A concrete surface crack marking identification method based on human-computer interaction guidance
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHIJIAZHUANG TIEDAO UNIV
- Filing Date
- 2026-05-18
- Publication Date
- 2026-08-07
AI Technical Summary
对于混凝土表面裂缝这种条状病害，由于其边缘模糊、与背景难以区分的特点，人工进行像素级精确标注耗时费力，导致可用的训练样本数量有限，制约了深度模型性能的提升
[0051]本发明方法充分利用人机交互信息的引导识别功能，通过构建裂缝参数识别引导机制，能够有效驱动网络在测试过程中从复杂背景中准确分割出裂缝目标，取得与人工标记真值图更为接近的高质量标记样本，在保持毫秒级实时响应速度的同时，裂缝识别精度和运算量显著优于既有深度学习分割网络。
Smart Images
Figure CN122528048A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and structural health and safety monitoring technology, specifically relating to a method for identifying crack markings on concrete surfaces based on human-computer interaction guidance. Background Technology
[0002] Cracks are a key indicator of early-stage deterioration in concrete structures, and effective detection and identification of crack defects are crucial for assessing structural performance degradation. However, the low signal-to-noise ratio of concrete surface crack images makes sample labeling time-consuming and laborious, hindering the training of high-performance crack detection and identification models. Existing crack detection and identification methods are mainly divided into traditional image processing methods and deep learning methods. Traditional image processing methods suffer from numerous false alarms and missed detections in complex environments and lack generalization ability. Deep learning-based crack segmentation methods, which have emerged in recent years, can automatically extract semantic features through multi-layer convolutional networks, achieving superior recognition results compared to traditional methods. However, current deep learning methods still face the following technical bottlenecks in practical engineering applications:
[0003] (1) High-precision models rely on a large number of high-quality labeled samples. For strip-shaped defects such as cracks on concrete surfaces, due to their blurred edges and difficulty in distinguishing them from the background, manual pixel-level accurate labeling is time-consuming and laborious, resulting in a limited number of available training samples, which restricts the improvement of deep model performance.
[0004] (2) Currently, mainstream crack segmentation networks have a large number of parameters and are computationally complex. For example, commonly used networks such as UNet and DeepLabV3+ have a large number of computational FLOPs and high hardware requirements. They are difficult to deploy directly on field detection terminals or mobile devices with limited computing power, which increases the cost and difficulty of real-time detection.
[0005] (3) Insufficient utilization of prior information from human-computer interaction. Currently, during the manual marking or detection of cracks, operators often draw target bounding boxes or click on several anchor points to indicate the location of cracks. This human-computer interaction information contains relevant information about the crack region and its differences from the background. However, most existing segmentation networks, during training and application, only use the above-mentioned human-computer interaction information as simple region constraints or weak supervision labels, failing to deeply analyze this prior information and organically embed it into the network's feature learning process. Specifically, the statistical differences and contrast relationships between the crack target region implied by the bounding box and its adjacent background region, as well as the crack direction and topological features indicated by the anchor points, are not used as explicit guiding signals to participate in the network's feature learning and decoding reconstruction. Therefore, the model is still prone to false alarms and missed detections when the image signal-to-noise ratio is low and there are water stains, shadows, or other interferences in the background. Summary of the Invention
[0006] To address the problems existing in the prior art, the present invention aims to provide a highly efficient method for marking and identifying cracks on concrete surfaces based on human-computer collaboration. This method parses human-computer interaction information into guiding signals and embeds them into a lightweight network, achieving high-precision automatic crack marking with extremely low computational overhead, significantly reducing the cost of manual marking.
[0007] The specific technical solution adopted in this invention is as follows:
[0008] A method for identifying crack markers on concrete surfaces based on human-computer interaction guidance includes the following steps:
[0009] Step S1: Obtain the original image containing the crack and the corresponding human-computer interaction information;
[0010] The human-computer interaction information includes a target bounding box for selecting the crack target, and multiple anchor points located inside the target bounding box and distributed along the crack direction;
[0011] Step S2: Generate crack parameter identification guidance signals based on human-computer interaction information and original images;
[0012] Step S3: Construct a lightweight recognition network and perform guided training;
[0013] The target parameter-guided AI recognition model is trained using the crack parameter recognition guidance signal obtained in step S2.
