A method and system for identifying seismic-resistant building structure types based on street view maps and deep learning

Abstract

This invention relates to the fields of artificial intelligence and earthquake engineering technology, and provides a method for identifying seismic-resistant building structure types based on street view maps and deep learning. The method includes acquiring and processing street view image data containing buildings, constructing an improved YOLOv11 target detection model, guiding the network to focus on key structural areas by fusing prior geometric attributes of the buildings, enhancing the ability to distinguish structural types, designing a structure-aware pruning strategy to retain feature channels that contribute significantly to structural identification for edge deployment needs, and deploying the optimized model on an edge computing device. Inputting a street view image enables rapid and automatic identification of building structure types. This method is designed for intelligent identification of four typical seismic-resistant building structure types in street view images: brick-wood structures, brick-concrete structures, frame structures, and civil engineering structures. It has advantages such as high structural identification accuracy, small model size, and high deployment efficiency, aiming to improve the automation and accuracy of building seismic performance assessment.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and earthquake engineering technology, specifically to a method and system for identifying the seismic-resistant structural types of buildings based on street view maps and deep learning. Background Technology

[0002] Earthquake disasters severely threaten human life and building safety. The seismic response characteristics of building structures directly affect their damage level and seismic performance. Therefore, monitoring the seismic response of seismically resistant building structures is of great significance for early identification of structural damage and post-earthquake safety assessment. Building structure type, as a key factor affecting seismic performance, is an input variable for most vulnerability assessment models. Efficiently and accurately obtaining information on the structure types of buildings over a wide area is crucial for seismic assessment. Traditional structure type identification relies on manual, building-by-building surveys, which are tedious, costly, and difficult to meet the needs of rapid identification of buildings over a wide area. Although remote sensing technology has become an important research direction due to its large-scale and high-timeliness characteristics, it lacks information on building facades due to the limitation of the overhead view. Multi-source data fusion and domain knowledge assistance are needed. Related research mainly includes machine learning or deep learning methods based on multi-source remote sensing images, as well as multimodal methods that integrate domain knowledge.

[0003] In recent years, the YOLO series of algorithms have been widely used in the field of object detection due to their advantages such as end-to-end modeling, and have been continuously optimized and upgraded. Street view images can more comprehensively reflect the details and true appearance of building facades compared to remote sensing images, and have a natural advantage in capturing structural clues. However, existing research has mostly focused on semantic analysis, and its application in seismic structure type identification is insufficient. In summary, existing structure identification methods have shortcomings such as the lack of facade information in remote sensing images, the insufficient application of street view images in seismic structure identification, and the large model size, which makes them difficult to deploy on edge devices.

[0004] To address the aforementioned technical problems, this invention proposes a method and system for identifying seismic-resistant building structure types based on street view maps and deep learning. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for identifying seismic-resistant building structure types based on street view maps and deep learning. This solves the problems of insufficient accuracy, reliance on remote sensing overhead views leading to missing structural details, low efficiency of manual surveys, and large and difficult-to-deploy models in existing methods for identifying seismic-resistant building structure types.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and system for identifying seismic-resistant building structure types based on street view maps and deep learning, comprising the following steps:

[0007] S1. Street View Image Acquisition and Preprocessing

[0008] Street view image data of the target area is collected, and image enhancement and preprocessing operations are performed, including brightness normalization, adaptive contrast enhancement, random occlusion simulation, size scaling and color dithering, to improve the robustness and recognition adaptability of the network.

[0009] S2. Building Structural Attribute Extraction and Embedding Coding

[0010] Structural attributes are estimated on the preprocessed images, and geometric and semantic features such as floor height, window opening ratio, roof shape, wall texture, structural cues, and facade proportions are extracted, encoded into structural attribute guidance vectors, and input into the network to guide the model to focus on key structural areas.

[0011] S3. Improved YOLO recognition model construction

[0012] Based on the YOLOv11 backbone network, a detection model integrating structure prior perception capability is designed, mainly including the following structure optimization modules: replacing the backbone convolution with a windmill-shaped convolution (PConv) to enhance the receptive field and spatial modeling capability; introducing a lightweight multidimensional joint attention module (Lite-MDJA) and dilated convolution to improve the multi-scale feature extraction effect; using a lightweight and efficient CARAFE upsampling module to enhance the feature decoding quality; and embedding a structure attribute guidance module (SAPA) in the middle of the backbone to realize the structure-guided attention mechanism.

[0013] S4. Teacher Model Training

[0014] On a street view building image dataset with structural labels, the improved YOLO recognition model is trained under supervision to obtain a high-performance teacher model as the source network for knowledge distillation.

[0015] S5. Model Compression and Structure-Guided Distillation Optimization

[0016] The teacher model undergoes structure-aware pruning, with importance assessment and redundancy reduction based on the contribution of each channel to structure recognition. A lightweight student model is constructed and trained using a structure-region-guided knowledge distillation strategy. The accuracy of the student model is improved through joint optimization of category distillation and feature distillation.

[0017] S6. Inference Deployment and Result Output

[0018] The optimized student model is deployed to edge devices, such as mobile terminals, seismic survey instruments, or drone platforms, to achieve rapid and automatic identification of building structure types and output structure category labels, target locations, and confidence scores, supporting visualization and integration with structural information management systems.

[0019] To achieve the above method, the present invention also provides a street view image building seismic structure type recognition system. This system supports efficient seismic structure recognition and edge deployment in various scenarios, and specifically includes the following modules:

[0020] 1. Street view image acquisition and preprocessing module

[0021] Functions: Collect street view image data of the target area, supporting multi-source input (such as ground cameras, vehicle-mounted equipment, drones, etc.); perform image preprocessing and enhancement, including brightness normalization, adaptive contrast adjustment, image cropping and scaling, random occlusion simulation, color perturbation, etc., to improve the robustness of the model in complex environments.

[0022] 2. Structural Attribute Extraction and Encoding Module

[0023] Function: Automatically estimates the structural attributes of a building based on image content or auxiliary information, including but not limited to the number of floors, floor-to-height ratio, window opening ratio, roof shape, wall texture, window type ratio, facade width-to-height ratio, structural line features, etc.; encodes the extracted results into structural attribute vectors, which are used as input to the structural guidance module to enhance the model's ability to focus on structural regions.

