Millimeter wave image dry hanging stone curtain wall metal pendant position detection system and method

Abstract

The invention provides a millimeter wave image dry-hanging stone curtain wall metal pendant position detection system and method, and the method comprises the following steps: designing a lightweight point-like target detection neural network, inputting a millimeter wave image I based on a U-Net architecture and in combination with a MobileNet module, and outputting a probability graph P of key point positions; designing a point loss function, and optimizing network training through weighted summation; performing advanced feature extraction and matching by using a spatial transformation network and comparison loss, automatically correcting the position and the scale of a key point region, and enhancing the similarity of feature vectors; and sub-pixel-level positioning is realized by adopting a Gaussian fitting method, and the high-reflection bright spots of the metal pendant are accurately positioned below the pixel level. According to the method, through lightweight network design, point loss function optimization, sub-pixel-level positioning and post-processing technologies, the detection efficiency, precision and robustness of the method are obviously superior to those of an existing method, and efficient and reliable technical support is provided for engineering application of millimeter wave images in the field of curtain wall detection.

Description

Technical Field

[0001] The present invention relates to the technical field of building detection, and in particular to a millimeter wave image dry-hanging stone curtain wall metal hanger position detection system and method. Background Art

[0002] In millimeter-wave imaging, metal materials, due to their high reflectivity, often appear as bright spots in images. Existing object detection methods, such as YOLO and Faster R-CNN, mostly use bounding boxes to mark and detect objects. These technologies are not only used for millimeter-wave image processing but are also widely applied to other imagery, such as optical and infrared, to locate the position and size of objects. However, in millimeter-wave imagery, metal pendants are primarily detected as point-like, highly reflective bright spots, rather than requiring complex information about the target's outer contour.

[0003] Deficiencies of existing technology: 1. Bounding box redundancy: Traditional object detection methods rely on bounding boxes to locate and identify objects. This approach is inefficient for point-like, highly reflective metal pendants. Because metal pendants appear as isolated bright spots in millimeter-wave images rather than solid objects with clear boundaries, using bounding boxes not only adds unnecessary computational complexity but can also lead to inaccurate positioning.

[0004] 2. Resolution Limitations and Noise Interference: Millimeter-wave imaging typically has lower resolution than optical or infrared images, making it difficult to clearly image small metal pendants. Combined with environmental noise and interference from the equipment itself, traditional algorithms are prone to missed detections or misjudgments, reducing detection reliability.

[0005] 3. Real-time performance and resource consumption: Existing deep learning-based object detection algorithms, such as YOLO and Faster R-CNN, perform well in many scenarios, but they require significant computational resources and time to process millimeter-wave imagery. Given that real-world applications (such as building facade inspections) often require rapid response times, the high computational cost of existing algorithms becomes a significant bottleneck.

[0006] Therefore, the existing technology has deficiencies and needs further improvement. Summary of the Invention

[0007] In response to the problems existing in the prior art, the present invention provides a millimeter wave image dry-hanging stone curtain wall metal hanger position detection system and method.

[0008] To achieve the above object, the specific solutions of the present invention are as follows: The present invention provides a method for detecting the position of metal hangers on a dry-hanging stone curtain wall based on millimeter wave images, the method comprising the following steps: S1, design a lightweight point target detection neural network MetalPointNet based on the U-Net architecture and combined with the MobileNet module. It inputs the millimeter wave image I and outputs a probability map P of key point locations, where P(i, j) represents the probability that the pixel (i, j) is a key point of a metal pendant. S2, design a point loss function, which includes positioning error loss, heatmap regression loss and L2 regularization term, and optimizes the training of the point target detection neural network through weighted summation; S3 uses the spatial transformer network (STN) and contrastive loss function for advanced feature extraction and matching, automatically corrects the position and scale of key point areas, and enhances feature vector similarity; S4, uses Gaussian fitting method to achieve sub-pixel positioning, accurately locating the highly reflective bright spots of metal pendants below the pixel level; S5, post-processing optimization of the detection results, removing noise and false positives through heat map adaptive thresholding and DBSCAN clustering algorithm, retaining valid detection points and visualizing candidate boxes.