[0014] Step S4: Use the target parameter-guided AI recognition model to output crack marking results;
[0015] The concrete surface image with the target bounding box and anchor points is input into the target parameter-guided AI recognition model, which outputs a binary image of crack markers to complete the recognition of crack markers on the concrete surface.
[0016] In step S1, the human-computer interaction information is generated by the operator using a labeling tool, and the target bounding box records the coordinates of its upper left corner (xmin, ymin) and lower right corner (xmax, ymax).
[0017] Step S2 includes the following steps:
[0018] S2.1 Using the target bounding box, automatically determine the crack target region and the local background reference region on the original image;
[0019] Crack target region: Consists of all pixels inside the target bounding box;
[0020] Local background reference region: Based on the target bounding box, expand outward by a preset pixel width d, where d is a natural number, to obtain an outward rectangle. Define the annular region inside the outward rectangle and located outside the original bounding box as the local background reference region.
[0021] S2.2 Extracts two types of identification guidance features from the crack target region and the local background reference region:
[0022] The first category is the regional statistical feature Vstyle, which is used to describe the intrinsic properties of the crack and the background themselves, including at least: the mean pixel brightness of the crack target area, the variance of the pixel brightness of the crack target area, the maximum width of the crack target area, and the mean and variance of the pixel brightness of the local background reference area.
[0023] The second category is local contrast features Vcontext, which is used to describe the difference between the crack and the adjacent background. It includes at least the difference between the mean pixel brightness difference and the difference between the pixel brightness variance of the crack target area and the local background reference area.
[0024] S2.3 Extract the shallow feature map output from the encoding stage of the preset lightweight recognition network. At the same time, encode the original human-computer interaction information to obtain the interaction parameter encoding vector. The interaction parameter encoding vector includes at least the normalized target bounding box coordinates and the average coordinates of all anchor points. After dimensionality reduction, the shallow feature map is concatenated with the interaction parameter encoding vector in the channel dimension to obtain the joint input vector.
[0025] S2.4 The joint input vector is fed into a dynamic gating unit, which consists of a globally average pooling layer, a multilayer perceptron, and a sigmoid activation function connected in sequence. The dynamic gating unit outputs an adaptive weight scalar α with a value between 0 and 1.
[0026] S2.5 Using the aforementioned adaptive weight α, the regional statistical features Vstyle and local contrast features Vcontext extracted in step S2.2 are weighted and fused to generate the final crack parameter identification guidance signal Vfind. The specific calculation formula is as follows:
[0027] ;
[0028] Among them, the regional statistical feature Vstyle and the local contrast feature Vcontext need to be unified to the same dimension through linear mapping before the above weighting operation is performed.
[0029] In step S2.1, after determining the crack target region and the local background reference region on the original image, a three-valued weighted map of the same size as the original image is generated, denoted as T(x,y).
[0030] The target crack target area is assigned a value of 0; the local background reference area is assigned a value of 255 or other non-zero constants; and the remaining far-field background areas are assigned a value of 1.
[0031] Step S3 includes the following steps:
[0032] S3.1 provides an encoder-decoder segmentation network with GhostNet as the backbone network;
[0033] In the encoding stage, convolutional attention modules are embedded after the outputs of each stage of GhostNet; a hollow spatial pyramid pooling module is set at the bottom of the network encoding stage; the decoding stage consists of multi-level upsampling layers that are symmetrical to the structure of the encoding stage, and each level of the decoding layer contains a multi-input concatenated guided decoding module.