[0024] 3. Improve the YOLO recognition model module

[0025] Functionality: Build a deep recognition network based on YOLOv11, integrating the following improved modules:

[0026] Windmill-shaped convolutional module (PConv): Enhances the ability to model small targets and local structural information;

[0027] Lightweight Multidimensional Joint Attention Module (LiteMDJA): Enhances feature selectivity and spatial attention capabilities;

[0028] Content-Aware Upsampling Module (CARAFE): Improves spatial reconstruction accuracy during the decoding stage;

[0029] Structural Attribute Guidance Module (SAPA): Introduces structural attribute vectors as attention guidance to improve the ability to extract structural features.

[0030] 4. Teacher Model Training Module

[0031] Function: Train an improved YOLO model on a dataset with structure labels to obtain a teacher model with high structure recognition accuracy, which can then be used as the source model for subsequent distillation optimization.

[0032] 5. Model Compression and Distillation Optimization Module

[0033] Function: Perform structure-aware pruning operations to identify and remove network channels or operators that contribute little to structure recognition; construct a lightweight student network and introduce a structure region-guided knowledge distillation strategy; use category distillation and multi-scale feature distillation to jointly optimize the student model performance, thereby maintaining recognition accuracy while compressing the number of model parameters.

[0034] 6. Results Output and Visualization Module

[0035] Functions: Outputs the identification results as structural type labels, target bounding boxes, and confidence scores. It supports overlay display on the original image or export as a vectorized structural information format (such as GeoJSON or Shapefile). It can be connected to seismic survey systems, urban BIM platforms, or post-earthquake rapid assessment systems to realize the application of results.

[0036] The system has end-to-end reasoning capabilities, covering the entire process from image acquisition, attribute modeling, structure recognition to model deployment and result management. It supports rapid structural type analysis of urban and rural building complexes, meeting the technical requirements of actual earthquake resilience planning and disaster risk management.

[0037] This invention provides a method and system for identifying seismic-resistant building structure types based on street view maps and deep learning. It has the following beneficial effects:

[0038] 1. This invention constructs an improved YOLO network that integrates windmill-shaped convolution, multi-dimensional joint attention mechanism, dilated convolution, and content-aware upsampling module, which can effectively enhance the ability to extract key structural features of buildings under multi-scale and complex backgrounds. Furthermore, by introducing a structural attribute guidance module, the model can integrate prior information on building geometry, significantly improving the ability to distinguish between brick-wood, brick-concrete, frame, and civil engineering structural types. In particular, it maintains good recognition performance under complex conditions such as shading, changes in lighting, and heterogeneous rural buildings, providing a reliable basis for the judgment of seismic performance of building structures.

[0039] 2. This invention employs a structure-aware pruning strategy, which retains channels and operators that significantly contribute to structure recognition during model compression, effectively reducing model redundancy and improving network operating efficiency. Combined with a knowledge distillation mechanism guided by structural regions, it can train a high-performance lightweight student model while preserving the expressive power of the teacher model, greatly reducing the number of model parameters and computational resource consumption.

[0040] 3. The optimized student model of this invention has a smaller computational load and memory footprint, and can run stably in edge computing environments such as mobile terminals, drone platforms, and handheld devices, meeting the real-time inference requirements of actual scenarios such as earthquake surveys, earthquake patrols, and post-earthquake investigations.

[0041] 4. The street view image building seismic performance structure recognition system constructed by this invention can replace the traditional manual building-by-building survey method, significantly improve the efficiency of building structure information acquisition, reduce labor costs, and has cross-regional generalization ability. It is applicable to various environments such as urban and rural areas, new and old buildings, and serves the construction of earthquake-resistant cities and spatially precise disaster risk management. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a typical street view image example of the present invention;

[0044] Figure 3 This is a schematic diagram of the seismic structure type identification model based on the improved YOLO network of the present invention.

[0045] Figure 4 This is a schematic diagram of the network skeleton of the windmill convolution module of the present invention;

[0046] Figure 5 This is a schematic diagram of the overall structure of the content-aware upsampling module of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example:

[0049] Please see the appendix Figure 1 This invention provides a method and system for identifying seismic-resistant building structure types based on street view maps and deep learning, comprising the following steps:

[0050] S1. Street View Image Acquisition and Preprocessing

[0051] In this embodiment, street view image acquisition mainly includes two methods: on-site shooting and web scraping. On-site shooting uses smartphones with high image clarity (such as Xiaomi 11 Ultra, iPhone 14 Pro, etc.) to collect street view images of urban and rural areas in Jiangsu Province, covering various building structure types, including brick-wood structures, brick-concrete structures, frame structures, and civil engineering structures. Shooting is conducted under clear weather conditions with unobstructed frontal views, and the shooting angle is controlled within ±30° of the frontal view to ensure clear structural details. Web image data mainly comes from public street view platforms (such as Gaode Street View, Baidu Street View, Tencent Street View, etc.), selecting representative street view building images from typical areas of Jiangsu Province (such as cities like Suqian, Taizhou, and Yancheng, and surrounding rural areas). Blurry, low-resolution, nighttime, or severely obstructed images are excluded during the screening process. This experiment collected a total of 5,419 street view images of urban and rural areas in Jiangsu Province, covering five typical earthquake-resistant building structure types: flat-roof structures, frame structures, brick-wood structures, brick-concrete structures, and civil engineering structures. Representative images of each structure are shown below. Figure 2 As shown.

[0052] The following preprocessing operations are performed on the acquired street view images:

[0053] Image scaling: Scale the original image to the standard input size of 640×640 pixels, maintaining the same aspect ratio, and padding the edges;

[0054] Brightness normalization: Linearly normalize the image so that the pixel brightness values ​​are distributed in the [0,1] interval;

[0055] Adaptive contrast enhancement: The CLAHE (Contrast-Limited Adaptive Histogram Equalization) algorithm is used to improve the contrast of local structures in the image;

[0056] Color perturbation enhancement: Randomly vary the saturation, hue, and brightness of the image (within ±20%) to improve the model's robustness to changes in lighting.

[0057] Random occlusion simulation: Several occlusion blocks are randomly inserted into the image (occlusion ratio controlled between 5% and 15%) to simulate occlusion situations such as vehicles and trees in reality;

[0058] Normalization and standardization: The image channels were standardized based on the ImageNet mean and standard deviation (mean = [0.485, 0.456, 0.406], standard deviation = [0.229, 0.224, 0.225]).

[0059] The above preprocessing process is implemented using Python scripts, with batch processing based on the OpenCV and Albumentations libraries. The enhanced images and corresponding structural labels (provided by manual annotation or field matching) are then used together for subsequent network training and feature extraction tasks.

