Highway pavement disease intelligent detection method and system based on multi-source data fusion and YOLO optimization algorithm

Abstract

The application discloses a kind of based on multi-source data fusion and YOLO optimization algorithm's highway pavement disease intelligent detection method and system.The method is by collecting multi-source data set, by time synchronization, spatial registration and resolution standardization processing, constructs unified data input set;Using the multi-scale feature extraction network of fusion deformable convolution and attention mechanism enhances the feature capture ability to irregular disease;Establish cross-modal feature interaction mechanism, realize visible light texture, infrared thermodynamics and the deep fusion of point cloud geometric feature;Using multi-objective optimization loss function carries out end-to-end training, improves model adaptability;Finally realize the automatic identification, positioning and classification of crack, pit and other diseases, and generate the structured detection report containing disease distribution, statistical evaluation and repair suggestion.The application effectively improves detection precision and efficiency, and provides reliable support for road intelligent maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road detection, in particular to a highway pavement disease intelligent detection method and system based on multi-source data fusion and YOLO optimization algorithm. BACKGROUND

[0002] As a core component of the national transportation network, the health condition of the highway pavement is directly related to driving safety, transportation efficiency and maintenance cost. The traditional manual visual detection method is not only inefficient and subjective, but also has safety hazards, and it is difficult to meet the rapid and accurate detection needs of large-scale road network.

[0003] In recent years, automated detection technology based on computer vision and deep learning has been widely applied, especially the target detection algorithm represented by YOLO, which has shown great potential in pavement disease identification. However, the existing methods still face many challenges: first, most of them rely on a single visible light sensor, and the detection performance drops sharply in complex environments such as light changes, shadow obstructions or water accumulation; second, the irregular shape and large scale difference of diseases such as cracks and potholes make the feature extraction and fusion ability of traditional convolutional neural networks insufficient, resulting in high miss detection rate for small targets and irregular defects; third, there is a lack of effective multi-source data fusion mechanism, and the complementary advantages of multi-modal data such as infrared thermal imaging and three-dimensional laser point cloud in representing internal abnormalities and accurate geometric shapes cannot be fully utilized.

[0004] In addition, the detection results of existing systems often lack precise association with geographic information, making it difficult to directly generate structured reports that can be used for maintenance decision-making, and there is a gap in the application chain from "detection" to "decision".

[0005] Therefore, there is an urgent need for an integrated intelligent detection method that can fuse multi-source information, adapt to complex environments, and achieve accurate identification and decision support. SUMMARY

[0006] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a highway pavement disease intelligent detection method and system with high detection accuracy, strong adaptability and multi-source data fusion.

[0007] In a first aspect, the present application provides a highway pavement disease intelligent detection method based on multi-source data fusion and YOLO optimization algorithm, which comprises:

[0008] S1, synchronously collecting multi-source data of road surface by multi-spectral imaging, infrared thermal sensing and three-dimensional laser scanning device, and performing time synchronization, space registration and resolution standardization processing to establish a registered multi-source data input set;

[0009] S2, a feature extraction framework based on a convolutional neural network is constructed, multi-scale feature fusion and deformable convolution are adopted to enhance the feature capturing capability for irregular diseases, and multi-level feature integration is realized through an attention mechanism and a feature pyramid;

[0010] S3, a multi-modal feature interaction mechanism is established, visible light texture features, infrared thermodynamic features and point cloud geometric features are deeply fused, and channel attention and spatial transformation are used to realize complementary enhancement of cross-modal features;

[0011] S4, an end-to-end training strategy is adopted, combined with a multi-scale supervision mechanism and a multi-objective optimization function,

[0012] the network parameters are optimized through gradient back propagation;

[0013] S5, the trained deep learning model is used to process multi-source data, and disease type classification, position coordinates and detection confidence are output;

[0014] S6, a structured detection report is generated based on the recognition result, including disease distribution map, quantitative statistical data and repair suggestions.

[0015] Optionally, in an implementation form of the first aspect of the present application, the time synchronization calibration, space coordinate unification and resolution standardization processing in S1 adopt a cascaded registration method based on feature matching, specifically including:

[0016] A high-precision GPS time module and a hardware trigger signal are used to realize microsecond-level time synchronization of multi-sensor data;

[0017] A group of reflective targets with known three-dimensional coordinates are arranged in the detection area to establish a unified spatial reference coordinate system;

[0018] Based on feature point extraction and descriptor matching algorithm, the feature points of multispectral images, infrared hot spots and laser point clouds are associated and registered;

[0019] Multi-resolution pyramids and interpolation algorithms are used to resample multispectral images and infrared thermal imaging data to a unified spatial resolution, and laser point clouds are grid processed to generate depth images;

[0020] Finally, a multi-source data cube with time alignment, spatial registration and unified resolution is generated.

[0021] Optionally, in an implementation form of the first aspect of the present application, the multi-scale feature fusion module in S2 adopts a hierarchical aggregation architecture, specifically including:

[0022] A deformable convolution layer is introduced in the encoder part, and the convolution kernel is adaptively covered on the irregular defect area through a learnable offset parameter;

[0023] A bidirectional feature pyramid network is designed to fuse low-level detail features and high-level semantic features through top-down and bottom-up paths, and a spatial attention mechanism is introduced at each fusion node.

[0024] An incremental upsampling strategy is adopted in the decoder part, and the multi-scale features of the encoder are fused with the corresponding level features of the decoder through a jump connection.

[0025] Finally, a multi-scale feature map containing rich context information is output.