[0034] S3.2 For each decoding layer in the decoding phase, perform the following guided embedding operation:
[0035] Channel weighting: The crack parameter identification guidance signal obtained in step S2 is passed through a fully connected layer and a Sigmoid activation function to generate a set of normalized channel weight vectors that are equal to the number of channels in the current skip connection feature map. The skip connection feature map is the output of each coding layer in the coding stage after processing the original image. The skip connection feature map is multiplied channel by channel using the normalized channel weight vector to obtain the channel weighted feature map.
[0036] Spatial weighting: The three-valued weight map generated in step S2.1 is scaled to the same spatial size as the current skip connection feature map through nearest neighbor interpolation; then, the spatial weights corresponding to the crack target region are extracted to obtain a binary spatial focusing mask; the binary spatial focusing mask is multiplied element-wise with the channel-weighted feature map to obtain a feature map after channel and spatial weighting.
[0037] Feature fusion: The feature map, which is weighted by both channel and space, is concatenated with the upsampled feature map from the previous decoding layer in the channel dimension. Then, it is fused through a convolutional layer and used as the output of the current decoding layer. This process continues to propagate upwards. After multiple decoding layers, the final output is the crack prediction probability map P.
[0038] S3.3 During the training of the encoder-decoder network, a joint loss function is constructed for supervision;
[0039] The specific formula for the joint loss function is as follows:
[0040] 1) Boundary loss function LBoundary: The product of the predicted crack probability map P and the Euclidean distance DG of the actual crack boundary is used as a penalty term. The specific loss function is as follows:
[0041] ;
[0042] Where N is the total number of pixels in the input image, and x and y are the horizontal and vertical coordinates of the pixels;
[0043] 2) Topological skeleton loss function LSkeleton: Calculate the Dice difference between the crack prediction skeleton Spred and the true labeled crack skeleton Sgt, and the Dice difference between the crack prediction probability map P and its prediction skeleton Spred; after weighting the above two Dice differences according to the following formula, it is used as the topological skeleton loss value constraint function.
[0044]
[0045] Where β is the weighting coefficient, with a value range of [0.01, 0.1];
[0046] 3) Joint loss function Ltotal: Introducing the basic loss function terms required for regular UNet network training, namely pixel classification loss LBCE and pixel region Dice loss function LDice, and weighting the boundary loss function LBoundary and the topological skeleton loss function LSkeleton to obtain the final loss function term, as shown in the following formula;
[0047] ;
[0048] Wherein, λ1 and λ2 are the weight coefficients of boundary loss and skeleton loss, respectively, and the value range of λ1 and λ2 is set to [0.01, 0.1].
[0049] S3.4 The above encoder-decoder network is iteratively trained using a pre-prepared training sample set until the total loss converges, resulting in a target parameter-guided AI recognition model. Each sample in the training sample set includes an original image, a corresponding pixel-level manually labeled ground truth map, and a corresponding set of human-computer interaction information.
[0050] The beneficial effects of this invention are:
[0051] The method of this invention makes full use of the guidance and recognition function of human-computer interaction information. By constructing a crack parameter recognition guidance mechanism, it can effectively drive the network to accurately segment crack targets from complex backgrounds during the test process, obtain high-quality labeled samples that are closer to the ground truth map labeled by humans, and maintain millisecond-level real-time response speed while crack recognition accuracy and computational load are significantly better than existing deep learning segmentation networks. Attached Figure Description
[0052] Figure 1 This is a schematic diagram of the method of the present invention;
[0053] Figure 2 Network architecture diagram of a target parameter-guided AI recognition model;
[0054] Figure 3 A diagram illustrating the generation process of the guidance signal for crack parameter identification;
[0055] Figure 4 A schematic diagram of a multi-input serial decoding module;
[0056] Figure 5 A graph showing the qualitative comparison results of AI recognition models; Detailed Implementation
[0057] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:
[0058] Specific embodiment 1, such as Figure 1 and Figure 2 As shown, this invention provides a method for identifying crack markers on concrete surfaces based on human-machine collaboration, comprising the following steps:
[0059] Step S1: Obtain the original image containing the crack and the corresponding human-computer interaction information;
[0060] The human-computer interaction information is generated by the operator using annotation tools, including:
[0061] A target bounding box for selecting crack targets, recording the coordinates of its upper left corner xmin, ymin and its lower right corner xmax, ymax;
[0062] Multiple anchor points are located inside the target bounding box and distributed along the crack direction. The operator only needs to click on several points on the crack; each anchor point is used to indicate the crack target itself.
