Multimodal Image Registration and Intelligent Diagnosis of Building Facade Debonding Defects via Visible–Infrared Thermal Fusion

2.1 Overall architecture of the method

The intelligent diagnosis method for building facade debonding defects proposed in this paper adopts a two-stage cascaded architecture of registration first and then fusion segmentation. The core objective is to achieve pixel-level accurate identification and localization of debonding regions through high-precision cross-modal image alignment and deep feature fusion. The method contains three core modules: preprocessing enhancement, cross-modal registration, and fusion segmentation. These modules work collaboratively and progressively, forming a complete end-to-end diagnosis system. The preprocessing enhancement module addresses the imaging differences between visible-light and infrared thermal imaging. Through targeted processing, it improves image quality, effectively suppresses noise, and highlights key features, providing reliable inputs for subsequent registration and segmentation tasks. The cross-modal registration module achieves pixel-level accurate alignment of dual-modal images, solving the core problem of feature mismatch in heterogeneous images and laying the foundation for effective multimodal feature fusion. The fusion segmentation module fully exploits the complementary information of the aligned dual-modal images, integrating the texture and structural details of visible-light images with the temperature anomaly features of infrared thermal imaging, thereby completing accurate segmentation and boundary optimization of debonding defects. The entire architecture is trained and optimized in an end-to-end manner to ensure the collaborative consistency among modules, significantly improving the accuracy, robustness, and engineering applicability of the diagnosis method. Figure 1 shows the overall architecture of the intelligent diagnosis method for facade debonding defects.

Figure 1. Overall architecture of the intelligent diagnosis method for facade debonding defects

2.2 Multimodal image preprocessing and enhancement

The single-channel temperature matrix of infrared thermal imaging is difficult for convolutional neural networks to effectively extract features from, and the visual discriminability of temperature differences is relatively low. Traditional pseudo-color mapping adopts fixed mapping rules and cannot adapt to the temperature distribution characteristics in different scenarios [15, 16]. To address this problem, an adaptive pseudo-color mapping strategy is proposed. According to the statistical characteristics of temperature in infrared thermal images, the mapping parameters are dynamically adjusted, converting the single-channel temperature matrix into three-channel visual features. The core mapping formula is:

$I_c(x, y)= \begin{cases}\frac{T(x, y)-T_{\min }}{T_{\operatorname{mid}}-T_{\min }} & T_{\min } \leq T(x, y)<T_{\operatorname{mid}} \\ 1 & T_{\operatorname{mid}} \leq T(x, y)<T_{\max } \\ \frac{T_{\max }-T(x, y)}{T_{\max }-T_{\operatorname{mid}}} & T(x, y) \geq T_{\max }\end{cases}$                (1)

where, I_c(x,y) is the pixel value of the mapped three-channel image, $c \in\{R, G, B\}$correspond to the red, green, and blue channels respectively, T(x,y) is the temperature value at pixel (x,y) in the infrared thermal image, and T_min, T_mid, and T_max are the minimum temperature, middle temperature, and maximum temperature of the image, respectively, which are adaptively determined by statistical analysis of the temperature histogram of the image. This strategy not only enhances the perception ability of human vision for temperature anomalies, but also converts single temperature information into multi-channel visual features, which is well adapted to the feature extraction requirements of convolutional neural networks. At the same time, to address the problem of temperature noise in infrared thermal images, an improved bilateral filtering algorithm is proposed. By introducing temperature gradient weights to adjust the filter kernel coefficients, the filtering formula is:

$\begin{gathered}I_{i n f r a}(x, y)=\frac{1}{K(x, y)} \sum_{i, j} G_{\sigma_s}(|(x, y)-(i, j)|) \\ G_{\sigma_r}(|T(x, y)-T(i, j)|) \cdot w_g \cdot T\left(i_j\right)\end{gathered}$              (2)

where, K(x,y) is the normalization coefficient, $G_{\sigma_s}$ and $G_{\sigma_r}$ are the Gaussian functions in the spatial domain and gray-level domain respectively, and w_g is the temperature gradient weight, which is calculated from the temperature gradient between pixel (x,y) and neighboring pixels. The larger the temperature gradient, the closer w_g is to 1. In this way, while smoothing noise in homogeneous regions, the temperature anomaly edges are preserved to the greatest extent, solving the edge blurring problem caused by traditional bilateral filtering.