[0060] Finally, a street view image dataset for identifying earthquake-resistant building structure types was constructed, containing 5,419 labeled images. This dataset covers five structural types: brick-concrete structures (3,404 images), brick-timber structures (3,017 images), frame structures (1,213 images), civil engineering structures (300 images), and flat-roof structures (225 images). This dataset basically reflects the typical structural characteristics of buildings in urban and rural areas. The dataset was proportionally divided into a training set (70%), a validation set (20%), and a test set (10%) to ensure the scientific rigor and generalization ability of the model evaluation.

[0061] S2. Building Structural Attribute Extraction and Coding

[0062] In this embodiment, to enhance the network model's ability to distinguish earthquake-resistant structural types, a structural attribute guidance mechanism is introduced. By constructing a "structural attribute vector" as auxiliary information, the backbone network is guided to focus on key structural areas of the building, thereby enhancing the extraction of structural clues and improving recognition accuracy and robustness. The structural attribute vector consists of observable or estimable building geometric attributes from street view images, mainly including the following eight categories of attribute features, as shown in Table 1:

[0063] Table 1. Structural Attribute Vectors of Earthquake-Resistant Buildings

[0064]

[0065]

[0066] The attribute extraction method is as follows:

[0067] Main auxiliary detection branch extracts explicit attributes (such as windows, roof, rain pipes): A lightweight branch network is designed next to the YOLO main network to perform multi-object detection tasks and obtain geometric information such as window frames and roof edges;

[0068] Image processing algorithms extract geometric structural attributes: such as using traditional image processing methods like Canny edge detection, Hough line detection, contour analysis, and affine projection to extract structural lines and roof shapes;

[0069] Implicit attributes, such as the number of building floors and window opening ratio, are obtained through image scale conversion and rule matching reasoning based on model prediction and rule calculation.

[0070] Vector normalization encoding: The above 8-dimensional structural attribute values ​​are uniformly normalized to the [0,1] interval to form a structural attribute vector.

[0071] The structural attribute vector is then input into the Structural Attribute Guidance Module (SAPA) as guiding weights for network attention generation, and is fused with the feature maps of the middle layers in the backbone. The SAPA module maps the attribute vector into attention guiding weights through a fully connected layer, and performs channel reweighting and spatial mask modulation operations with the middle layer semantic features (such as layer C3) to achieve structural semantic guidance.

[0072] Through the above methods, the model can obtain guidance from key structural attributes during both the training and inference stages, significantly improving the recognition accuracy in categories with weak structural information such as brick and wood, and earth and wood, and enhancing the network's generalization ability under complex street scene conditions.

[0073] S3. Improved YOLO recognition model construction

[0074] In this embodiment, the building structure type recognition network is designed based on the YOLOv11 backbone architecture. Addressing the challenges of weak structural cues, complex textures, varying scales, and limited deployment resources in street view images, four key structural optimizations were implemented for the YOLOv11 network. First, the C3k2 module in the backbone network was replaced with a windmill-shaped convolutional module (PConv) to more effectively model small target features, expand the receptive field, and reduce overall computational overhead. Second, the lightweight upsampling operator CARAFE was used in the neck network to replace the standard upsample operation of YOLOv11, improving the detail preservation capability of feature map reconstruction. Finally, a structural attribute guidance module (SAPA) was introduced and embedded between the backbone output and the neck fusion layer (located between SPPF and C2PSA) to fuse structural attribute vectors and generate guiding attention weights. Simultaneously, before the head network, i.e., in the feature fusion stage, a multi-scale dilated joint attention (MDJA) module is added. This module utilizes dilated convolutions to capture multi-scale feature information, enhancing the model's receptive field. The classification loss function in YOLOv11 is replaced with the adaptive threshold focus loss function (ATFL), effectively improving the model's accuracy in recognizing building structures. This is particularly suitable for complex environments in street scene images where the target is small, heavily occluded, and has strong background interference, significantly improving the model's robustness and generalization ability. The specific network structure is as follows: Figure 3 As shown.

[0075] The improvements to the YOLO network in this invention mainly include the following four aspects:

[0076] Improvement 1: The backbone convolution is replaced with a windmill-shaped convolution module (c3k2_AP).

[0077] An improved module, C3k2_AP, was introduced into the YOLOv11 backbone network to replace the original C3k2 module (e.g., Figure 4 (As shown). This module effectively expands the receptive field while controlling the number of model parameters by stacking multiple APBottleneck units. The core of the APBottleneck module lies in introducing a pinwheel convolution (PConv) operation to enhance the directionality and local structure awareness of feature representation. The PConv module applies asymmetric padding in four directions (Top, Bottom, Left, Right) to the input feature map in the spatial dimension. This padding method simulates a cross-shaped expansion similar to a "pinwheel," effectively expanding the convolution receptive area and introducing directional prior while maintaining the alignment of the original image. Subsequently, the padding results in each direction are fed into two directional convolution kernels of different shapes for feature extraction. The resulting four sets of directional feature maps are concatenated in the channel dimension to construct an intermediate tensor with enhanced directional information. This tensor is then integrated through a standard fusion convolution (Conv2D+BatchNorm+SiLU) to suppress redundant features and enhance semantic consistency. Finally, to improve the stability of feature representation and gradient flow, the PConv structure introduces a Shortcut residual connection on the main path, which performs element-wise addition on the original input and the fused output to form a complete residual learning path.

[0078] Compared with traditional convolution, PConv constructs multi-directional perception capabilities through cross-shaped extended paths, which not only effectively expands the receptive field, but also significantly improves the robustness and accuracy of the model in recognizing small target building structures in complex street scenes, especially showing stronger discriminative power in cases of blurred edges and uneven scale.

[0079] Improvement point two: Introduce the Structure Attribute Guidance Module (SAPA)

[0080] To enable the network to proactively focus on the facade geometry and structural details most directly related to the seismic load-bearing system, a Structural Property Guidance Module (SAPA) was embedded between the main output and the neck fusion layer.

[0081] The SAPA module specifically includes the following structure:

[0082] Attribute encoding mapping unit: The eight-dimensional structural attribute vector extracted in step S2 is used as input, and the dimension is gradually increased to the same dimension as the number of feature channels in the middle layer through two fully connected networks to generate channel attention weights, and then compressed to the (0,1) interval through the Sigmoid activation function;

[0083] Channel modulation mechanism: The above channel weights are fused with the C3k2 output feature map in the form of channel-by-channel multiplication to realize the prior modulation of the structural attributes on the visual channel response;

[0084] Spatial attention fusion structure: Based on the above channel modulation, a set of continuous convolution operations are introduced, including a 3×3 depthwise separable convolution layer to capture local spatially salient regions, followed by a 1×1 convolution layer for channel dimension integration and compression, and finally outputting a modulation map of the same size as the original feature map.