[0009] Furthermore, in step S2, the total loss function of the point-wise loss function is defined as: ; in, is the positioning error loss based on MAE, is the heatmap regression loss, is the L2 regularization term, Adjust parameters for your experiment.

[0010] Furthermore, in step S3, the transformation matrix T of the spatial transformer network (STN) is: ; in, 、 is the scaling factor, 、 is the translation amount, which is used to automatically adjust the position and scale of the region of interest.

[0011] Furthermore, in step S4, sub-pixel positioning includes: Centered on the pixel-level detection results, a 3×3 local heatmap region is extracted; Assuming that the local area obeys a two-dimensional Gaussian distribution, the Gaussian distribution parameters are fitted by the least squares method to solve the sub-pixel coordinates. ,in: ; is the probability value of the local heat map.

[0012] Furthermore, in step S5, post-processing optimization includes: Heatmap adaptive thresholding: Adaptive thresholding , , using the threshold to filter probability points, where are the global mean and standard deviation of the image, ; DBSCAN clustering: By setting the neighborhood radius And the minimum number of points MinPts, cluster the detection points, and retain the clusters with points greater than MinPts as valid results.

[0013] The present invention also provides a system for detecting the position of metal hangers for dry-hanging stone curtain walls based on millimeter wave images, which is used to implement the above method. The system includes: An image acquisition module, used to obtain millimeter wave images of dry-hanging stone curtain walls; The point target detection neural network module uses the U-Net architecture combined with the MobileNet module to input millimeter wave images and output key point location probability maps; Point loss function module, including positioning error calculation unit, heat map regression loss calculation unit and L2 regularization unit, is used to optimize network training; Advanced feature extraction and matching module, integrating spatial transformer network (STN) and contrast loss calculation unit to correct key point regions and enhance feature similarity; Sub-pixel positioning module, which realizes sub-pixel coordinate calculation based on Gaussian fitting method; The post-processing optimization module includes a heat map adaptive thresholding unit and a DBSCAN clustering unit to remove noise and false positives.

[0014] Furthermore, the output of the point target detection neural network module is: ; in, It is a lightweight convolutional network (MobileNet). This is the enhanced millimeter wave image.

[0015] Furthermore, the contrast loss (ContrastiveLoss) function of the contrast loss calculation unit is: ; in, is the eigenvector distance, For labels, is the marginal value, which is used to measure the similarity of feature vectors.

[0016] Furthermore, the sub-pixel positioning module fits the local heat map area through a two-dimensional Gaussian distribution model, and the Gaussian distribution expression is: ; in , The error function E of the local heat map area is obtained by solving it using the least squares method; is the amplitude, , is the center of the Gaussian distribution, i.e., the sub-pixel coordinate, 、 is the standard deviation.

[0017] Furthermore, the DBSCAN clustering unit determines the similarity between points by Euclidean distance. When the number of points in the neighborhood is ≥MinPts, they are grouped into the same cluster, isolated noise points are removed, and valid detection clusters are retained; are the two points detected, is the neighborhood radius, and MinPts is the minimum number of points required to form a dense area.

[0018] The technical solution of the present invention has the following beneficial effects: 1. Detection efficiency is significantly improved By designing a lightweight point object detection network (MetalPointNet), combined with the MobileNet module and U-Net architecture, the model complexity is significantly reduced. Experiments show that compared with the traditional YOLOv5 model: Parameters reduced by 66%: The model parameters are reduced from 7.2M in YOLOv5 to 2.4M; Inference speed increased by 3.2 times: Single-frame image processing time reduced from 25ms to 7.8ms (NVIDIA Jetson TX2 platform), meeting real-time detection needs; Reduced computing resource usage: The floating-point operations (FLOPs) are reduced to 0.8G, making it suitable for edge computing device deployment.