[0026] Optionally, in an implementation form of the first aspect of the present application, the deformable convolution layer adopts an adaptive receptive field adjustment mechanism, specifically including:

[0027] An offset learning branch is introduced in each convolution layer to learn two-dimensional spatial offsets through a standard convolution layer, so that the convolution kernel can adaptively adjust the sampling position according to the input features;

[0028] A bilinear interpolation algorithm is used to realize feature value sampling of non-integer coordinates, ensuring the spatial continuity of the offset feature mapping;

[0029] An offset constraint mechanism is set to limit the range of offsets through a regularization term to prevent unstable feature extraction;

[0030] A gradient clipping strategy is used in the training process to ensure the training stability of the offset learning branch;

[0031] By stacking multi-scale deformable convolution components, a feature extraction network with adaptive spatial perception ability is constructed.

[0032] Optionally, in an implementation form of the first aspect of the present application, the multi-modal feature fusion in S3 adopts a multi-level semantic guidance and attention enhancement strategy, specifically including:

[0033] A multi-level semantic guidance decoder is introduced to realize effective fusion of the shallow detail features of the encoder and the deep semantic features of the decoder through a jump fusion module;

[0034] A multi-dimensional attention module is constructed to integrate channel attention blocks, global attention blocks and multi-scale attention blocks, wherein:

[0035] The channel attention block adopts a double-path structure to generate channel weights;

[0036] The global attention block realizes spatial attention modeling through batch normalization and parallel paths;

[0037] The multi-scale attention block adopts a multi-branch asymmetric depth separable convolution to capture multi-scale features;

[0038] The visible light texture feature, the infrared thermal anomaly feature and the point cloud geometry feature are integrated through a weighted fusion mechanism to form an enhanced composite feature representation;

[0039] A deep supervision training strategy is adopted, and intermediate outputs of multiple decoding stages are used for auxiliary training.

[0040] Optionally, in an implementation form of the first aspect of the present application, the cross-modal feature fusion in S3 adopts a complementary enhancement strategy based on an attention mechanism, specifically including:

[0041] The visible light texture feature, the infrared thermal anomaly feature and the point cloud geometry feature are integrated through a weighted fusion mechanism to fully utilize the complementary advantages of each modal data;

[0042] A channel reweighting mechanism is adopted to dynamically adjust the weight coefficients of feature channels according to the importance of different modal features;

[0043] A spatial attention transformation is introduced to highlight the feature response of important regions through a spatial weight map and suppress background interference;

[0044] Finally, an enhanced composite feature representation is formed to improve the feature learning ability of the model for complex road defects and the cross-modal fusion effect.

[0045] Optionally, in an implementation form of the first aspect of the present application, the optimization objective function adopted in S4 is a multi-feature fusion loss function, and its basic expression is:

[0046] L fusion =η·L feature +(1-η)·w γ ·L reconstruction ,

[0047] Wherein, L feature is a feature preservation loss component, L reconstruction is a reconstruction accuracy loss component, η is a fusion coefficient, and w γ is a spatial distribution weight coefficient;

[0048] The feature preservation loss component adopts a multi-level feature similarity calculation method:

[0049]

[0050] Wherein, S is the number of feature levels, and respectively represent the estimated result and the true result at the jth feature level, and FSM(·) is a feature similarity measurement function;

[0051] The reconstruction accuracy loss component adopts an absolute value difference calculation method:

[0052]

[0053] wherein D1, D2, D3 represent three dimensions of data respectively, R(d1, d2, d3) and G(d1, d2, d3) represent values of reconstructed data and real data at position (d1, d2, d3) respectively;

[0054] The calculation formula of the feature similarity index FSM is as follows:

[0055]

[0056] wherein μ A , μ B are mean values of data blocks A and B, are variances of data blocks A and B, σ AB is a covariance of data blocks A and B, E1 and E2 are stabilization parameters.

[0057] Optionally, in an implementation form of the first aspect of the present application, the intelligent identification and positioning of defects in S5 comprises:

[0058] Multi-scale defect detection from fine-grained cracks to large-area pits and grooves is achieved by using the enhanced composite features after multi-scale feature map and cross-modal fusion;

[0059] A GPU-accelerated parallel processing pipeline is established to perform real-time synchronous processing on multi-source data, and complete the whole process in the mobile detection process;

[0060] A classification head based on attention mechanism is used to accurately classify cracks, pits and grooves, repaired areas and ruts, and a regression prediction head is used to output the boundary box coordinates and direction angle information of the defects;

[0061] The detection results from three modalities of visible light, infrared and point cloud are confidence weighted and fused;

[0062] The multi-modal detection results are verified for spatial consistency by using a unified spatial coordinate system, and false detection and repeated detection are eliminated;

[0063] The identification threshold is dynamically adjusted according to environmental conditions, a non-maximum suppression algorithm is used to optimize the detection results, and the type label, accurate geographic coordinates, size information and detection confidence of the defects are output.

[0064] Optionally, in an implementation form of the first aspect of the present application, the detection result generation and output of S6 specifically comprises:

[0065] The defect identification result is fused and visualized with the original multi-source data to generate a comprehensive defect distribution map containing visible light images, infrared thermal maps and three-dimensional point cloud data, and the defect types are labeled with differentiated colors.

[0066] Perform multi-dimensional statistical analysis on defect information, including statistics on the distribution by type, severity and spatial location, and generate a heat map of defect distribution at the road segment level;

[0067] Establish a maintenance urgency assessment model and conduct priority assessment by combining defect size, confidence level, and distribution characteristics;

[0068] Generate a structured test report, including an execution summary, detailed test data, statistical charts, and maintenance recommendations.

[0069] Secondly, embodiments of this application provide an intelligent detection system for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm, applied to the intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm as described in the first aspect. The system includes:

[0070] The data acquisition module is used to simultaneously acquire multi-source data of the road surface through a multispectral imaging device, an infrared thermal sensing device, and a three-dimensional laser scanning device;

[0071] The data preprocessing module is used to perform time synchronization calibration, spatial coordinate unification, and resolution standardization on the collected multi-source data to establish a fully registered multi-source data input set;

[0072] The feature extraction module is used to build a feature extraction framework based on convolutional neural networks. It adopts multi-scale feature fusion and deformable convolution to enhance the feature capture capability of irregular diseases, and realizes multi-level feature integration through attention mechanism and feature pyramid.