[0085] Fusion mechanism: The modulation map and the backbone feature map are fused pixel by pixel (either residual connection or weighted summation is acceptable) to form an intermediate representation of the fused attribute prior and image structure features, which is then input into the subsequent FPN path for multi-scale fusion and structure detection.

[0086] This module design enables the network to maintain a high response to structural attribute-dominated areas such as "building height, window ratio, and roof shape" during subsequent inference stages, significantly improving the classification ability and stability of weak texture targets such as brick and wood and earth and wood.

[0087] Improvement point 3: Integrate the content-aware upsampling module (CARAFE)

[0088] To improve the spatial information fidelity and structural semantic restoration capability during the feature map upsampling process, this invention introduces a content-aware reassembly of features (CARAFE) module in the feature fusion stage of the YOLOv11 network to replace the original nearest neighbor interpolation operation.

[0089] The overall structure of the CARAFE module is shown in the attached figure. Figure 5 As shown, it mainly consists of two parts: a convolutional kernel prediction module and a content-aware reconstruction module.

[0090] In the kernel prediction module, the input feature map first undergoes a channel compression operation to reduce the number of redundant channels, and is then input to the content encoder to extract local texture and construction information. Subsequently, the spatial dimension is expanded through the pixel shrug module, and the kernel normalizer outputs a set of dynamically learnable adaptive upsampling kernel weights, with each kernel corresponding to a spatial location in the input feature map.

[0091] In the content-aware reconstruction module, the system extracts local neighborhood image patches centered on each pixel. These patches are then weighted and reconstructed point-by-point with the predicted dynamic convolutional kernel at the i-th position, completing the upsampling and reconstruction operation to generate a feature map with higher resolution and stronger semantic consistency. The CARAFE module offers the following advantages:

[0092] It can adapt to changes in the semantic context of input features, enabling information modeling of asymmetric and irregular regions;

[0093] Without introducing significant computational resource overhead, it effectively enhances the clarity of feature edges and the ability to restore details;

[0094] In particular, it can more clearly restore the structural features of buildings such as edges, window outlines, facade lines, beams and columns in street view images, enhancing the credibility and discriminative power of subsequent classification judgments.

[0095] Improvement point four: Introduce a lightweight multidimensional joint attention module (Lite-MDJA)

[0096] To further enhance structural discrimination capabilities while maintaining model lightweightness, this invention also introduces a lightweight multi-dimensional joint attention module, called the Lite-MDJA module (LiteMulti-DimensionalJointAttention), in each detection branch (i.e. before 11Detect). This module is used to enhance the ability to express structural information at different levels, especially the ability to model subtle structural clues in low-rise old buildings such as brick and wood and earthen buildings.

[0097] This module consists of two main sub-modules: the Lightweight Multi-Scale Spatial Attention Sub-module (LiteGMSA) and the Lightweight Channel Enhancement Sub-module (LiteAGC). The LiteGMSA module performs channel normalization using GroupNorm and introduces channel-by-channel learnable weight vectors. It then uses depthwise separable convolutions to generate a spatial attention map, and finally forms an attention mask using the Silu activation function and a Sigmoid gating mechanism to spatially enhance the input feature map. The LiteAGC module simulates traditional channel attention mechanisms, using two paths for enhancement: one path compresses channels using 1×1 convolutions and then restores the channel count to achieve global perception modeling; the other path introduces grouped convolutions and ReLU nonlinear activations to simulate deep semantic perception. Both paths are multiplied by the input feature map after Sigmoid mapping and then fused with the original input. The entire Lite-MDJA module outputs through a residual structure, enabling the enhanced features to possess stronger structural semantic expressive power without significantly increasing computational cost. This module can effectively improve the ability to distinguish the features of structurally weak targets (such as brick and wood, civil engineering buildings), and is especially suitable for edge deployment scenarios on resource-constrained platforms.

[0098] Through the synergistic effect of the four structural improvement modules, this invention significantly improves the recognition ability of key areas of building structures in street view images while maintaining the original detection speed advantage of the YOLO network, providing a basic guarantee for subsequent model compression and deployment.

[0099] To improve the accuracy of YOLOv11 in identifying seismic structures in complex street scene environments, an Adaptive Threshold Focal Loss (ATFL) function was introduced into the classification branch to replace the original BCEWithLogitsLoss. This loss function is specifically designed to address issues such as severe imbalance between foreground and background samples, blurred target boundaries, and strong background interference in object detection tasks.

[0100] This method makes three key improvements to the original binary cross-entropy loss: foreground-background decoupling calculation, confidence-aware adaptive weighting, and focus learning for ambiguous targets. Regarding foreground-background decoupling, let the original predicted logits of the network output be... The corresponding true label is y∈{0,1} n , where n is the number of samples in each batch. We divide the loss calculation process into two non-overlapping subsets: foreground samples (i.e., y... i =1) and background samples (i.e., y) i =0). This separation mechanism effectively avoids interference from the background dominant gradient on the learning signal of subtle foreground regions, which helps to improve the discrimination performance of small targets and low-contrast structures.

[0101] Regarding confidence-perceived weighting, for each sample i, the confidence calibration probability is first calculated. The definition is as follows:

[0102]

[0103] Where σ(·) represents the Sigmoid activation function, ensuring that… It reflects the confidence level of the model in predicting the current sample.

[0104] Subsequently, a sliding estimate of the batch average confidence level is introduced, and the dynamic focus factor is updated by a weighted combination of the historical mean and the current mean:

[0105]

[0106] The sliding coefficient β = 0.05 is used to smooth the historical confidence level estimate. Then, based on this estimate, the dynamic focusing factor γ is defined. t :

[0107]

[0108] This dynamic factor changes dynamically with the overall confidence level during training, enabling the focusing mechanism to adaptively adjust during the learning process.

[0109] To further emphasize the low-confidence fuzzy samples, a modulation factor m is introduced. i Defined as:

[0110]

[0111] This modulation factor can assign higher training weights to occluded, blurred, or class-bound architectural targets, thereby enhancing the model's ability to learn from difficult samples.

[0112] Finally, the improved Adaptive Threshold Focus Loss Function (ATFL) takes the form of:

[0113]

[0114] in This represents the standard binary cross-entropy loss between predicted logits and the true labels.

[0115] S4. Teacher Model Training

[0116] In this embodiment, the improved YOLO recognition model uses YOLOv11m as its backbone architecture and integrates the windmill-shaped convolutional module, the structure attribute guidance module (SAPA), the content-aware upsampling module (CARAFE), and the lightweight multidimensional joint attention module (Lite-MDJA) proposed in this invention. This network serves as the teacher model, trained on a fully annotated street view building structure image dataset. The resulting model will form the basis for subsequent knowledge distillation and structural pruning.