Visible-light images are easily affected by illumination angle, shadows, and other factors, resulting in uneven illumination, which leads to bias in texture structure feature extraction and affects subsequent registration accuracy [17]. Based on Retinex theory [18], an improved illumination balancing algorithm is proposed. It breaks the limitation that traditional algorithms only adjust overall illumination, and combines the structural prior obtained by Canny edge extraction to guide illumination correction. First, the visible-light image is decomposed into a reflectance component R(x,y) and an illumination component L(x,y), with the decomposition formula:

$I_{v i s}(x, y)=R(x, y) \cdot L(x, y)$                  (3)

For the decomposed illumination component, the structural prior mask M(x,y) obtained by Canny edge detection is introduced. In the mask, the pixel value of edge regions is 1 and that of non-edge regions is 0. Adaptive illumination compression is performed using the following formula:

$L^{\prime}(x, y)=L(x, y)^{\alpha M(x, y)+\beta(1-M(x, y))}$               (4)

where, α and β are the illumination compression coefficients for edge regions and non-edge regions respectively, $\alpha \in(0,1)$ and $\beta \in(0, \alpha)$, which are determined through adaptive iterative optimization. This ensures that the illumination compression in edge regions is smaller than that in non-edge regions, eliminating uneven illumination while avoiding distortion of edge structures. The corrected reflectance component R′(x,y)=I_vis(x,y)/L′(x,y) is processed by gray normalization and then used as the input for the subsequent registration task, providing accurate structural reference for cross-modal registration.

Through targeted innovative processing for the two modal images, the preprocessing and enhancement module effectively suppresses noise interference and solves problems such as uneven illumination and blurred temperature anomaly edges. It significantly improves image quality and the distinguishability of key features, laying a solid foundation for the efficient operation of subsequent cross-modal registration and fusion segmentation modules, and ensuring the accuracy and robustness of the entire diagnosis system.

2.3 End-to-end registration network based on cross-modal feature mutual guidance

Dual-branch Feature Mutual-Guided Registration Network (DFMGR-Net) adopts an encoder-decoder end-to-end architecture. The core innovation lies in the collaborative design of dual-branch feature extraction and hierarchical cross-modal interaction, achieving precise extraction and mutual enhancement of heterogeneous image features. Figure 2 shows the architecture of the dual-branch feature mutual-guidance registration network. As shown in the figure, the encoder is divided into two parallel branches for visible-light and infrared modalities. The differential network design adapts to the feature characteristics of the two modalities while ensuring alignment of feature hierarchies, providing the basis for subsequent attention interaction. The visible-light branch uses a lightweight 50-layer Residual Network (ResNet50) as the backbone network, removing the fully connected layers and retaining all residual modules. Multi-stage residual learning extracts multi-scale texture and structural features, and the output feature map is denoted as $F_{v i s}^l \in \mathrm{R}^{C_l \times H_l \times W_l}$, where l represents the encoder feature hierarchy, and C_l, H_l, W_l are the channel number, height, and width of the corresponding hierarchy. The infrared branch is a customized lightweight convolutional neural network. Considering the low-texture characteristic of temperature distribution features, it adopts 3 × 3 small convolution kernels and sparse channels, setting feature hierarchies fully matched with the visible-light branch. It outputs multi-scale temperature distribution feature maps $F_{i n f}^l \in \mathrm{R}^{C_l \times H_l \times W_l}$, ensuring temperature feature extraction accuracy while reducing computation by more than 40%. Features from both branches at each hierarchy are synchronously input into the cross-modal attention interaction module. The enhanced features are channel-concatenated and sent to the decoder’s deformation field estimation module, forming an “feature extraction – interaction enhancement – deformation estimation” end-to-end registration pipeline.