[0063] The original image containing the crack and the corresponding human-computer interaction information together constitute the input for subsequent steps.
[0064] Step S2: Generate crack parameter identification guidance signal based on human-computer interaction information;
[0065] The specific details of this step are as follows:
[0066] S2.1 Using the target bounding box, the crack target region and the local background reference region are automatically determined on the original image.
[0067] Crack target region: Consists of all pixels inside the target bounding box.
[0068] Local background reference region: Based on the target bounding box, an outwardly extended rectangle with a preset pixel width d (d is a natural number, taken as d=10 in this embodiment) is obtained. The annular region inside this extended rectangle and located outside the original target bounding box is defined as the local background reference region. The local background reference region does not overlap with the crack target region. This definition is based on the fact that the local background immediately adjacent to the crack is most similar to the crack region in visual features and is the main source of false alarms, thus serving as the most discriminative reference region.
[0069] To facilitate subsequent spatial guidance, this invention further generates a ternary weight map of the same size as the original image, denoted as T(x,y): pixels inside the target bounding box (crack target region) are assigned a value of 0; the annular local background reference region is assigned a value of 255 (or other non-zero constant); and the remaining far-field background region is assigned a value of 1. The ternary weight map clearly delineates three levels: crack target, adjacent background, and irrelevant region, providing a refined structured prior for spatial focusing in the decoding stage.
[0070] S2.2 Extracts two types of identification guidance features from the crack target region and the local background reference region:
[0071] The first category is regional statistical features (Vstyle), which are used to describe the intrinsic properties of the crack and the background themselves. These include at least: the mean pixel brightness of the crack target area, the variance of the pixel brightness of the crack target area, the maximum width of the crack target area, and the mean and variance of the pixel brightness of the local background reference area.
[0072] The second category is local contrast features (Vcontext), which specifically characterize the difference between the crack and the adjacent background. It includes at least the difference between the mean pixel brightness difference and the variance pixel brightness difference between the crack target area and the local background reference area.
[0073] These features are derived directly from the spatial range defined by human-computer interaction information, providing physically interpretable numerical prior constraints for target recognition.
[0074] S2.3 Extract the shallow feature map output from the encoding stage of the preset lightweight recognition network (e.g., the feature map output from the first or second convolutional layer of the backbone network). Simultaneously, encode the original human-computer interaction information to obtain an interaction parameter encoding vector. This vector contains at least the normalized target bounding box coordinates and the average coordinates of all anchor points. After dimensionality reduction processing, the shallow feature map is concatenated with the interaction parameter encoding vector along the channel dimension to obtain a joint input vector.
[0075] S2.4 The joint input vector is fed into a dynamic gating unit. The dynamic gating unit consists of a globally average pooling layer, a multilayer perceptron (consisting of two fully connected layers and an intermediate activation function), and a sigmoid activation function connected in sequence. This unit outputs an adaptive weight scalar α with a value between 0 and 1.
[0076] S2.5 Utilizes the aforementioned adaptive weight α to perform weighted fusion of the regional statistical features Vstyle and local contrast features Vcontext extracted in step S2.2, generating the final crack parameter identification guidance signal, such as... Figure 3 As shown. The specific calculation formula is:
[0077]
[0078] Among them, the regional statistical feature Vstyle and the local contrast feature Vcontext need to be unified to the same dimension through linear mapping before the above weighting operation is performed.