[0117] Model training was performed on a well-configured high-performance computing platform. The training system ran on a Windows operating system and was developed using the PyTorch framework (version 2.10). The computing platform included an Intel Xeon Gold 6226R @ 2.90GHz processor, 256GB of memory, and an NVIDIA GeForce RTX 4090 graphics card, providing excellent training acceleration and video memory support.

[0118] The training data includes images of various typical street scenes and buildings, labeled according to five types of earthquake-resistant structures, specifically:

[0119] 0-Masonry-concrete_1F (Single-story brick-concrete structure)

[0120] 1-Reinforced-concrete (framework structure)

[0121] 2-Masonry-timber (brick and wood structure)

[0122] 3-Masonry-concrete_ML (Multi-layer brick-concrete structure)

[0123] 4-Earthen-wood (earthenwood structure)

[0124] The training process used the following key hyperparameter configurations: input image size: 640×640 pixels; initial learning rate: 0.001, eventually decaying to 0.01; total training epochs: 100 epochs; batch size: 16; optimizer: AdamW; weight decay factor: 0.0005.

[0125] To enhance the model's generalization ability and robustness in complex street scene scenarios, various data augmentation strategies were introduced during the training phase. Image augmentation included HSV color perturbation (random perturbation of hue, saturation, and brightness within a range of ±15%), Mosaic mosaic augmentation (multiple images stitched together to construct complex backgrounds), MixUp enhancement (image blending and label weighting), and **random erasing**, effectively simulating different lighting conditions, occlusion, and background interference in reality, thereby improving the model's practical performance.

[0126] During training, the IoU threshold was set to 0.7 to filter positive and negative samples, ensuring the reliability of foreground sample selection. Regarding the loss function, the classification branch adopted the proposed Adaptive Threshold Focal Loss (ATFL) function to enhance the model's attention to difficult-to-classify samples (such as civil engineering and brick-and-wood construction) and ambiguous boundary samples; the regression branch used the CIoU (Complete IoU) loss function to improve the localization accuracy of the bounding boxes.

[0127] The model saves a weight snapshot every 10 training rounds. If the validation set metrics do not improve significantly in 50 consecutive training rounds, the early stopping mechanism is automatically triggered to prevent overfitting and shorten the training time.

[0128] The model performance was evaluated using a comprehensive assessment of three mainstream object detection metrics:

[0129] Precision(P): Represents the proportion of correctly identified structural types of targets in the model's detection results;

[0130] Recall(R): Represents the proportion of correctly detected targets out of the actual targets;

[0131] meanAveragePrecision (mAP): Calculated at an IoU threshold of 0.5 (mAP@50) and an IoU range of 0.5–0.95 (mAP@50–95), respectively, comprehensively reflecting the model's localization and classification capabilities.

[0132] The aforementioned metrics comprehensively evaluate the model's accuracy, stability, and generalization performance in the task of identifying seismic-resistant building structure types. After training and optimization, this teacher model will serve as a knowledge source for subsequent distillation and edge deployment of student models.

[0133] S5. Model Compression and Distillation Optimization

[0134] To achieve efficient deployment of the model on edge computing devices and reduce inference costs while maintaining recognition performance, this embodiment further compresses and optimizes the model through knowledge transfer after the teacher model is trained. This mainly includes two stages: structure-aware pruning and knowledge distillation guided by structure regions.

[0135] First, in the structure-aware pruning stage, this invention introduces a channel pruning strategy based on discriminative scoring. This strategy aims to retain key feature channels that contribute significantly to structure recognition while eliminating redundant, weak, or low-relevance channels, thereby reducing the number of model parameters and computational load. The specific method is as follows:

[0136] In the trained teacher model, the output channels of each convolutional layer are evaluated, and their gradient sensitivity to the output probability of each type of structure (brick and wood, earth and wood, frame, etc.) is calculated.

[0137] By combining the mean activation of feature channels in positive samples with the class gradient distribution, a channel importance score is defined:

[0138]

[0139] Where f i This represents the activation feature of the i-th channel. Describes the loss function for category c.

[0140] Set the pruning rate r∈(0,1) to retain the top (1-r) channels with the highest scores, and generate a sparse structure network;

[0141] After pruning, the network is fine-tuned to restore some performance, serving as a lightweight student model structure.

[0142] Subsequently, in the knowledge transfer phase, a structure-region-guided distillation strategy was employed to train and optimize the aforementioned student model, ensuring that it retained its key structural discrimination capabilities while reducing model size. The distillation mechanism used in this phase is as follows:

[0143] Feature distillation: Select the mid-layer feature map of the teacher model (such as C3k2 or SPPF output) as the distillation source, and fit the feature map of the same layer position in the student network using mean squared error (MSE) loss, so that the student model learns the structural semantic expression ability of the teacher model layer by layer.

[0144] Attention distillation: The attention weight map generated by the Structural Attribute Guidance Module (SAPA) is extracted from the teacher model and used as a structural region heatmap for supervision, which constrains the student model to focus on key regions, thereby improving the discrimination performance of the student network on difficult-to-classify structures such as brick and wood and earth.

[0145] Category distillation: KL divergence loss is applied to the classification head output to make the probability distribution of structure type predicted by the student model closer to that of the teacher model, thereby improving the stability of fuzzy category discrimination.

[0146] The entire distillation training process is implemented using the PyTorch framework. The training strategy is basically the same as that of the teacher model, except that the distillation loss and structure heatmap guidance modules are explicitly added. The resulting lightweight student model has the following characteristics:

[0147] The number of model parameters decreased significantly (by approximately 45% to 60%);

[0148] Inference speed is increased by more than 40%, while significantly reducing memory usage and power consumption;

[0149] It can run smoothly on typical edge devices (such as Jetson Xavier NX, mobile terminal CPU+GPU collaborative environment) and supports real-time structure recognition tasks;

[0150] The accuracy metrics (P, R, mAP) are highly similar to those of the teacher model, with the accuracy drop controlled within 3%.

[0151] Through the above-mentioned combined pruning and distillation strategy, this embodiment successfully constructs an edge model that balances deployment performance and recognition capabilities. It is suitable for complex application environments such as urban blocks, rural building surveys, and pre-earthquake investigations, and meets the edge application needs of actual seismic structure surveys.