The core objective of the cross-modal attention interaction module is to establish a bidirectional guidance mechanism between visible-light structural features and infrared temperature features, solving the problem of weak correlation and high difficulty in aligning heterogeneous modality features. This module is embedded in each feature hierarchy of the encoder to achieve multi-scale feature mutual enhancement, ensuring that both low-level details and high-level semantics can complete effective cross-modal information transfer. The core logic is to constrain the spatial positioning of infrared features by the spatial structure information of visible-light images, while using infrared temperature response information to select key feature regions in visible-light images, forming a complementary enhancement closed loop between modalities.

Figure 2. Dual-branch feature mutual-guidance registration network architecture

The core implementation of the module is the joint computation of channel attention and spatial attention, achieving fine feature selection through the collaboration of the two weights. First, the input feature maps of the dual branches are channel-aligned, then channel attention weights and spatial attention weights are computed separately. The channel attention weights are calculated based on global feature statistics. For visible-light features $F_{v i s}^l$ and infrared features $F_{\text {inf }}^l$, global average pooling is performed to obtain channel-level feature vectors $v_{ {vis}}^l \in \mathrm{R}^{C_l}$ and $v_{{inf}}^l \in \mathrm{R}^{C_l}$, which are mapped by shared two-layer fully connected layers. Finally, Sigmoid activation produces the channel attention weights:

$\begin{aligned} & W_{c, { vis }}^l=\sigma\left(W_2 \cdot \operatorname{ReLU}\left(W_1 \cdot v_{i n f}^l\right)\right), \\ & W_{c, i n f}^l=\sigma\left(W_2 \cdot \operatorname{ReLU}\left(W_1 \cdot v_{v i s}^l\right)\right)\end{aligned}$           (5)

where, W₁ and W₂ are fully connected layer weights, and σ is the Sigmoid activation function. The spatial attention weights focus on local region features. The aligned dual-modal feature maps are channel-concatenated, reduced in dimension by a 1×1 convolution, then spatial correlations are extracted via a 3×3 depth convolution, followed by Sigmoid activation to generate spatial weight maps:

$\begin{gathered}W_{s, v i s}^l=\sigma\left(\operatorname{Conv}_{3 \times 3}\left(\operatorname{Conv}_{1 \times 1}\left(F_{v i s}^l \oplus F_{i n f}^l\right)\right)\right), \\ W_{s, i n f}^l=W_{s, v i s}^l\end{gathered}$                    (6)

where, $\oplus$ denotes channel concatenation and Conv_k×k represents convolution with kernel size k × k.

Bidirectional feature enhancement is realized by element-wise multiplication of the attention weights with the original feature maps. The joint channel and spatial attention weights are: $W_{v i s}^l=W_{c, v i s}^l \otimes W_{s, v i s,}^l, W_{i n f}^l=W_{c, i n f}^l \otimes W_{s, i n f}^l$, where $\oplus$ denotes element-wise multiplication. The final enhanced features are: $F_{v i s}^l=F_{v i s}^l \cdot W_{v i s}^l, F_{i n f}^l=F_{i n f}^l \cdot W_{i n f}^l$.

This process allows the infrared branch to receive the high-precision spatial structure constraints from visible-light images, precisely locating temperature anomaly positions, while the visible-light branch focuses on thermal anomaly regions indicated by infrared features, filtering irrelevant wall texture noise. The enhanced dual-modal features have improved heterogeneity correlation, providing high-quality input for the decoder’s dense deformation field estimation and significantly improving cross-modal registration accuracy and robustness.