[0073] The feature fusion module is used to establish a multimodal feature interaction mechanism, which deeply fuses visible light texture features, infrared thermodynamic features and point cloud geometric features, and achieves cross-modal feature complementarity enhancement through channel attention and spatial transformation;

[0074] The model training module is used to optimize network parameters through gradient backpropagation by employing an end-to-end training strategy, combining multi-scale supervision mechanisms and multi-objective optimization functions.

[0075] The disease identification module is used to process multi-source data using a trained deep learning model and output disease type classification, location coordinates, and detection confidence.

[0076] The report generation module is used to generate structured inspection reports based on the identification results, including disease distribution maps, quantitative statistical data, and maintenance recommendations.

[0077] Thirdly, embodiments of this application provide an electronic device, including:

[0078] processor;

[0079] Memory used to store processor-executable instructions;

[0080] The processor is configured to implement the intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm as described in the first aspect when executing the instructions.

[0081] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm as described in the first aspect.

[0082] This invention discloses an intelligent detection method and system for highway pavement defects based on multi-source data fusion and the YOLO optimization algorithm. The method simultaneously acquires multi-source data from the road surface using multispectral imaging, infrared thermal sensing, and 3D laser scanning devices. After time synchronization, spatial registration, and resolution standardization, a unified data input set is constructed. A multi-scale feature extraction network integrating deformable convolution and attention mechanisms is employed to enhance the feature capture capability for irregular defects. A cross-modal feature interaction mechanism is established to achieve deep fusion of visible light texture, infrared thermodynamics, and point cloud geometric features. End-to-end training is performed using a multi-objective optimization loss function to improve model adaptability. Finally, the method achieves automated identification, location, and classification of defects such as cracks and potholes, generating a structured detection report containing defect distribution, statistical assessment, and maintenance suggestions. This invention effectively improves detection accuracy and efficiency, providing reliable support for intelligent road maintenance.

[0083] The beneficial effects of this invention are as follows: through multi-source data fusion and deep learning optimization, high-precision, fully automatic, and real-time detection of highway pavement defects is achieved, significantly improving the identification accuracy and positioning accuracy of defects such as cracks and potholes; the method is highly adaptable, can effectively cope with complex road environments, greatly improve detection efficiency, and provide quantitative and visualized data support for maintenance decisions, reducing labor costs and subjective errors. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of a method for intelligent detection of highway pavement defects based on multi-source data fusion and YOLO optimization algorithm, provided as an embodiment of this application.

[0085] Figure 2 This is a structural diagram of a deep learning model provided in an embodiment of this application.

[0086] Figure 3 The architecture diagram of the intelligent detection system for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm provided in this application.

[0087] Figure 4 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0088] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0089] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0090] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, terms such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiments or designs described as exemplary or for example in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present related concepts in a specific manner.

[0091] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0092] Example 1

[0093] Figure 1 This is a schematic flowchart illustrating a method for intelligent detection of highway pavement defects based on multi-source data fusion and YOLO optimization algorithm, provided as an embodiment of this application. Figure 1 As shown, a smart detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm includes:

[0094] S1. Multi-source data of the road surface is simultaneously acquired using multispectral imaging, infrared thermal sensing, and 3D laser scanning devices. This data is then processed through temporal synchronization, spatial registration, and resolution standardization to establish a fully registered multi-source data input set. This provides a standardized data foundation for subsequent fusion and recognition.

[0095] Specifically, in this embodiment, the time synchronization calibration, spatial coordinate unification, and resolution standardization processing in S1 adopts a cascaded registration method based on feature matching, which specifically includes:

[0096] A high-precision GPS timing module and hardware trigger signals are used to achieve microsecond-level time synchronization of multi-sensor data acquisition. Specifically, through multi-sensor time synchronization, a high-precision GPS timing module and hardware trigger signals are used to achieve microsecond-level time synchronization of multispectral, infrared, and laser sensor data acquisition, ensuring strict alignment of different modal data in the time dimension and providing timing consistency assurance for subsequent fusion.

[0097] A unified spatial reference coordinate system is established by arranging a group of reflective targets with known three-dimensional coordinates in the detection area. These reflective targets utilize calibration plates with specific geometric patterns and retro-reflective materials. This unified spatial coordinate system provides a common spatial reference for multi-source data, resolving the issue of inconsistent spatial reference systems among different sensors.

[0098] Based on feature point extraction and descriptor matching algorithms, feature points from multispectral images, infrared hotspots, and laser point clouds are correlated and registered. Cross-modal feature point correlation and registration achieves accurate spatial correspondence between different modal data.

[0099] A multi-resolution pyramid and interpolation algorithm is employed to resample multispectral images to a uniform spatial resolution, convert infrared thermal imaging data to the same pixel spacing as the multispectral images, and generate depth images by meshing laser point cloud data. By standardizing the resolution of multi-source data, resolution inconsistencies caused by sensor differences are eliminated.

[0100] The final result is a multi-source data cube that is temporally aligned, spatially registered, and has a uniform resolution, providing standardized input for subsequent feature fusion. Integrating the aforementioned processing results, a multi-source data cube that is temporally aligned, spatially registered, and has a uniform resolution is finally generated, providing standardized, high-quality data input for subsequent feature fusion and disease identification.

[0101] S2. Construct a feature extraction framework based on convolutional neural networks. Employ multi-scale feature fusion and deformable convolution to enhance the feature capture capability for irregular diseases. Achieve multi-level feature integration through attention mechanisms and feature pyramids. Through adaptive feature extraction and fusion, construct a feature extraction framework based on convolutional neural networks. Utilize multi-scale feature fusion and deformable convolution to enhance the feature capture capability for irregular diseases (such as cracks and pits). Achieve multi-level feature integration through attention mechanisms and feature pyramids to improve the discriminative power of feature representation.