[0152] S6. Edge Reasoning and Result Output

[0153] In this embodiment, the student model, optimized through pruning, compression, and knowledge distillation, possesses the ability to balance structural recognition accuracy and computational efficiency. It can be deployed on various edge computing devices, including but not limited to mobile processors (such as ARM+GPU), embedded AI devices (such as NVIDIA Jetson Xavier NX), and industrial-grade laptops or in-vehicle computing units. During deployment, the model is loaded into a local inference environment (such as TensorRT, ONNXRuntime, or PyTorchModule) and, combined with a street view image acquisition module, achieves end-to-end automatic identification of seismic-resistant structure types.

[0154] During operation, users acquire street view images via front-end terminals or import existing local images. The system automatically feeds these images into the model for real-time inference. The inference process includes image preprocessing, automatic extraction of structural attribute vectors, feature encoding and fusion, structural type recognition, and result visualization. The overall inference time is controlled within 0.05–0.10 seconds per image, achieving near real-time feedback.

[0155] The model output includes the following:

[0156] Structural type labels: Predicted category numbers for five structural types (e.g., 0–4), which can be mapped to specific types such as “single-story brick-concrete”, “frame structure”, “brick-timber structure”, “multi-story brick-concrete”, and “civil-timber structure”.

[0157] Target location box: The building area is marked using a bounding box, and the returned format is (x_min, y_min, x_max, y_max);

[0158] Confidence score: The probability value (range 0–1) of the model's prediction of the structure, used for subsequent confidence filtering or alarm threshold setting;

[0159] Optional display of structural attributes: including estimated number of floors, window ratio, roof shape and other structural attribute values, which can be used as supplementary information in the output results;

[0160] Output format: Supports exporting to text or spatial formats such as CSV, GeoJSON, and Shapefile, facilitating integration with Geographic Information Systems (GIS) or earthquake information management platforms.

[0161] In addition, the system can be deployed in vehicle-mounted equipment, earthquake inspection terminals, drone platforms and other scenarios to realize inspection-type, remote or high-frequency structure type detection tasks, and batch processing recognition, sliding window inference and continuous image recognition modes can be set as needed to meet different application requirements.

[0162] The inference results can be displayed in real time on the client interface, or uploaded to the server platform or seismic structure survey database via API or Socket, for applications such as pre-earthquake disaster prevention planning, building structure filing, and seismic reinforcement assessment.

[0163] This invention not only features fast response, low resource consumption, and strong adaptability in the edge reasoning stage, but also ensures the recognition accuracy in complex street scene environments, greatly improving the efficiency and wide applicability of seismic structure type data acquisition, and has good practical application value.

[0164] This invention also provides a street view image building seismic structure type recognition system based on an improved YOLO network and knowledge distillation, the structure of which is as follows: Figure 1 As shown, it mainly includes the following six functional modules:

[0165] 1. Street view image acquisition and preprocessing module

[0166] Function Description: Used to acquire street view building images from multiple sources and perform standardized preprocessing and enhancement operations on the image data.

[0167] Input parameters: Raw street view image (formats include JPG / PNG, etc.); acquisition methods include handheld device shooting, vehicle-mounted camera system, drone shooting, street view platform capture, etc.

[0168] Output parameters: A uniformly sized enhanced image (640×640 pixels) in tensor array format for input to subsequent modules;

[0169] Specific operations include brightness normalization, adaptive contrast enhancement, color perturbation (HSV), Mosaic mosaic enhancement, MixUp blending enhancement, and random erasure, which are used to enhance the robustness and adaptability of the model.

[0170] 2. Structural Attribute Extraction and Encoding Module

[0171] Functional description: Extract geometric and semantic feature information related to structure type from images and encode it into a structure attribute vector;

[0172] Input parameter: Preprocessed street view image tensor;

[0173] Output parameters: Structural attribute vector, including structural cues such as floor height estimation, window ratio, roof shape, wall texture, and facade aspect ratio;

[0174] Technical implementation: By combining lightweight window detection, image contour extraction, geometric relationship analysis and shallow CNN auxiliary module, attributes are extracted and normalized and encoded as input to the structural attribute guidance module.

[0175] 3. Improve the YOLO recognition model module

[0176] Function Description: Construct and run an improved YOLOv11 network integrating multiple structural enhancement mechanisms to achieve target detection and classification of building structure types;

[0177] Input parameters: Image tensor + corresponding structural attribute vector;

[0178] Output parameters: predicted label (category) of the building structure target, bounding box location, and confidence score;

[0179] Module composition: The backbone network integrates a windmill-shaped convolutional module (PConv) to enhance local perception; the middle is embedded with a structural attribute guidance module (SAPA) to fuse attribute guidance information; the upsampling path integrates a content-aware upsampling module (CARAFE) to improve feature reconstruction; and a lightweight multidimensional joint attention module (Lite-MDJA) is connected before the detection head to enhance structural discrimination ability.

[0180] Output format: predicted structure type number (0–4), bounding box coordinates, and category confidence score.

[0181] 4. Teacher Model Training Module

[0182] Function Description: Used to train a high-precision YOLO teacher model on a labeled training dataset, providing a knowledge source for subsequent distillation;

[0183] Input parameters: training images and structure labels; hyperparameter settings (learning rate, batch size, number of epochs, etc.);

[0184] Output parameters: The weight file of the trained YOLO teacher model;

[0185] Training features: Adaptive Threshold Focus Loss (ATFL) is used for classification loss, CIoU is used for regression loss, and auxiliary data augmentation strategies are used to enhance the model's generalization ability.

[0186] 5. Model Compression and Distillation Optimization Module

[0187] Function Description: Performs pruning and compression of the teacher model and knowledge transfer learning to obtain a lightweight student model with efficient deployment capabilities;

[0188] Input parameters: teacher model weights, pruning threshold, distillation coefficient;

[0189] Output parameters: Distillation-optimized student model weights;

[0190] Functional details: The pruning stage evaluates channel importance and removes redundant channels; the distillation stage uses structural region heatmaps for attention supervision and employs KL divergence and MSE for feature and class distillation.

[0191] Deployment features: Model parameters are reduced by approximately 50%, inference time is shortened by more than 40%, and it is compatible with edge platforms such as Jetson / mobile GPUs.

[0192] 6. Results Output and Visualization Module

[0193] Function Description: Processes and displays the recognition results, and supports the export of structural data and platform integration;

[0194] Input parameters: Structure type labels, bounding boxes, and confidence scores of the model's inference output;

[0195] Output parameters: Visualized image (overlaying predicted bounding boxes and labels); Structure classification result file (supports CSV, JSON, and GeoJSON formats); Vector data of structural types that can be used for spatial database access;

[0196] Supported features include: real-time display, batch export, API interface upload, and GIS platform embedding.