[0079] Step S3: Construct a lightweight recognition network and perform guided training. This step includes the following:
[0080] S3.1 provides an encoder-decoder segmentation network with GhostNet as its backbone. In the encoding stage, a Convolutional Attention Module (CBAM) is embedded after the outputs of each GhostNet stage to improve the model's ability to perceive cracked targets. At the bottom of the encoding stage, a Spatial Pyramid Pooling Module (ASPP) is set, employing parallel processing of dilated convolutions with different dilation rates to capture multi-scale contextual information. The decoding stage consists of multi-level upsampling layers symmetrical to the encoding stage structure, and each decoding layer contains a multi-input concatenated guided decoding module, such as... Figure 2 As shown.
[0081] S3.2 For each decoding layer in the decoding phase, the following guided embedding operation is performed, such as... Figure 4 As shown:
[0082] Channel weighting: The crack parameter identification guidance signal obtained in step S2 is passed through a fully connected layer and a sigmoid activation function to generate a set of normalized channel weight vectors equal to the number of channels in the current skip connection feature map. The skip connection feature map is the output of each coding layer in the encoding stage after processing the original image. The normalized channel weight vector performs channel-by-channel multiplication on the skip connection feature map from the corresponding stage of the encoder, thereby adaptively enhancing channels related to crack features and suppressing background interference channels in the channel dimension.
[0083] Spatial weighting: The ternary weight map generated in step S2.1 is scaled to the same spatial size as the current skip connection feature map using nearest neighbor interpolation. Then, the spatial weights corresponding to the crack target region are extracted (positions with a value of 0 in the ternary map are weighted to 1, and other positions are weighted to 0), resulting in a binary spatial focusing mask. This binary spatial focusing mask is multiplied element-wise with the channel weighted feature map, forcing the network to focus only on the crack target region selected by the operator during decoding, completely blocking background pixels outside the box, thus forming a precise "rectangular focusing window" in the spatial dimension. Although the local background reference region does not directly participate in spatial weighting, its pixel features are injected into the calculation of the region statistical feature Vstyle and the local contrast feature Vcontext, indirectly guiding the network to distinguish between cracks and adjacent background through prior parameters.
[0084] Feature fusion: The feature map, which is weighted by both channel and spatial dimensions, is concatenated with the upsampled feature map from the previous decoding layer in the channel dimension, and then fused through a convolutional layer as the output of the current decoding layer and continues to be passed up.
[0085] S3.3 describes the construction of a joint loss function for supervision during network training. This joint loss function consists of four parts: pixel classification loss function LBCE, Dice loss function LDice, boundary loss function LBoundary, and topological skeleton loss function LSkeleton. The specific formula is as follows:
[0086] 1) Boundary loss function (LBoundary): The product of the crack prediction probability map P and the Euclidean distance DG of the actual crack boundary is used as a penalty term, which can supervise the network prediction results to be closer to the actual crack boundary range. The specific loss function is as follows.
[0087] ;
[0088] Where N is the total number of pixels in the input image, and x and y are the horizontal and vertical coordinates of the pixels;
[0089] 2) Topology skeleton loss (LSkeleton): Calculate the Dice difference between the predicted crack skeleton Spred and the true labeled crack skeleton Sgt, as well as the Dice difference between the predicted crack probability map P and its predicted skeleton Spred. The two Dice differences are weighted according to the following formula and used as the constraint function for the topology skeleton loss value, which can supervise the network to identify the crack skeleton morphology and the complete content of the target region.
[0090]
[0091] Wherein, β is the weight coefficient, with a value range of [0.01, 0.1], which is determined by automatic iterative updates during network training.
[0092] 3) Joint Loss Function (Ltotal): Introducing the basic loss function terms required for regular UNet network training (i.e., pixel classification loss LBCE and pixel region Dice loss function LDice), and then weighting the boundary loss function LBoundary and the topological skeleton loss function LSkeleton to obtain the final loss function term, as shown in the following formula.