[0102] Specifically, in this embodiment, the multi-scale feature fusion module in S2 adopts a hierarchical aggregation architecture, specifically including:

[0103] Deformable convolutional layers are introduced in the encoder section. Through learnable offset parameters, the convolutional kernels adaptively cover irregular defect areas, enhancing the model's ability to perceive features of irregular morphological defects such as crack edges and pit contours, and improving the flexibility of feature extraction.

[0104] The deformable convolutional layer employs an adaptive receptive field adjustment mechanism, specifically including: introducing an offset learning branch in each convolutional layer. This branch learns a two-dimensional spatial offset through a standard convolutional layer, enabling the convolutional kernel to adaptively adjust the sampling position according to the specific shape of the input features, thereby dynamically adjusting the receptive field to better match the geometric structure of irregular diseases. A bilinear interpolation algorithm is used to sample feature values ​​for non-integer coordinates, ensuring accurate feature value calculation at the sampling points after applying the learned offset, maintaining the spatial continuity of the feature mapping, and avoiding information distortion or grid effects caused by coordinate offsets. By introducing constraints such as regularization terms, the learning range of the offset is limited to prevent excessive offsets during training, which could lead to unstable feature extraction or convergence difficulties, thus ensuring the robustness of the model's training.

[0105] During training, a gradient truncation strategy is employed for the offset learning branch to control the gradient explosion problem, ensuring the training stability of this branch and enabling the network to effectively learn reasonable spatial transformations. By stacking multiple multi-scale deformable convolutional components, a deep feature extraction network with adaptive spatial awareness is constructed, enabling the network to progressively capture complex and irregular geometric features of road defects from local to global perspectives, significantly improving the model's ability to represent morphologically variable defects.

[0106] A bidirectional feature pyramid network is designed, fusing low-level detailed features and high-level semantic features through bidirectional paths from top to bottom and bottom to top. A spatial attention mechanism is introduced at each fusion node to dynamically adjust feature weights at different spatial locations, achieving effective integration and enhancement of multi-scale features. Progressive upsampling and skip connections are employed. In the decoder section, a progressive upsampling strategy is used, while skip connections are used to deeply fuse the multi-scale features of the encoder with the corresponding layer features of the decoder. This is combined with depthwise separable convolutions to reduce computational complexity, improving feature representation while preserving detailed information. The output is a multi-scale feature map, ultimately containing rich contextual information. This provides a feature representation with strong discriminative and multi-scale perception capabilities for subsequent defect recognition tasks, supporting detection needs from fine-grained cracks to large-area pits.

[0107] By employing deformable convolutions to adapt to irregular shapes, bidirectional feature pyramids to fuse multi-scale information, and skip connections to preserve detailed features, a feature extraction network was constructed that can effectively capture and integrate multi-scale, irregular road disease features, significantly improving the model's ability to represent complex diseases.

[0108] S3. Establish a multimodal feature interaction mechanism to deeply fuse visible light texture features, infrared thermodynamic features, and point cloud geometric features. Through channel attention and spatial transformation, cross-modal feature complementarity enhancement is achieved, overcoming the limitations of single-modal data.

[0109] Specifically, in this embodiment, the multimodal feature fusion in S3 employs a multi-level semantic guidance and attention enhancement strategy, specifically including:

[0110] A multi-level semantic-guided decoder is introduced, which effectively fuses shallow detail features from the encoder with deep semantic features from the decoder through a skip fusion module. This skip fusion module employs three parallel paths to process detail features, semantic features, and attention weights for the generation channel, respectively. The introduction of this multi-level semantic-guided decoder enhances the richness and discriminative power of feature representation.

[0111] A multi-dimensional attention module is constructed, which sequentially integrates channel attention blocks, global attention blocks, and multi-scale attention blocks, respectively implementing: Channel attention block: using a dual-path structure (adaptive global average pooling and adaptive global max pooling) to generate channel weights, highlighting important feature channels; Global attention block: realizing spatial attention modeling through batch normalization and parallel path structure, focusing on key regions; Multi-scale attention block: using multi-branch asymmetric depthwise separable convolution to capture multi-scale features, enhancing the model's ability to perceive diseases at different scales.

[0112] By integrating visible light texture features, infrared thermal anomaly features, and point cloud geometric features through a weighted fusion mechanism, an enhanced composite feature representation is formed. This fully leverages the complementary advantages of each modality of data to improve the robustness and information integrity of feature representation.

[0113] A deep supervised training strategy is adopted, which utilizes the intermediate outputs of multiple decoding stages for auxiliary training, thereby improving the model's ability to learn features of complex road defects, accelerating convergence and improving model performance, and ensuring the stability and effectiveness of the training process.

[0114] The cross-modal feature enhancement and fusion of S3 adopts a fusion strategy based on multi-level guidance and attention mechanisms, specifically including:

[0115] By employing a weighted fusion mechanism, visible light texture features, infrared thermal anomaly features, and point cloud geometric features are integrated. This fully leverages the complementary advantages of each modal data in terms of surface texture, internal thermodynamic properties, and three-dimensional geometry, resulting in a more comprehensive composite feature representation. Specifically, visible light texture features provide high-resolution surface texture information; infrared thermal anomaly features reflect differences in the thermodynamic properties within the material; and point cloud geometric features provide accurate three-dimensional spatial geometric information.

[0116] A channel reweighting mechanism is adopted to dynamically adjust the weight coefficients of feature channels according to the importance of different modal features. This dynamic adjustment of feature channel weight coefficients highlights information-rich modal features and suppresses redundant or noisy feature channels, thereby achieving adaptive feature selection and enhancement.