[0197] The overall system supports multi-source input of street view images, deep recognition driven by structural attributes, lightweight optimization of multi-targets, and flexible deployment output, and is suitable for various application scenarios such as earthquake survey, urban and rural structure filing, and earthquake disaster risk management.

[0198] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying seismic-resistant building structure types based on street view maps and deep learning, characterized in that, Includes the following steps: S1: Collect street view image data of the target area and perform image enhancement and preprocessing operations, including brightness normalization, adaptive contrast enhancement, random occlusion simulation, size scaling and color dithering, to improve network robustness and recognition adaptability. S2: Structural attribute extraction and encoding. The street view image is input into the image processing module to extract geometric and semantic features related to the building structure, including but not limited to the number of building floors, window ratio, aspect ratio, roof shape, wall texture and structural cues, and encode them into structural attribute vectors. S3: Construct an improved YOLO recognition model. The model is based on YOLOv11, integrates a windmill-shaped convolution module in the backbone network, integrates a content-aware upsampling module CARAFE in the feature fusion path, introduces a lightweight multidimensional joint attention module Lite-MDJA in front of the classification detection head, and embeds a structural attribute guidance module in the middle of the backbone network to fuse the aforementioned structural attribute vectors and guide the network to focus on key structural regions. S4: Train the improved YOLO model as a teacher model on the labeled dataset, use the adaptive threshold focus loss function ATFL as the loss function for the classification branch, and combine it with IoU weighted bounding box regression loss to optimize the model; S5: Perform structure-aware pruning and knowledge distillation on the teacher model, filter and prune feature channels based on channel importance scores, construct a lightweight student model, and achieve the transfer and optimization of structure discrimination ability through feature distillation, category distillation and structural region heatmap guidance. S6: Deploy a lightweight student model on an edge computing device. After inputting a street view image, output the building's structure type label, target location box, and confidence score. The inference results are visualized and output as a structure classification data file.

2. A building seismic structure type identification system based on street view maps and deep learning, using the building seismic structure type identification method based on street view maps and deep learning as described in claim 1, characterized in that, include: The street view image acquisition and preprocessing module is used to acquire image data and perform image enhancement and normalization operations; The structural attribute extraction and encoding module is used to extract the geometric attributes and semantic features of buildings from images and encode them into structural attribute vectors. An improved YOLO recognition model module was developed, including a windmill-shaped convolution, a structural attribute guidance module, a multi-dimensional joint attention module, and a content-aware upsampling module, to perform building structure type recognition. The teacher model training module is used to train the improved YOLO model on the labeled dataset and optimizes the classification performance using an adaptive threshold focus loss function. The model compression and distillation optimization module is used to prune the model structure based on channel importance scores and train student models in conjunction with a structure-guided distillation mechanism. The results output and visualization module is used to format and visualize the structure recognition results, and supports the export of structure data or platform integration.

3. The method for identifying seismic-resistant building structure types based on street view maps and deep learning according to claim 1, characterized in that, The structural attribute extraction and encoding process described in step S2 includes the following: Visual feature information related to building structure type is extracted from the preprocessed street view image, including but not limited to the number of building floors, floor-to-height ratio, building width-to-height ratio, window opening ratio, window arrangement pattern, roof shape, wall texture, visible structural lines, and rainwater pipe distribution. Attribute estimation is performed using auxiliary networks or image processing algorithms. The number of building floors is estimated by window detection and linear fitting, the window opening ratio is calculated by the ratio of the target window area to the overall facade area, and the roof shape is determined by the roof edge angle to determine whether it is a flat roof or a sloping roof. The aforementioned attributes are normalized and encoded into structural attribute vectors, which serve as input to the subsequent structural guidance module for channel weighting and spatial attention guidance of the mid-level semantic feature map of the recognition model. The structural attribute vector is input to the structural attribute guidance module SAPA, which generates an attention weight map through a fully connected mapping, modulates the backbone output feature map, and guides the model to focus on structurally significant regions.

