity.
3. Stage II (MCFF Module): Inserted at the 21st layer in the YOLOv8 neck, MCFF fuses hierarchical multi-scale features by treating feature maps as tensors contracted along all three modes with learned weights, enhancing cross-dimensional interactions.
4. Stage III (GAM Module): Placed at the 3rd block of the backbone, GAM enhances feature maps by sequentially applying MLP for channel attention and large-kernel convolution for spatial attention, preserving channel–spatial correlations.
5. Stage IV (YOLOv8 Detection Head): Standard components consist of CSP-Darknet-style convolutional blocks, SPPF for receptive field expansion, C2f layers for further bottlenecking, and detection heads for bounding boxes, class scores, and DFL.
6. Stage V (MS COCO Transfer Learning): Pretraining on the COCO dataset, followed by fine-tuning on augmented GPR images, ensures efficient convergence and improved out-of-domain generalization.

The forward pass propagates input through backbone convolutional blocks, with the GAM module inserted at an early stage, followed by up-sampling and concatenation in the neck, with MCFF applied before detection heads aggregate fused feature representations. Key tensor locations include $f_3$ ([B,256,32,32]) for GAM and $c_2$ ([B,768,32,32]) for MCFF application.

2. Multi-modal Chain Feature Fusion (MCFF)

MCFF is formulated to overcome limitations in conventional feature fusion (such as simple addition or concatenation), which may neglect complex cross-dimensional dependencies intrinsic to GPR defect morphology. Each input feature map tensor $X\in\mathbb{R}^{I\times J\times K}$ is processed via chained Einstein (mode-$n$) products with learned matrices:

$Y = X\times_1 W_1 \times_2 W_2 \times_3 W_3$

where $W_1\in\mathbb{R}^{I\times R_1}$, $W_2\in\mathbb{R}^{J\times R_2}$, $W_3\in\mathbb{R}^{K\times R_3}$. In expanded form:

$$Y_{pqr} = \sum_{i=1}^I\sum_{j=1}^J\sum_{k=1}^K X_{ijk} (W_1)_{ip}(W_2)_{jq}(W_3)_{kr}$$

For cases where $R_1=I, R_2=J, R_3=K,$ a residual connection preserves coarse-grained context:

$Y = X + X\times_1 W_1 \times_2 W_2 \times_3 W_3$

MCFF is applied after the second up-sampling in the YOLOv8 neck, ensuring that both shallow and deep feature representations contribute to final detection. This addresses scale- and morphology-variant defect structures in GPR imaging (Lv et al., 25 Dec 2025).

3. Global Attention Mechanism (GAM)

The GAM module is introduced to address the insufficiency of conventional Squeeze-and-Excitation or CBAM blocks, which may ignore fine-grained channel-spatial interactions. GAM preserves the 3D structure $F\in\mathbb{R}^{C\times H\times W}$, applying:

• Channel Attention: Permute $F$ to $F'\in\mathbb{R}^{H\times W\times C}$, unfold into vectors, then apply a two-layer MLP with reduction:

$Z = \sigma(W_2\,\delta(W_1\,\text{vec}(F')))$

with $$W_1\in\mathbb{R}^{(C/r)\times C},\ W_2\in\mathbb{R}^{C\times (C/r)}$$, $\delta(\cdot)$ = ReLU, $\sigma$ = Sigmoid. Output is reshaped back to $\mathbb{R}^{C\times H\times W}$.

• Spatial Attention: Sequential $7\times 7$ convolutions yield a spatial map:

$$M_s(F) = \sigma(\text{Conv}_{7\times7}(\text{ReLU}(\text{Conv}_{7\times7}(F))))$$

• Refinement: Combined channel and spatial maps modulate features multiplicatively:

$F' = F\odot M_c(F)\odot M_s(F)$

This process enhances channel–spatial relationships critical for accurate subsurface defect localization.