[0093]
[0094] Here, λ1 and λ2 are the weight coefficients of the boundary loss and skeleton loss, respectively. In actual network training, the loss functions LBCE and LDice are responsible for the global gradient update of the network, while the loss functions LBoundary and LSKeleton serve as regularization constraints. If the weight values are set too large, it can easily lead to Euclidean distance gradient explosion or excessive shrinkage of skeleton features. To achieve more stable model optimization training, the range of values for the two weight parameters λ1 and λ2 is set to [0.01, 0.1], and the specific values are determined through iterative updates during network training. Specifically, λ = 0.01 + 0.09σ(w), where σ(*) is the Sigmoid function, and w is a free parameter automatically learned by the network. This design avoids the difficulty of manual parameter tuning and effectively prevents problems such as boundary distance gradient explosion or excessive shrinkage of the skeleton.
[0095] S3.4 The network is iteratively trained using a pre-prepared training sample set until the total loss converges. Each sample in the training sample set includes an original image, a corresponding pixel-level manually labeled ground truth image, and a set of corresponding human-computer interaction information. The AdamW optimizer is used during training, with an initial learning rate ranging from [1×10⁻⁻⁴]. 6 1×10⁻ 4 The specific learning rate is dynamically adjusted using a cosine annealing strategy. The batch size is set to 4, the maximum number of training epochs is 2000, and an early stopping strategy is used to prevent overfitting during training.
[0096] Pixel-level manually labeled ground truth images are binary images created by manually and precisely outlining the actual crack regions on the original image. In these images, the crack pixel value is 255, and the background pixel value is 0. When input into a deep learning network for loss calculation, this image is normalized, mapping the actual crack pixel value to 1 and the background pixel value to 0. Manually labeled ground truth images play a crucial supervisory role during the training phase, comparing the results with the model's predictions to optimize training parameters.
[0097] The Euclidean distance DG of the true crack boundary for calculating the boundary loss, the true labeled crack skeleton Sgt of the topological skeleton loss, and the calculation benchmarks of the basic classification loss functions LBCE and LDice are all directly calculated or derived from this ground truth map.
[0098] Step S4: Output crack marking results
[0099] For an image of a concrete surface to be processed, the operator only needs to provide simple human-computer interaction information (i.e., draw a target bounding box and click several anchor points along the crack). The image and interaction information are then input into a trained lightweight guided recognition network, which can automatically output a high-precision pixel-level binary image of crack markings, where crack pixels are marked as 1 and background pixels as 0.
[0100] Specific Implementation Example 2, Effect Verification, as follows: Figure 5 As shown;
[0101] Images of cracks in subway tunnels were selected as experimental data, totaling 14,410 original images (512×512 resolution), covering various working conditions such as dryness, water seepage, mud stains, and shadows. To meet the training requirements of the AI recognition network, the labeled samples were randomly divided in an 8:1:1 ratio to obtain training, validation, and test sample sets.
[0102] To verify the effectiveness of the proposed target parameter guided AI recognition model (GCS-Net), three representative deep learning networks in the field of image semantic segmentation—UNet, UNet++, and DeepLabV3+—were selected for comparative experiments. All comparative experiments were conducted in a uniformly configured computing environment, specifically an Intel i7-12700K processor and an NVIDIA RTX 4080 GPU with 16GB of RAM. The target parameter guided AI recognition model proposed in this application is implemented using the PyTorch 2.3.1 computing framework.
[0103] Quantitative comparison results show that:
[0104] UNet has a computational cost of 21.71 GFLOPs, a response time of 4.86 ms, and a Dice coefficient of 0.711.
[0105] The computational cost of UNet++ is 64.11 GFLOPs, the response time is 8.37 ms, and the Dice coefficient is 0.723.
[0106] DeepLabV3+ has a computational cost of 31.70 GFLOPs, a response time of 7.10 ms, and a Dice factor of 0.742.
[0107] The computational cost of the method of this invention is only 3.06 GFLOPs, the response time is 16.49 ms, and the Dice coefficient reaches 0.829.