4. The method for identifying seismic-resistant building structure types based on street view maps and deep learning according to claim 1, characterized in that, The improved YOLO recognition model described in step S3 includes the following structural improvement modules: The improved YOLO recognition model is based on YOLOv11m as its backbone. Its structure includes a backbone network, a structural attribute guidance module, and a feature fusion and upsampling module. The following structural improvements are introduced at each stage: The windmill-shaped convolutional module PConv replaces the standard convolutional block in the backbone: it applies asymmetric padding in four directions to the input feature map in the spatial dimension. Then, the padding result in each direction is fed into two convolutional kernels of different shapes for feature extraction. The resulting four sets of directional feature maps are concatenated in the channel dimension to construct an intermediate tensor with enhanced directional information. This tensor is then integrated through a standard fusion convolution. Finally, the PConv structure introduces a Shortcut residual connection on the main path to perform element-wise weighting of the original input and the fusion output to form a complete residual learning path. The Structural Attribute Guidance Module (SPA) is embedded between the trunk and the neck. The SPA module specifically includes the following structure: Attribute encoding mapping unit: The eight-dimensional structural attribute vector extracted in step S2 is used as input, and the dimension is gradually increased to the same dimension as the number of feature channels in the middle layer through two fully connected networks to generate channel attention weights, and then compressed to the (0,1) interval through the Sigmoid activation function; Channel modulation mechanism: The above channel weights are fused with the C3k2 output feature map in the form of channel-by-channel multiplication to realize the prior modulation of the structural attributes on the visual channel response; Spatial attention fusion structure: Based on the above channel modulation, a set of continuous convolution operations are introduced, including a 3×3 depthwise separable convolution layer to capture local spatially salient regions, followed by a 1×1 convolution layer for channel dimension integration and compression, and finally outputting a modulation map of the same size as the original feature map. Fusion mechanism: The modulation map and the backbone feature map are fused pixel by pixel to form an intermediate representation of the fusion attribute prior and image structure features, which is then input into the subsequent FPN path for multi-scale fusion and structure detection; The content-aware upsampling module CARAFE replaces the standard upsampling path. The CARAFE module consists of two parts: ① Convolutional kernel prediction module: The input feature map first undergoes channel compression to reduce the number of redundant channels and is then input to the content encoder to extract local texture and construction information; subsequently, the spatial dimension is expanded through the pixel rearrangement module, and a set of dynamically learnable adaptive upsampling kernel weights is output by the convolutional kernel normalizer, with each convolutional kernel corresponding to a spatial location in the input feature map; ② Feature reconstruction module: The system extracts local neighborhood image patches centered on each pixel and performs point-by-point weighted reconstruction with the predicted i-th position dynamic convolutional kernel Wi to complete the upsampling and recombination operation, thereby generating a feature map with higher resolution and stronger semantic consistency. The lightweight multidimensional joint attention module Lite-MDJA is embedded before each detection head: The Lite-MDJA module comprises three substructures: LiteGMSA: It uses GroupNorm for normalization, and generates a spatial weight map by using 1×1 depthwise convolution + SiLU activation and sigmoid to perform positional weighting on the input feature map; LiteAGC consists of an uplink path and a downlink path, both of which are 1×1 convolution → ReLU → convolution. The downlink path uses depthwise convolution (Depthwise Conv) for compressed computation. Residual fusion unit: The attention features of the two paths mentioned above are added to the original input as residuals, and then fused by a 1×1 convolution to improve the structural consistency when detecting small targets; An adaptive threshold focus loss function (ATFL) was introduced in the classification branch to replace the original BCEWithLogitsLoss. This loss function is designed to solve problems such as severe imbalance between foreground and background samples, blurred target boundaries, and strong background interference in object detection tasks. This method makes three key improvements to the original binary cross-entropy loss: foreground-background decoupling calculation, confidence-aware adaptive weighting, and focus learning for fuzzy targets. Regarding foreground-background decoupling, let the original predicted logits of the network output be... The corresponding true label is y∈{0,1} n Where n is the number of samples in each batch, the loss calculation process is divided into two non-overlapping subsets: foreground samples, i.e., y i =1 and background sample, i.e., y i =0; Regarding confidence-perceived weighting, for each sample i, the confidence calibration probability is first calculated. The definition is as follows: Where σ(·) represents the Sigmoid activation function, ensuring that… It reflects the confidence level of the model in predicting the current sample. Subsequently, a sliding estimate of the batch average confidence level is introduced, and the dynamic focus factor is updated by a weighted combination of the historical mean and the current mean: where the sliding coefficient takes β = 0.05 to smooth the historical confidence estimate, and the dynamic focus factor γ is defined according to the estimate t : This dynamic factor changes dynamically with the overall confidence level during training, enabling the focusing mechanism to adaptively adjust during the learning process; To further emphasize the low confidence ambiguous samples, a modulation factor m is introduced i defined as: This modulation factor can assign higher training weights to occluded, blurred, or class-bound architectural targets, thereby enhancing the model's ability to learn from difficult samples. Finally, the improved adaptive threshold focus loss function (ATFL) takes the form of: in This represents the standard binary cross-entropy loss between predicted logits and the true labels.

5. The method for identifying seismic-resistant building structure types based on street view maps and deep learning according to claim 1, characterized in that, The model compression and distillation optimization process described in step S5 includes the following: After the teacher model is trained, model compression and lightweight transfer training are performed based on its complete structure. This mainly includes two stages: structure-aware pruning and knowledge distillation guided by structure regions, in order to improve the deployment efficiency and structure recognition performance of the model on edge devices. The structure-aware pruning stage includes: In the teacher model, the impact of the output channels of each convolutional layer on the classification results of each structure type is evaluated separately. By combining the mean activation value of each feature channel in positive samples with the gradient response of its corresponding class prediction loss function, a channel importance score function is constructed to quantify the contribution of the channel to discriminativity. Set a pruning rate, retain the top n% of key channels in terms of score, delete redundant, weakly correlated or negatively contributing channels, and fine-tune the pruned network structure to restore performance, resulting in a lightweight student model structure. The knowledge distillation stage includes: Feature distillation: Select the mid-layer semantic feature map in the teacher model as the distillation source, and make the corresponding layer output of the student model perform feature matching through mean squared error loss (MSE), thereby improving its ability to represent local structural texture. Attention distillation: Extract the channel attention map and spatial mask generated by the structural attribute guidance module SAPA in the teacher model, and use them as structural region attention supervision signals to guide the student model to focus on the salient areas related to the structure in the building facade; Category distillation: The probability distribution of the classifier head outputs of the teacher and student models is aligned using KL divergence loss, so that the student model can maintain the discriminative boundary characteristics of the teacher model in multi-class recognition; The distillation optimization training was completed using the same hyperparameter settings and data augmentation strategies as the teacher model, and the final generated student model has the following characteristics: The number of model parameters is reduced by more than 45% compared to the original model, the inference speed is increased by no less than 40%, and the runtime memory usage and power consumption are effectively reduced. The model can run smoothly on typical edge computing platforms; The accuracy indicators are basically consistent with the teacher model, and the accuracy decrease is controlled within 3%. It possesses high reliability suitable for seismic building structure type identification tasks and supports edge deployment requirements in multiple scenarios such as urban and rural areas.

6. The method for identifying seismic-resistant building structure types based on street view maps and deep learning according to claim 1, characterized in that, The edge reasoning and result output process described in step S6 includes: The lightweight student model optimized in step S5 is deployed on an edge computing device, which includes, but is not limited to, a mobile terminal processor, an embedded artificial intelligence chip, an in-vehicle terminal, or an industrial-grade edge inference platform. The student model runs in a local inference environment, which is a deployment framework that supports deep learning model inference and is selected from at least one of ONNXRuntime, TensorRT, and PyTorchMobi. The reasoning process includes the following steps: Receive the image to be detected input from the street view image acquisition module, or read existing street view image data locally; Perform image preprocessing operations, including size normalization, brightness / contrast adjustment, and color space conversion; Automatically extract and encode structural attribute vectors, then fuse and modulate them with the image backbone feature map; Multi-scale target detection and structure type identification are completed using the improved YOLO recognition model described in step S3. The output structure recognition results include: a. Structural type prediction labels, ranging from five categories: brick-wood, brick-concrete, frame, civil engineering, and brick-concrete multi-story buildings; b. The target location bounding box, in the format (x_min, y_min, x_max, y_max); c. The prediction confidence level for each target, ranging from 0 to 1; d. Optional outputs: Structural attribute parameters such as estimated number of floors, window ratio, and roof style; e. Supports exporting recognition results to structured formats such as CSV, GeoJSON, and Shapefile; The inference latency is controlled within 0.1 seconds, enabling near real-time processing and supporting operating modes such as batch inference, sliding window detection, and continuous image input. The identification results can be displayed directly on the client terminal, or uploaded to a remote server or seismic structure survey platform via API or network interface, and applied to scenarios such as pre-earthquake risk prediction, building structure information database, and seismic capacity assessment.

Images