The core function of the dense deformation field estimation module is to generate a high-precision dense displacement field that adapts to the complex spatial transformations of building facades, solving the core problem that traditional convolution cannot model non-rigid deformation. This module adopts deformable convolution to construct a hierarchical deformation estimation structure. Extra-learned offsets guide the convolution kernel for adaptive non-grid point sampling, flexibly capturing perspective deformation, subtle wall deformation, and local geometric distortions. The output of deformable convolution is calculated as:

$y(x, y)=\sum_{k=1}^K w_k \cdot x\left(x+\Delta x_k, y+\Delta y_k\right)$           (7)

where, y(x,y) is the output feature at pixel (x,y), K is the number of convolution sampling points, w_k is the convolution weight of the k-th sample point, and Δx_k, Δy_k are the extra learned offsets in [-1,1], adaptively learned from the concatenated enhanced dual-modal features. The offsets allow convolution kernels to dynamically adjust sampling positions according to local image geometry, effectively adapting to non-rigid deformation of building facades without preset grids. The enhanced dual-modal features are processed through 4 layers of deformable convolution hierarchically, gradually improving deformation modeling capability. The module finally outputs a dense displacement field Δ(x,y)=(Δx(x,y),Δy(x,y)) of the same size as the input images, achieving pixel-level accurate alignment between visible-light and infrared images, providing a reliable spatial alignment foundation for subsequent multimodal feature fusion.

To ensure that the registration results satisfy both visual structural alignment and the physical distribution of infrared temperature anomalies, an innovative loss function considering visual similarity and physical consistency is designed. Through multi-loss collaborative constraints, registration accuracy and reliability are improved, differing from existing loss designs that only focus on visual matching. The total loss function is a weighted sum of multi-scale structural similarity loss, mutual information loss, and temperature consistency loss:

$L_{\text {total }}=\lambda_1 L_{M S-S S I M}+\lambda_2 L_{M I}+\lambda_3 L_{T C}$             (8)

where, λ₁, λ₂, and λ₃ in [0,1] are weights for each loss component, determined via adaptive iterative optimization, ensuring coordinated effect of all loss components.

The multi-scale structural similarity loss constrains the structural alignment of registered images, reducing structural distortion caused by modality heterogeneity. It is computed based on multi-scale image decomposition, comparing the luminance, contrast, and structure of registered and reference images layer by layer:

$L_{M S-S S I M}=1-\prod_{i=1}^N \operatorname{SSIM}_i\left(I_{r e g}, I_{r e f}\right)$           (9)

where, N is the number of decomposition scales (set to 4 in this paper), SSIMi is the structural similarity coefficient at scale i, I_reg is the registered image, and I_ref is the reference image. The mutual information loss measures the statistical dependency between registered cross-modal images, quantifying fusion consistency:

$L_{M I}=-I\left(I_{r e g, v i s}, I_{r e g, i n f r a}\right)=-\sum_u \sum_v p(u, v) \log \frac{p(u, v)}{p(u) p(v)}$             (10)

where, I( , ) is mutual information, p(u,v) is the joint probability density of registered visible $I_{ {reg,vis}}$ and infrared $I_{ {reg,infica}}$ images, and p(u), p(v) are their marginal probabilities.

The temperature consistency loss is the core innovative component, constraining the deformation field to not distort the physical distribution of infrared temperature anomalies, ensuring that the registered infrared image accurately reflects the temperature characteristics of debonding regions. It is computed by the difference in temperature gradients before and after registration:

$L_{T C}=\frac{1}{H \times W} \sum_{x=1}^H \sum_{y=1}^W\left|\nabla T_{r e g}(x, y)-\nabla T_{o x i}(x, y)\right|$          (11)

where, H, W are the height and width of the infrared image, ∇T_ori(x,y) is the temperature gradient of the original infrared image at pixel (x,y), and ∇T (x,y) is the temperature gradient of the registered infrared image, computed by the Sobel operator. This loss effectively suppresses distortion of temperature anomaly regions by the deformation field, ensuring that temperature distribution features of debonding regions are preserved during registration, providing accurate temperature feature support for subsequent defect segmentation.