[0108] Qualitative comparison results show that the UNet and DeepLabV3+ models are prone to missing detections in fine crack details because they lose a significant amount of shallow spatial detail during multiple downsampling processes. The UNet and UNet++ models, lacking an attention-based computation mechanism, are more sensitive to background noise, leading to false alarms in the recognition results. In summary, the method of this invention fully utilizes the guided recognition function of human-computer interaction information. By constructing a crack parameter recognition guidance mechanism, it can effectively drive the network to accurately segment crack targets from complex backgrounds during testing, achieving high-quality labeled samples that more closely approximate the ground truth image.
[0109] In summary, the target parameter-guided AI recognition model proposed in this invention achieves a Dice segmentation accuracy of 0.829 under lightweight conditions with a computational cost of only 3.06 GFLOPs. While maintaining a millisecond-level real-time response speed, its crack recognition accuracy is significantly better than that of existing deep learning segmentation networks.
Claims
1. A method for identifying crack markers on concrete surfaces based on human-computer interaction guidance, characterized in that, Includes the following steps: Step S1: Obtain the original image containing the crack and the corresponding human-computer interaction information; The human-computer interaction information includes a target bounding box for selecting the crack target, and multiple anchor points located inside the target bounding box and distributed along the crack direction; Step S2: Generate crack parameter identification guidance signals based on human-computer interaction information and original images; Step S3: Construct a lightweight recognition network and perform guided training; The target parameter-guided AI recognition model is trained using the crack parameter recognition guidance signal obtained in step S2. Step S4: Use the target parameter-guided AI recognition model to output crack marking results; The concrete surface image with the target bounding box and anchor points is input into the target parameter-guided AI recognition model, which outputs a binary image of crack markers to complete the recognition of crack markers on the concrete surface.
2. The method for identifying crack markers on concrete surfaces based on human-computer interaction guidance according to claim 1, characterized in that: In step S1, the human-computer interaction information is generated by the operator using a labeling tool, and the target bounding box records the coordinates of its upper left corner (xmin, ymin) and lower right corner (xmax, ymax).
3. The method for identifying crack markers on concrete surfaces based on human-computer interaction guidance according to claim 1, characterized in that: Step S2 includes the following steps: S2.1 Using the target bounding box, automatically determine the crack target region and the local background reference region on the original image; Crack target region: Consists of all pixels inside the target bounding box; Local background reference region: Based on the target bounding box, expand outward by a preset pixel width d, where d is a natural number, to obtain an outward rectangle. Define the annular region inside the outward rectangle and located outside the original bounding box as the local background reference region. S2.2 Extracts two types of identification guidance features from the crack target region and the local background reference region: The first category is the regional statistical feature Vstyle, which is used to describe the intrinsic properties of the crack and the background themselves, including at least: the mean pixel brightness of the crack target area, the variance of the pixel brightness of the crack target area, the maximum width of the crack target area, and the mean and variance of the pixel brightness of the local background reference area. The second category is local contrast features Vcontext, which is used to describe the difference between the crack and the adjacent background. It includes at least the difference between the mean pixel brightness difference and the difference between the pixel brightness variance of the crack target area and the local background reference area. S2.3 Extract the shallow feature map output from the encoding stage of the preset lightweight recognition network. At the same time, encode the original human-computer interaction information to obtain the interaction parameter encoding vector. The interaction parameter encoding vector includes at least the normalized target bounding box coordinates and the average coordinates of all anchor points. After dimensionality reduction, the shallow feature map is concatenated with the interaction parameter encoding vector in the channel dimension to obtain the joint input vector. S2.4 The joint input vector is fed into a dynamic gating unit, which consists of a globally average pooling layer, a multilayer perceptron, and a sigmoid activation function connected in sequence. The dynamic gating unit outputs an adaptive weight scalar α with a value between 0 and 1. S2.5 Using the aforementioned adaptive weight α, the regional statistical features Vstyle and local contrast features Vcontext extracted in step S2.2 are weighted and fused to generate the final crack parameter identification guidance signal Vfind. The specific calculation formula is as follows: ； Among them, the regional statistical feature Vstyle and the local contrast feature Vcontext need to be unified to the same dimension through linear mapping before the above weighting operation is performed.