4. Adversarial Data Augmentation via DCGAN

A deep convolutional GAN framework (DCGAN) is utilized to synthesize realistic GPR images, addressing data scarcity and maintaining defect morphology:

• Generator $G$: Projects $z\in\mathbb{R}^{100}$ to $256\times256\times1$ via a sequence of deconvolutional layers, batch normalization, and Tanh activation.
• Discriminator $D$: Processes $256\times256\times1$ images through convolution, LeakyReLU, global average pooling, and sigmoid layer.

The adversarial loss functions with one-sided label smoothing are: $$L_D = -\mathbb{E}_{x\sim p_{\text{data}}}[\log D(x)] - \mathbb{E}_{z\sim p_z}[\log(1-D(G(z)))]$$

$L_G = -\mathbb{E}_{z\sim p_z}[\log D(G(z))]$

On-the-fly horizontal flips and Gaussian noise ($\sigma=0.05$) are applied only to discriminator inputs for regularization, preserving integrity in generated morphologies such as hyperbolas and high-amplitude echo bands.

5. Transfer Learning and Training Strategy

Transfer learning is implemented by pretraining MCGA-Net (modified for single-channel input) on MS COCO for 50 epochs ($\text{lr}=0.01$, Adam, batch size 16), then fine-tuning with the augmented GPR dataset (GPR-Aug) for 160 epochs, using a learning-rate schedule of $0.01\rightarrow 0.001$ and batch size 8. Preprocessing includes ISDFT-based interference suppression, low-pass filtering, and background subtraction. Loss comprises box_loss (Complete-IoU), cls_loss (binary cross-entropy for $K=3$ classes), and dfl_loss (distribution focal loss). Augmentations in detection comprise random flips, small color jitter, and mosaic composition.

6. Quantitative Performance and Empirical Evaluation

Empirical results demonstrate notable gains attributed to each MCGA-Net component:

Model/Step	Precision (%)	Recall (%)	mAP@50 (%)
YOLOv8 (GPR-Ori)	75.4	76.9	80.9
YOLOv8 + DCGAN (GPR-Aug)	89.1	89.6	92.1
MCGA-Net (GPR-Aug)	92.8	92.5	95.9
MCGA-Net + COCO	92.8	94.0	96.7

In robustness evaluations under Gaussian noise ($\sigma=25$), YOLOv8’s mean confidence drops by ∼10%, whereas MCGA-Net’s confidence remains stable or increases by 2–3%. MCGA-Net outperforms Faster R-CNN, SSD, YOLOv3-tiny, YOLOv5, YOLOv6, and YOLOv8 in both precision and mAP on the test set. For small-target and low-amplitude object detection, MCGA-Net increases crack/cavity confidence by approximately 10% relative to YOLOv8, evidencing improved sensitivity to weak signals.

Ablation results indicate sequential improvements contributed by each stage:

• +DCGAN: +13.7 P, +15.6 R, +13.9 mAP
• +MCFF: +3.5 P, +0.9 mAP
• +GAM: +3.3 P
• +COCO pretrain: +1.2 R, +0.8 mAP Final: P=92.8%, R=94.0%, mAP@50=96.7% (Lv et al., 25 Dec 2025).

7. Implementation Specifics and Data Flow

The pseudocode for MCGA-Net’s forward pass details the placement of MCFF and GAM within the YOLOv8-style backbone and neck. GAM is applied at $f_3$ ([B,256,32,32]); MCFF at $c_2$ ([B,768,32,32]). Three detection heads operate at progressively different scales to support detection from large, medium, to small targets. Residual connections within MCFF preserve original context, and feature refinement via GAM ensures channel-spatial coupling is maintained.

The architecture enables integration of adversarial augmentation, chain-based multi-mode fusion, and context-aware channel–spatial weighting, establishing a new state of the art for GPR-based road defect detection by balancing computational efficiency, accuracy, and robustness (Lv et al., 25 Dec 2025).