2.4 Multi-scale adaptive fusion and intelligent debonding defect segmentation

MAF-SegNet takes the registered visible-light and infrared dual-modal images as input, adopting a dual encoder-decoder architecture. The core innovation lies in hierarchical dual-modal feature fusion and detail preservation mechanisms, focusing on pixel-level precise segmentation and boundary optimization of debonding defects. The specific architecture is shown in Figure 3. The dual encoders are designed to match the feature hierarchies of the registration network, separately processing the two modality images to fully exploit complementary information: the visible-light encoder continues the lightweight Residual Network structure, focusing on extracting wall texture, edges, and other structural features, outputting multi-scale feature maps $F_{v i s}^s \in \mathrm{R}^{C_s \times H_s \times W_s}$; the infrared encoder adopts a customized lightweight Convolutional Neural Network (CNN), focusing on capturing temperature anomaly-related features, outputting corresponding scale feature maps $F_{i n f}^s \in \mathrm{R}^{C_s \times H_s \times W_s}$, where s is the feature scale, and C_s, H_s, W_s are the channel number, height, and width at the corresponding scale. The decoder uses a layer-by-layer upsampling structure. Through skip connections, the dual-modal fused features of each encoder scale are integrated with the corresponding decoder layer features, gradually restoring spatial resolution while effectively suppressing gradient vanishing, ensuring that features of subtle debonding regions and defect edges are not lost. The network finally outputs a debonding segmentation probability map with the same size as the input image, achieving precise defect localization and complete segmentation.

The channel-spatial attention fusion module is the core innovative component of MAF-SegNet. Its main objective is to solve the problems in traditional dual-modal feature concatenation, such as feature redundancy and the drowning of useful information. Through the collaborative effect of dual attention mechanisms, fine-grained fusion of multi-modal information is realized, enhancing the discriminative capability for debonding defects. The module schematic is shown in Figure 4. The module is embedded at every feature scale layer in the encoder and decoder, enabling precise feature selection and fusion at multiple scales, ensuring that both low-level detail features and high-level semantic features are effectively integrated, fully leveraging the complementary advantages of dual-modal images.

The channel attention mechanism is used to adaptively select the most critical feature channels for debonding diagnosis, achieving channel-level weight allocation based on global feature statistics. First, the dual-modal feature maps are concatenated along the channel dimension to obtain the initial fused feature $F_{c a t}^s=F_{v i s}^s \oplus F_{i n f}^s$. Then, global average pooling is applied to extract the channel-level global feature vector $v^s \in \mathrm{R}^{2 C_s}$. The feature vector is mapped and dimension-reduced through two fully connected layers, and finally Sigmoid activation generates the channel attention weights $W_c^s$, calculated as:

$W_c^s=\sigma\left(W_b \cdot \operatorname{ReLU}\left(W_a \cdot v^s\right)\right)$               (12)

where, $W_a \in \mathrm{R}^{\left(2 C_s\right) / 4 \times 2 C_s}, W_b \in \mathrm{R}^{2 C_s \times\left(2 C_s\right) / 4}$ are the weights of the two fully connected layers, σ is the Sigmoid activation function. The weight vector is multiplied element-wise with the initial fused feature, effectively enhancing the texture, edge, and temperature-related features critical for debonding diagnosis while suppressing redundant channel interference.