[0117] By introducing spatial attention transformation, the feature responses of important regions are highlighted through spatial weight maps, background interference is suppressed, and the model pays more attention to potential disease areas, thereby improving the signal-to-noise ratio of feature expression and the ability to perceive key regions.

[0118] The final result is an enhanced composite feature representation. A deep supervised training strategy is employed, utilizing intermediate outputs from multiple decoding stages for auxiliary training. This improves the model's feature learning ability for complex road defects and its cross-modal feature fusion performance. The resulting enhanced composite feature representation effectively enhances the model's feature learning ability for complex road defects and its cross-modal feature fusion performance, providing more discriminative feature inputs for subsequent high-precision defect identification.

[0119] By employing weighted fusion to achieve modal complementarity, channel reweighting to achieve adaptive feature selection, and spatial attention to highlight key regions, an efficient and adaptive cross-modal feature enhancement strategy was constructed. This strategy significantly improves the quality and efficiency of multi-source feature fusion and enhances the model's ability to perceive and discriminate feature defects in complex road environments.

[0120] S4. An end-to-end training strategy is adopted, combining multi-scale supervision mechanisms and multi-objective optimization functions, and the network parameters are optimized through gradient backpropagation. This end-to-end model optimization training strategy, combined with multi-scale supervision mechanisms and multi-objective optimization functions, optimizes network parameters through gradient backpropagation, improving the model's adaptability to complex road environments and its detection accuracy.

[0121] When fusing visible light, infrared, and point cloud data, the model needs to simultaneously achieve two seemingly contradictory goals: Goal 1 (high-level semantic consistency): The fused features must retain and highlight the semantic information most effective for identifying defects in each modality (e.g., crack texture, thermal anomalies of potholes, geometric depth of ruts). The model should not fuse for the sake of fusion, leading to feature "distortion" or "homogenization." Goal 2 (low-level detail authenticity): The fusion process must retain as much fine detail as possible from the original data. For example, pixel-level edges of minute cracks, subtle thermal gradient changes, etc., are crucial for accurate location and assessment of defect severity. A single traditional loss function (such as using only L1 / L2 reconstruction loss) cannot adequately satisfy both goals simultaneously; therefore, a composite loss function needs to be designed.

[0122] In S4, a multi-feature fusion loss function is used for network training. This loss function is composed of a weighted feature preservation loss component and a reconstruction accuracy loss component. The feature preservation loss component is calculated through multi-level feature similarity to maintain the feature consistency of the image. The reconstruction accuracy loss component is calculated through absolute value difference to ensure the quality of detail reconstruction. The loss function also introduces a fusion coefficient to balance the weights of the two losses and uses a spatial distribution weight coefficient to adjust the contribution of the reconstruction accuracy loss according to the pixel spatial location.

[0123] Specifically, in this embodiment, the optimization objective function used in S4 is a multi-feature fusion loss function, the basic expression of which is:

[0124] L fusion =η·L feature +(1-η)·w γ ·L reconstruction ,

[0125] Among them, L feature The feature-preserving loss component is used to maintain the feature consistency of the image;

[0126] L reconstruction The component representing the reconstruction accuracy loss is used to ensure the quality of detailed reconstruction; η is a fusion coefficient, ranging from 0.6 to 0.9, used to balance the weights of feature preservation loss and reconstruction accuracy loss, w γThe spatial distribution weights are used to adjust the contribution of reconstruction accuracy loss based on pixel spatial location. A composite loss function is constructed, consisting of a weighted sum of feature preservation loss components and reconstruction accuracy loss components, to simultaneously optimize feature consistency and detail reconstruction quality during training, guiding the model to learn more accurate and robust feature representations.

[0127] The feature preservation loss component employs a multi-level feature similarity calculation method:

[0128]

[0129] Where S is the number of feature levels. and Let FSM(·) represent the estimated result and the true result at the j-th feature level, respectively. FSM(·) is the feature similarity measurement function. By calculating feature similarity at multiple levels, the structural consistency of the image in the feature space is maintained, ensuring that the fused features retain the key attributes of the original data and preventing feature distortion.

[0130] The reconstruction accuracy loss component is calculated using the absolute value difference method:

[0131]

[0132] Here, D1, D2, and D3 represent the three dimensions of the data, and R(d1,d2,d3) and G(d1,d2,d3) represent the values ​​of the reconstructed data and the real data at positions (d1,d2,d3), respectively. By calculating the absolute value difference, we ensure that the reconstruction result is as close as possible to the real data at the pixel level, thereby improving the model's ability to preserve detailed information and the accuracy of the reconstruction result.

[0133] The formula for calculating the Feature Similarity Index (FSM) is as follows:

[0134]

[0135] Where, μ A μ B Let A and B be the mean of the data blocks. Let σ be the variance of data blocks A and B. AB Let E1 be the covariance of data blocks A and B, and E2 be stabilization parameters used to prevent the denominator from being zero. Their values ​​depend on the dynamic range of the data. E1 = (k1L) 2 E2 = (k2L) 2 , where L is the dynamic range of the data values, k1 = 0.01, k2 = 0.03.

[0136] The spatial distribution weighting coefficient w γ :

[0137]

[0138] Where q represents the coordinates of the current pixel, q0 represents the coordinates of the image center, and γ is the standard deviation parameter of the Gaussian distribution, controlling the spatial range of the weight distribution. The contribution of the reconstruction accuracy loss is adjusted according to the pixel's spatial location, making the model focus more on the reconstruction quality of important regions (such as potential defect areas), thus improving training efficiency and model performance in key regions.

[0139] The fusion coefficient η is adaptively determined in the following manner:

[0140]

[0141] Where, δ feature δ reconstruction These represent the standard deviations of the feature preservation loss and reconstruction accuracy loss on the validation dataset. An adjustable fusion coefficient is used to dynamically balance the weights of the feature preservation loss and reconstruction accuracy loss, adjusting the optimization direction according to different tasks or data characteristics to enhance the adaptability and flexibility of the loss function.