4. The method for identifying crack markers on concrete surfaces based on human-computer interaction guidance according to claim 3, characterized in that: In step S2.1, after determining the crack target region and the local background reference region on the original image, a three-valued weighted map of the same size as the original image is generated, denoted as T(x,y). The target crack target area is assigned a value of 0; the local background reference area is assigned a value of 255 or other non-zero constants; and the remaining far-field background areas are assigned a value of 1.
5. The method for identifying crack markers on concrete surfaces based on human-computer interaction guidance according to claim 3, characterized in that: Step S3 includes the following steps: S3.1 provides an encoder-decoder segmentation network with GhostNet as the backbone network; In the encoding stage, convolutional attention modules are embedded after the outputs of each stage of GhostNet; a hollow spatial pyramid pooling module is set at the bottom of the network encoding stage; the decoding stage consists of multi-level upsampling layers that are symmetrical to the structure of the encoding stage, and each level of the decoding layer contains a multi-input concatenated guided decoding module. S3.2 For each decoding layer in the decoding phase, perform the following guided embedding operation: Channel weighting: The crack parameter identification guidance signal obtained in step S2 is passed through a fully connected layer and a Sigmoid activation function to generate a set of normalized channel weight vectors that are equal to the number of channels in the current skip connection feature map. The skip connection feature map is the output of each coding layer in the coding stage after processing the original image. The skip connection feature map is multiplied channel by channel using the normalized channel weight vector to obtain the channel weighted feature map. Spatial weighting: The three-valued weight map generated in step S2.1 is scaled to the same spatial size as the current skip connection feature map through nearest neighbor interpolation; then, the spatial weights corresponding to the crack target region are extracted to obtain a binary spatial focusing mask; the binary spatial focusing mask is multiplied element-wise with the channel-weighted feature map to obtain a feature map after channel and spatial weighting. Feature fusion: The feature map, which is weighted by both channel and space, is concatenated with the upsampled feature map from the previous decoding layer in the channel dimension. Then, it is fused through a convolutional layer and used as the output of the current decoding layer. This process continues to propagate upwards. After multiple decoding layers, the final output is the crack prediction probability map P. S3.3 During the training of the encoder-decoder network, a joint loss function is constructed for supervision; The specific formula for the joint loss function is as follows: 1) Boundary loss function LBoundary: The product of the predicted crack probability map P and the Euclidean distance DG of the actual crack boundary is used as a penalty term. The specific loss function is as follows: ； Where N is the total number of pixels in the input image, and x and y are the horizontal and vertical coordinates of the pixels; 2) Topological skeleton loss function LSkeleton: Calculate the Dice difference between the crack prediction skeleton Spred and the true labeled crack skeleton Sgt, and the Dice difference between the crack prediction probability map P and its prediction skeleton Spred; after weighting the above two Dice differences according to the following formula, it is used as the topological skeleton loss value constraint function. ； Where β is the weighting coefficient, with a value range of [0.01, 0.1]; 3) Joint loss function Ltotal: Introducing the basic loss function terms required for regular UNet network training, namely pixel classification loss LBCE and pixel region Dice loss function LDice, and weighting the boundary loss function LBoundary and the topological skeleton loss function LSkeleton to obtain the final loss function term, as shown in the following formula; ； Wherein, λ1 and λ2 are the weight coefficients of boundary loss and skeleton loss, respectively, and the value range of λ1 and λ2 is set to [0.01, 0.1]. S3.4 The above encoder-decoder network is iteratively trained using a pre-prepared training sample set until the total loss converges, resulting in a target parameter-guided AI recognition model. Each sample in the training sample set includes an original image, a corresponding pixel-level manually labeled ground truth map, and a corresponding set of human-computer interaction information.