The spatial attention mechanism focuses on potential debonding regions, strengthening the discriminative ability of local features and compensating for the channel attention mechanism's neglect of local spatial information. The channel-attention-weighted feature is first reduced in dimension by a 1×1 convolution, reducing computation while achieving preliminary channel information fusion. Then, a 3×3 depth convolution extracts local spatial correlation features, capturing spatial differences between debonding regions and surrounding background. Sigmoid activation generates the spatial attention weight map $\boldsymbol{W}_s^s \in \mathrm{R}^{1 \times H_s \times W_s}$, calculated as:

$W_s^s=\sigma\left(\operatorname{Conv}_{3 \times 3}\left(\operatorname{Conv}_{1 \times 1}\left(F_{c a t}^s \cdot W_c^s\right)\right)\right)$                (13)

where, Conv_k×k denotes convolution with kernel size k×k. The final fine-grained fused feature is obtained through the collaborative effect of dual weights: $F_{\text {fusion }}^s=\left(F_{{cat}}^s \cdot W_c^s\right) \otimes W_s^s$, where $\oplus$ denotes element-wise multiplication. This fusion preserves the structural detail advantages of visible-light images while fully utilizing the temperature anomaly features of infrared images, effectively suppressing irrelevant background noise and significantly enhancing the discriminative capability of fused features for debonding defects, providing high-quality input for accurate decoder segmentation.

Figure 3. Multi-scale adaptive fusion segmentation network architecture

Figure 4. Channel-spatial attention fusion module schematic

To overcome the inherent limitation of convolutional neural networks' local receptive field and enhance the network's perception of large-scale debonding defects, a lightweight Transformer module is embedded at the deepest encoder layer. Through self-attention, global spatial correlation information of the image is captured, enabling complete recognition of large-scale debonding regions. The module adopts channel compression and sparse attention design, maintaining long-range dependency capture while effectively controlling computation to meet the real-time requirements of segmentation tasks. The core self-attention calculation is:

$\operatorname{Attn}(Q, K, V)=\operatorname{Softmax}\left(\frac{Q K^T}{\sqrt{d_s}}\right) V$              (14)

where, Q, K, V are the query, key, and value matrices, obtained by linear transformation of the deepest encoder fused features, and d_k is the dimension of the query matrix. For lightweight design, a channel grouping strategy divides feature channels into 4 groups, each computing self-attention independently, and then concatenating along channels to fuse results, reducing computational complexity while preserving feature correlation across channels. This module effectively captures the global distribution features of large-scale debonding regions, avoiding incomplete segmentation due to local receptive field limitations, and enhancing global contrast between debonding regions and background, improving feature discriminability.

To address common problems in existing segmentation methods, such as blurred debonding edges and discontinuous segmentation results [19, 20], an edge loss is introduced at the network output layer to precisely constrain the segmentation accuracy of defect edges, forcing the network to learn clear boundaries of debonding regions. Edge loss is calculated based on the edge difference between segmentation results and ground truth. First, the Sobel operator extracts edges from the ground truth to generate an edge mask MÎ{0,1}, where edge pixels are 1 and non-edge pixels are 0. The edge loss uses a binary cross-entropy form, penalizing segmentation errors only at edge regions, calculated as:

$L_{\text {edge }}=-\frac{1}{N_{\text {edge }}} \sum_{x=1}^H \sum_{y=1}^W M(x, y)\left|\begin{array}{c}Y(x, y) \log P(x, y) \\ +(1-Y(x, y)) \log (1-P(x, y))\end{array}\right|$            (15)

where, N_edge is the total number of edge pixels, Y(x,y) is the ground truth, and P(x,y) is the network output segmentation probability map. Edge loss collaborates with the main segmentation loss; the main loss ensures overall segmentation accuracy of debonding regions, while edge loss focuses on edge detail optimization, effectively solving boundary blur and discontinuity issues, improving completeness and accuracy of debonding defect segmentation.

The combination of long-range dependency enhancement and edge constraint enables complete perception of large-scale debonding regions and precise segmentation of fine edges, compensating for traditional segmentation networks' shortcomings in global association and edge details. The lightweight Transformer module combined with edge loss ensures computational efficiency while significantly improving segmentation performance, allowing the network to adapt to debonding defects of different scales and shapes, enhancing robustness.