[0142] The above composite loss function is designed to achieve the following four core advantages: 1. Dual protection:

[0143] Simultaneously constraining high-level semantics and low-level details ensures that the fused features are both "useful" (facilitating recognition) and "realistic" (preserving details). 2. Adaptive balancing: through coefficients η and weights w γ This allows the model to intelligently allocate optimization priorities, rather than mechanically executing a single objective. 3. Human-centered design: The FSM function is designed to align with human visual perception; L1 Loss tends to preserve sharp edges, and Gaussian weights reflect the cognitive focus on the road surface center, making the optimization process closer to real-world application needs. 4. Stable training: The stabilization parameters E1 and E2 in FSM, along with potential gradient truncation of offsets, collectively ensure that the training process of such a complex multi-component loss function is stable and convergent.

[0144] The design of this loss function fully considers the core challenges of multi-source data fusion tasks. It is not a simple summation of losses, but a systematic and collaborative optimization framework designed to guide the model to learn a fusion feature representation that is both highly discriminative and rich in detail, thereby laying a solid foundation for subsequent high-precision disease identification.

[0145] S5. Utilize the trained deep learning model to process multi-source data, outputting disease type classification, location coordinates, and detection confidence. Through intelligent disease identification and localization, the trained model processes multi-source data, outputting disease type classification, precise location coordinates, and detection confidence, achieving automated, high-precision identification and localization of typical road diseases. For example...Figure 2 The diagram shown is a structural diagram of a deep learning model provided in an embodiment of this application.

[0146] Specifically, in this embodiment, the intelligent defect identification and localization of S5 includes:

[0147] Multi-scale defect detection mechanism: The multi-scale feature maps extracted by the adaptive feature extraction network constructed in S2 are then input into the deep learning model trained in S4 after cross-modal feature enhancement and fusion to achieve multi-scale defect detection from fine-grained cracks to large-area pits, thereby improving the model's ability to identify defects of different sizes.

[0148] Real-time processing pipeline: Establish a GPU-accelerated parallel processing pipeline to perform real-time synchronous processing of multi-source data collected by S1, ensuring that the entire process of data acquisition, feature extraction, defect identification and localization is completed while the mobile inspection platform is in operation.

[0149] Defect Classification and Location: An attention-based classification head is used to accurately classify defect types, including four categories: cracks, potholes, repair areas, and ruts. At the same time, a regression prediction head outputs the bounding box coordinates and orientation angle information of the defects, enabling accurate identification and location of defects.

[0150] Confidence fusion strategy: Confidence-weighted fusion of detection results from visible light, infrared, and point cloud modalities is performed. The specific formula is as follows:

[0151] C final =w v ·C v +w t ·C t +w p ·C p ,

[0152] Among them, C v C t C p w represents the detection confidence levels for visible light, infrared, and point cloud modes, respectively. v w t w p The weighting coefficients for each modality are used; the detection results from the visible light, infrared, and point cloud modalities are weighted and fused with confidence levels to comprehensively utilize the advantages of each modality and improve the reliability of the final result.

[0153] Spatial Consistency Verification: Utilizing the unified spatial coordinate system established in S1, spatial consistency verification is performed on the multimodal detection results to eliminate false detections and duplicate detections, ensuring positioning accuracy reaches the centimeter level. Adaptive Threshold Adjustment: The recognition threshold is dynamically adjusted based on ambient lighting conditions, road surface materials, and detection speed to ensure stable detection performance under different working conditions. Result Optimization and Post-processing: The detection results are optimized using a non-maximum suppression algorithm, overlapping detection boxes are merged, and the results are refined based on the geometric features and spatial distribution of the defects. Finally, the defect type label, precise geographic coordinates, size information, and detection confidence level are output.

[0154] By employing a multi-scale detection mechanism, a real-time processing pipeline, multi-modal confidence fusion, spatial consistency verification, and adaptive optimization strategies, a complete, efficient, and reliable intelligent identification and localization system for road defects was constructed, enabling accurate and real-time detection of various types of road defects.

[0155] S6. Generate a structured inspection report based on the identification results, including a defect distribution map, quantitative statistical data, and maintenance recommendations. Through structured report generation and decision support, a structured inspection report containing a defect distribution map, quantitative statistical data, and maintenance recommendations is generated based on the identification results, providing data support for road maintenance decisions and achieving effective transformation of inspection results into maintenance actions.

[0156] Specifically, in this embodiment, the generation and output of the detection result in step S6 includes:

[0157] Multimodal data fusion display: Based on the unified spatial coordinate system established in S1, the defect identification results output by S5 are fused and visualized with the original multi-source data to generate a comprehensive defect distribution map containing visible light images, infrared thermal maps and three-dimensional point cloud data, in which different types of defects are marked with different colors;

[0158] Quantitative statistical analysis: Multi-dimensional statistical analysis is performed on the defect information obtained from S5 detection, including:

[0159] Statistical analysis was performed on the number, area, and distribution density of defects by type; statistical analysis was also performed on a severity-based basis.

[0160] The S5 output is used to classify the detection confidence level and defect size; a road segment-level defect distribution heatmap is generated based on spatial location distribution statistics; maintenance priority assessment: a maintenance urgency assessment model is established by combining defect type, size, confidence level, and spatial distribution characteristics.

[0161] P = η1·S + η2·C + η3·D,

[0162] Where S is the defect size score, C is the confidence score, D is the distribution density score, and η1, η2, and η3 are weighting coefficients.

