Center-Enhanced CAM Consistency Module

• The paper introduces CECCM, a module that refines CAMs through center-focused enhancement and consistency loss to improve pseudo-label fidelity in weakly supervised detection.
• It employs a method combining center of mass calculations with Gaussian weighting to align classifier and reconstructor activations for stable spatial predictions.
• Ablation studies show that integrating CECCM enhances detection metrics such as F1-score and IoU by stabilizing CAM activations and improving central localization in tasks like road crack detection.

The Center-Enhanced CAM Consistency Module (CECCM) is a specialized architectural element within weakly supervised image analysis pipelines, designed primarily to refine class activation maps (CAMs) for dense prediction tasks such as road crack detection. By applying center-focused enhancement and consistency constraints, CECCM addresses instability in CAM activations under adversarial or collaborative training regimes and improves the spatial reliability of pseudo-labels used for downstream pixel-wise supervision. This approach is central to frameworks such as WP-CrackNet, advancing the quality of weakly supervised localization through mathematically grounded spatial weighting and cross-stream regularization (Ma et al., 20 Oct 2025).

1. Purpose and Motivation

CECCM is motivated by two core challenges observed in weakly supervised detection:

1. Activation Instability: In collaborative or adversarial networks (e.g., between classifier and reconstructor components), the spatial focus of CAMs can drift, which degrades localization and pseudo-label quality.
2. Centrality Bias: Many structural targets, such as road cracks, are spatially compact; focusing CAM responses on the central, most salient regions can improve both pseudo-labeling fidelity and model convergence.

By enforcing center-focused spatial weighting and inter-stream consistency, CECCM stabilizes attention, aligns activation regions, and results in more reliable guidance for dense detectors.

2. Technical Structure and Mathematical Formulation

The mathematical operations in CECCM are as follows. Let $M_C$ and $M^{(\mathrm{rec})}_C$ denote the CAMs from the classifier and reconstructor, respectively. The module processes these maps through the following sequence:

1. Center of Mass Calculation: For a CAM $M$, the center is

$$\mu_x = \frac{\sum_{x, y} x \cdot M(x, y)}{\sum_{x, y} M(x, y)}, \quad \mu_y = \frac{\sum_{x, y} y \cdot M(x, y)}{\sum_{x, y} M(x, y)}$$

1. Center-Weighted Gaussian Map Construction: For each spatial location $(x, y)$,

$$G(x, y) = \exp \left( -\frac{(x-\mu_x)^2 + (y-\mu_y)^2}{2\sigma^2} \right)$$

where $\sigma$ governs the focus of localization.

1. Application of Center Gaussian Weight: The center-enhanced CAMs are

$$M_{CG} = M_C \odot G_C, \qquad M^{(\mathrm{rec})}_{CG} = M^{(\mathrm{rec})}_C \odot G^{(\mathrm{rec})}_C$$

where $\odot$ denotes element-wise multiplication.

1. Consistency Loss: CECCM imposes a constraint penalizing divergence between these enhanced maps:

$$L_{CEC} = \frac{1}{H \cdot W} \sum_{x, y} |M_{CG}(x, y) - M^{(\mathrm{rec})}_{CG}(x, y)|$$

This loss enforces spatial and structural agreement between the classifier and reconstructor activations, refined to the most central regions.

3. Integration into Weakly Supervised Detection Pipelines

In WP-CrackNet (Ma et al., 20 Oct 2025), CECCM operates within a tri-partite architecture: a classifier produces CAMs, a reconstructor generates complementary CAMs via feature inversion, and a detector trains on pseudo-labels derived from the post-processed CAMs. The CECCM module is situated at the interface between the classifier and reconstructor, performing the following roles:

• It processes both CAM variants through center enhancement and computes $L_{CEC}$ as a regularizer during training.
• The module checks activation stability across collaborative updates, mitigating drifts caused by adversarial objectives.
• The refined, center-weighted CAMs are subsequently post-processed (e.g., with denseCRF) and used as pseudo-labels for pixel-level fine-tuning in the detection branch.

This design ensures tight coupling between coarse image-level supervision and dense pixel-wise predictions, leveraging spatial priors to maximize localization quality.

4. Performance Enhancement and Empirical Results

Ablation studies indicate that insertion of CECCM results in measurable improvements in standard weakly supervised detection metrics:

• Integration of CECCM raises both F1-score and Intersection over Union (IoU) when compared to versions that lack this module.
• CAMs emerging from pipelines with CECCM show broader, more stable coverage and improved spatial aggregation over crack regions, as detailed in qualitative and quantitative assessments (Ma et al., 20 Oct 2025).

The progression observed in ablation figures demonstrates that, while classifiers alone typically highlight only fragmentary or maximally discriminative regions, adversarial collaboration broadens these regions. The addition of CECCM enforces central spatial coherence, facilitating higher-quality pseudo-label generation.

5. Relationship to Related Methods and Prior Work

CECCM is related to, but distinct from, earlier advances in consistency-based semi-supervised learning. In semi-supervised classification, Grad-CAM consistency loss has been proposed, penalizing differences between attention maps produced from original and augmented views of an input by means of an $L_2$ metric (Lee et al., 2021). While that line of work includes structural flexibility—allowing attention at arbitrary layers and thus variable semantic scale—the CECCM specifically adapts these ideas to the demands of weakly supervised dense localization:

• It imposes spatial centrality via explicit Gaussian weighting, a mechanism absent from generic CAM consistency methods.
• The focus on inter-stream consistency (classifier vs. reconstructor) as opposed to input-wise consistency (original vs. augmented) reflects the different demands of image localization versus classification.

A plausible implication is that the adaptability of target layers in Grad-CAM consistency could be combined with center-focused mechanisms like CECCM to create broader, task-agnostic localization consistency modules. However, CECCM's bias toward spatial centrality may particularly benefit domains where salient objects are spatially contiguous and centrally distributed.

6. Limitations and Future Extensions

Empirical results show that CECCM particularly excels in domains with prominent central features (e.g., elongated cracks in roads). In scenes with multi-object or off-center targets, central bias may reduce sensitivity to non-central activations. It is possible that augmenting CECCM with dynamic or multi-modal spatial priors could increase its flexibility. Furthermore, explicit integration of CECCM with scale-adaptive or hierarchical attention modules could extend its applicability across diverse weakly supervised tasks.

7. Visualization and Interpretation

CECCM’s effects are illustrated in qualitative figures (see “Ablation Study Results” in (Ma et al., 20 Oct 2025)), where CAMs are shown at successive pipeline stages:

• Classifier-only CAMs typically highlight sparse discriminative regions.
• Adversarial training expands activation coverage but may be unstable.
• CAMs post-CECCM exhibit pronounced spatial aggregation and centralization, correlating with improved pseudo-label quality and detector accuracy.

Although the literature does not provide a dedicated diagram for CECCM, its computational steps are summarized in the WP-CrackNet pipeline figures, and its effect is visually evident in the comparative localization maps.

In sum, the Center-Enhanced CAM Consistency Module (CECCM) is an analytically grounded, empirically validated module for enhancing weakly supervised segmentation pipelines, achieving improved pseudo-label precision and stability by integrating center-enhanced spatial focus with inter-stream consistency constraints (Ma et al., 20 Oct 2025). Its development is informed by, and extends, the theoretical foundation laid by Grad-CAM consistency in consistency-based semi-supervised learning (Lee et al., 2021).