To further optimize the coarse segmentation output, eliminate isolated false positives, fill small missed regions, and improve segmentation completeness and accuracy, conditional random field (CRF) is used for post-processing. Unlike existing methods that perform no post-processing or only simple threshold-based segmentation, CRF fully utilizes spatial neighborhood relationships between pixels for fine adjustment of segmentation results. CRF takes the network output coarse segmentation probability map as input, defines an energy function to describe pixel relationships, and minimizes the energy function to obtain the optimal segmentation:

$E(\mathrm{X})=E_{\text {unarv }}(\mathrm{X})+\lambda E_{\text {pairwise }}(\mathrm{X})$               (16)

where, X is the final segmentation result, E_unary is unary potential computed directly from the network output segmentation probability map, penalizing deviation from the coarse segmentation, E_pairwise is pairwise potential describing spatial relationships between adjacent pixels, and λ is the weight balancing unary and pairwise potentials.

The pairwise potential is calculated using a Gaussian kernel considering both spatial distance and intensity similarity of adjacent pixels:

$E_{ {pairwise}}(\mathrm{X})=\sum_{i \neq j} \exp \left(-\frac{\left|x_i-x_j\right|^2}{2 \sigma_z^2}-\frac{\left|I\left(x_i\right)-I\left(x_j\right)\right|^2}{2 \sigma_\tau^2}\right) \cdot \delta\left(X_i, X_j\right)$              (17)

where, xi, xj are coordinates of adjacent pixels, I(x_i), I(x_j) are the corresponding pixel intensity values, σ_s, σ_r are standard deviations of spatial and intensity Gaussian kernels, and δ( ) is an indicator function, 0 if X_i=X_j, 1 otherwise. This pairwise potential effectively suppresses isolated false positives, strengthens category consistency of adjacent pixels, and fills small missed regions, making segmentation results more consistent with actual debonding defect shapes. Through CRF post-processing, false detection and missed detection rates are significantly reduced, boundary smoothness and regional completeness are further improved, providing an accurate segmentation basis for subsequent quantitative analysis of debonding defects.

This work addresses core issues of strong concealment of building exterior wall debonding defects, low efficiency of traditional detection, insufficient multi-modal image registration accuracy, and imprecise feature fusion, proposing a multi-modal image registration and debonding defect intelligent diagnosis method that fuses visible and infrared thermal imaging. The method adopts a two-stage cascaded architecture of registration followed by fusion segmentation. The core innovations include a dual-branch feature mutually guided registration network and a multi-scale adaptive fusion segmentation network, constructing a registration loss function considering both visual similarity and physical consistency, and building a large-scale UAV dual-modal debonding dataset. Experimental results show that the proposed registration network significantly outperforms existing mainstream methods in RMSE and MI, maintaining high-precision pixel-level alignment even in large-scale deformation scenarios. The segmentation network performs outstandingly on Dice coefficient and boundary F1 score, achieving precise segmentation and boundary optimization of debonding defects of different scales and types, controlling false positive and false negative rates within 5%, while maintaining high detection speed and environmental robustness, effectively addressing the core bottlenecks of existing methods.

This work provides a new technical path and theoretical support for cross-modal image registration and intelligent building defect detection, with significant academic and engineering practical value. Academically, the proposed cross-modal attention interaction mechanism, temperature consistency loss, and channel-spatial attention fusion module effectively solve industry problems of low heterogeneous image registration accuracy and imprecise feature fusion, providing referable design ideas for similar multi-modal registration and segmentation tasks. In engineering practice, this method does not rely on inspector experience, showing significantly higher detection efficiency and accuracy than traditional methods, adapting to complex detection scenarios of different seasons, lighting, and wall materials. It can be widely applied in intelligent detection of building exterior wall debonding, reducing detection costs and improving detection safety, providing reliable technical support for building structural safety maintenance. Future work can further optimize network lightweight design to improve deployment on embedded devices and extend the method to other building defect types, promoting the further development of intelligent building inspection technology.