[0163] Structured Report Generation: Generates industry-standard inspection reports, including: Executive Summary: Highlighting major defects and urgent repair needs; Detailed Inspection Data: Includes the precise location, size, type, and image evidence of all defects; Statistical Charts: Defect distribution histograms, trend curves, and comparative analysis; Repair Recommendations: Providing specific repair plans and construction suggestions based on priority assessment; Data Traceability and Correlation: Establishing a correlation between inspection results and the original data collected by S1, supporting retrospective verification of inspection results, with each defect traceable to corresponding multi-source data sources; Decision Support Function: Providing trend analysis based on historical data, evaluating defect development trends and maintenance effectiveness by comparing multiple inspection results, providing data support for long-term maintenance planning; Automated Output Interface: Supports one-click export of inspection results, with output formats including PDF reports, Excel spreadsheets, GIS map files, and standard data interfaces for integration with road maintenance management systems.

[0164] By employing multimodal visualization, multidimensional statistical analysis, intelligent priority assessment, structured report generation, and systematic data management, a complete system for generating and outputting test results has been constructed, enabling full-process support from data to decision-making and providing a comprehensive and reliable solution for road maintenance management.

[0165] Example 2

[0166] like Figure 3 As shown, this application provides an architecture diagram of an intelligent detection system for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm. It is applied to the intelligent detection system for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm as described in Embodiment 1, including a data acquisition module 11, a data preprocessing module 12, a feature extraction module 13, a feature fusion module 14, a model training module 15, a defect identification module 16, and a report generation module 17.

[0167] Data acquisition module 11 is used to simultaneously acquire multi-source data of the road surface through a multispectral imaging device, an infrared thermal sensing device, and a three-dimensional laser scanning device;

[0168] Data preprocessing module 12 is used to perform time synchronization calibration, spatial coordinate unification and resolution standardization on the collected multi-source data to establish a fully registered multi-source data input set;

[0169] Feature extraction module 13 is used to construct a feature extraction framework based on convolutional neural network. It adopts multi-scale feature fusion and deformable convolution to enhance the feature capture capability of irregular diseases, and realizes multi-level feature integration through attention mechanism and feature pyramid.

[0170] The feature fusion module 14 is used to establish a multimodal feature interaction mechanism, which deeply fuses visible light texture features, infrared thermodynamic features and point cloud geometric features, and achieves cross-modal feature complementarity enhancement through channel attention and spatial transformation;

[0171] Model training module 15 is used to optimize network parameters through gradient backpropagation by adopting an end-to-end training strategy, combining a multi-scale supervision mechanism and a multi-objective optimization function.

[0172] The disease identification module 16 is used to process multi-source data using a trained deep learning model and output disease type classification, location coordinates and detection confidence.

[0173] The report generation module 17 is used to generate a structured inspection report based on the identification results, which includes a disease distribution map, quantitative statistical data, and maintenance recommendations.

[0174] Figure 4 This is an electronic device provided in one embodiment of this application. For example... Figure 4 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0175] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0176] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0177] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0178] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0179] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0180] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0181] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0182] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for intelligent detection of pavement defects on highways based on multi-source data fusion and YOLO optimization algorithm, characterized in that, Includes the following steps: S1. Multi-source data of road surface are collected simultaneously through multispectral imaging, infrared thermal sensing and three-dimensional laser scanning devices, and time synchronization, spatial registration and resolution standardization are performed to establish a fully registered multi-source data input set; S2. Construct a feature extraction framework based on convolutional neural networks, employing multi-scale feature fusion and deformable convolution to enhance the feature capture capability for irregular diseases. Multi-level feature integration is achieved through attention mechanisms and feature pyramids. The multi-scale feature fusion adopts a hierarchical aggregation architecture, specifically including: Deformable convolutional layers are introduced in the encoder section, and the convolutional kernels adaptively cover irregular defect areas through learnable offset parameters; We design a bidirectional feature pyramid network that fuses low-level detail features and high-level semantic features through top-down and bottom-up paths, and introduces a spatial attention mechanism at each fusion node. In the decoder section, a progressive upsampling strategy is adopted, which uses skip connections to deeply fuse the multi-scale features of the encoder with the corresponding hierarchical features of the decoder. The final output is a multi-scale feature map containing rich contextual information; S3. Establish a multimodal feature interaction mechanism to deeply fuse visible light texture features, infrared thermodynamic features and point cloud geometric features, and achieve cross-modal feature complementarity enhancement through channel attention and spatial transformation; S4. An end-to-end training strategy is adopted, combining a multi-scale supervision mechanism and a multi-objective optimization function. The network parameters are optimized through gradient backpropagation. The optimization objective function used in S4 is a multi-feature fusion loss function, the basic expression of which is: ， in, To preserve the loss components for the features, To reconstruct the accuracy loss components, The fusion coefficient is... Spatial distribution weighting coefficients; The feature preservation loss component employs a multi-level feature similarity calculation method: ， in, The number of feature levels, and They represent the first Estimated results and true results at each feature level This is a feature similarity measurement function; The reconstruction accuracy loss component is calculated using the absolute value difference method: ， in, These represent the three dimensions of the data. and These represent the reconstructed data and the actual data in terms of location, respectively. The value at that location; The feature similarity index The calculation formula is: ， in, Let A and B be the mean of the data blocks. Let A and B be the variances. Let the covariance of data blocks A and B be... For stabilization parameters; S5. Use the trained deep learning model to process multi-source data and output disease type classification, location coordinates and detection confidence. S6. Generate a structured inspection report based on the identification results, including a disease distribution map, quantitative statistical data, and maintenance suggestions.

2. The intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm according to claim 1, characterized in that, The time synchronization calibration, spatial coordinate unification, and resolution standardization processing in S1 employs a cascaded registration method based on feature matching, specifically including: Microsecond-level time synchronization of multi-sensor data is achieved by using a high-precision GPS timing module and hardware trigger signals; A unified spatial reference coordinate system is established by arranging a group of reflective targets with known three-dimensional coordinates in the detection area; Based on feature point extraction and descriptor matching algorithms, feature points of multispectral images, infrared hot spots and laser point clouds are associated and registered. Multi-resolution pyramid and interpolation algorithms are used to resample multispectral images and infrared thermal imaging data to a uniform spatial resolution, and laser point clouds are meshed to generate depth images. The final result is a multi-source data cube that is time-aligned, spatially registered, and has uniform resolution.

3. The intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm according to claim 2, characterized in that, The deformable convolutional layer employs an adaptive receptive field adjustment mechanism, specifically including: An offset learning branch is introduced into each convolutional layer. The two-dimensional spatial offset is learned through the standard convolutional layer, enabling the convolutional kernel to adaptively adjust the sampling position according to the input features. A bilinear interpolation algorithm is used to sample the feature values ​​of non-integer coordinates to ensure that the offset feature mapping maintains spatial continuity. An offset constraint mechanism is set up to limit the range of the offset through a regularization term to prevent unstable feature extraction. A gradient truncation strategy is employed during training to ensure the training stability of the offset learning branch. By stacking multi-scale deformable convolutional components, a feature extraction network with adaptive spatial awareness is constructed.

4. The intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm according to claim 1, characterized in that, The multimodal feature fusion in S3 employs a multi-level semantic guidance and attention enhancement strategy, specifically including: A multi-level semantic-guided decoder is introduced, and the shallow detail features of the encoder and the deep semantic features of the decoder are effectively fused through a skip fusion module. Construct a multi-dimensional attention module that integrates channel attention blocks, global attention blocks, and multi-scale attention blocks, where: The channel attention block uses a dual-path structure to generate channel weights; Global attention blocks achieve spatial attention modeling through batch normalization and parallel paths; Multi-scale attention blocks employ multi-branch asymmetric depthwise separable convolutions to capture multi-scale features; Visible light texture features, infrared thermal anomaly features, and point cloud geometric features are integrated through a weighted fusion mechanism to form an enhanced composite feature representation; A deep supervised training strategy is adopted, which utilizes the intermediate outputs of multiple decoding stages for auxiliary training.

5. The intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm according to claim 1, characterized in that, The cross-modal feature fusion in S3 employs a complementary enhancement strategy based on an attention mechanism, specifically including: By integrating visible light texture features, infrared thermal anomaly features, and point cloud geometric features through a weighted fusion mechanism, the complementary advantages of each modality of data can be fully utilized. A channel reweighting mechanism is adopted to dynamically adjust the weight coefficients of feature channels according to the importance of different modal features; A spatial attention transformation is introduced to highlight the feature responses of important regions through a spatial weight map, thereby suppressing background interference. Ultimately, this results in an enhanced composite feature representation, improving the model's ability to learn features about complex road defects and its cross-modal fusion performance.

6. The intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm according to claim 1, characterized in that, The intelligent defect identification and localization in S5 includes: By utilizing multi-scale feature maps and enhanced composite features obtained through cross-modal fusion, multi-scale defect detection from fine-grained cracks to large-area pits can be achieved. Establish a GPU-accelerated parallel processing pipeline to perform real-time synchronous processing of multi-source data and complete the entire process during motion detection. An attention-based classification head is used to accurately classify cracks, potholes, repair areas, and ruts, and a regression prediction head outputs the bounding box coordinates and orientation angle information of the defects. Confidence-weighted fusion is performed on the detection results from three modalities: visible light, infrared, and point cloud. A unified spatial coordinate system is used to verify the spatial consistency of multimodal detection results, thereby eliminating false detections and duplicate detections. The identification threshold is dynamically adjusted according to environmental conditions, and the detection results are optimized using a non-maximum suppression algorithm. The output includes the defect type label, accurate geographic coordinates, size information, and detection confidence level.

7. The intelligent detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm according to claim 1, characterized in that, The generation and output of the detection results in S6 specifically includes: The defect identification results are fused and visualized with the original multi-source data to generate a comprehensive defect distribution map that includes visible light images, infrared thermal maps and three-dimensional point cloud data, and the defect types are marked with differentiated colors. Perform multi-dimensional statistical analysis on defect information, including statistics on the distribution by type, severity and spatial location, and generate a heat map of defect distribution at the road segment level; Establish a maintenance urgency assessment model and conduct priority assessment by combining defect size, confidence level, and distribution characteristics; Generate a structured test report, including an execution summary, detailed test data, statistical charts, and maintenance recommendations.

8. A smart detection system for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm, applied to the smart detection method for highway pavement defects based on multi-source data fusion and YOLO optimization algorithm as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition module is used to simultaneously acquire multi-source data of the road surface through a multispectral imaging device, an infrared thermal sensing device, and a three-dimensional laser scanning device; The data preprocessing module is used to perform time synchronization calibration, spatial coordinate unification, and resolution standardization on the collected multi-source data to establish a fully registered multi-source data input set; The feature extraction module is used to build a feature extraction framework based on convolutional neural networks. It adopts multi-scale feature fusion and deformable convolution to enhance the feature capture capability of irregular diseases, and realizes multi-level feature integration through attention mechanism and feature pyramid. The feature fusion module is used to establish a multimodal feature interaction mechanism, which deeply fuses visible light texture features, infrared thermodynamic features and point cloud geometric features, and achieves cross-modal feature complementarity enhancement through channel attention and spatial transformation; The model training module is used to optimize network parameters through gradient backpropagation by employing an end-to-end training strategy, combining multi-scale supervision mechanisms and multi-objective optimization functions. The disease identification module is used to process multi-source data using a trained deep learning model and output disease type classification, location coordinates, and detection confidence. The report generation module is used to generate structured inspection reports based on the identification results, including disease distribution maps, quantitative statistical data, and maintenance recommendations.

Images