[0019] 2. Detection accuracy is significantly improved To address the low resolution and high noise characteristics of millimeter-wave images, the following technologies are used to improve sub-pixel positioning accuracy and robustness: Sub-pixel positioning error reduction: After using Gaussian fitting technology, the positioning error is reduced from ±1.5 pixels of the traditional method to ±0.1 pixels (error reduction of 93%); Improved false positive rate and missed detection rate: In the measured data set, the false positive rate (FPR) was reduced from 8.7% to 1.2%, and the missed detection rate (FNR) was reduced from 6.5% to 0.9%; Heatmap regression and contrastive loss optimization: The key point detection accuracy (AP@0.5) reaches 98.4%, an increase of 21% over Faster R-CNN.

[0020] 3. Enhanced algorithm robustness The spatial transformer network (STN) and post-processing optimization techniques effectively address common geometric deformations and noise interference in millimeter-wave images: Adaptive correction capability: STN can maintain key point detection accuracy (AP@0.5 ≥ 96%) under conditions of image rotation ±15° and scaling ±20%; Noise suppression effect: Based on DBSCAN clustering and adaptive thresholding, the noise false detection point removal rate is as high as 95% in low-quality images with a signal-to-noise ratio (SNR) of 10dB; Feature matching robustness: Contrastive Loss reduces the mismatch rate to below 3% in similar metal interference scenarios.

[0021] 4. Practical application value Improved engineering efficiency: In curtain wall inspection projects, the time for a single inspection is greatly reduced; Economic optimization: Supports deployment of low-computing-power devices, reducing equipment costs compared to high-computing-power platforms; Safety assurance: By locating the hidden dangers of metal hangers with high precision, we can effectively prevent curtain wall falling accidents and improve public safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the overall flow chart of the present invention; Figure 2 It is the point target detection network feature map of the present invention; Figure 3 It is a comparison diagram of the test results of the present invention and the actual object. DETAILED DESCRIPTION

[0023] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It will be understood that the specific embodiments described herein are merely intended to explain the present invention rather than to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only show portions related to the present invention rather than all of the present invention.

[0024] Combine Figure 1-Figure 3 As shown, the present invention provides a method for detecting the position of metal hangers of dry-hanging stone curtain walls based on millimeter wave images, the method comprising the following steps: S1, design a lightweight point target detection neural network MetalPointNet based on the U-Net architecture and combined with the MobileNet module. It inputs the millimeter wave image I and outputs a probability map P of key point locations, where P(i, j) represents the probability that the pixel (i, j) is a key point of a metal pendant. S2, design a point loss function, which includes positioning error loss, heatmap regression loss and L2 regularization term, and optimizes the training of the point target detection neural network through weighted summation; S3 uses the spatial transformer network (STN) and contrastive loss function for advanced feature extraction and matching, automatically corrects the position and scale of key point areas, and enhances feature vector similarity; S4, uses Gaussian fitting method to achieve sub-pixel positioning, accurately locating the highly reflective bright spots of metal pendants below the pixel level; S5, post-processing optimization of the detection results, removing noise and false positives through heat map adaptive thresholding and DBSCAN clustering algorithm, retaining valid detection points and visualizing candidate boxes.

[0025] In step S2, the total loss function of the point-wise loss function is defined as: ; in, is the positioning error loss based on MAE, is the heatmap regression loss, is the L2 regularization term, Adjust parameters for your experiment.

[0026] In step S3, the transformation matrix T of the spatial transformer network (STN) is: ; in, 、 is the scaling factor, 、 is the translation amount, which is used to automatically adjust the position and scale of the region of interest.

[0027] In step S4, sub-pixel positioning includes: Centered on the pixel-level detection results, a 3×3 local heatmap region is extracted; Assuming that the local area obeys a two-dimensional Gaussian distribution, the Gaussian distribution parameters are fitted by the least squares method to solve the sub-pixel coordinates. ,in: ; is the probability value of the local heat map.

[0028] In step S5, post-processing optimization includes: Heatmap adaptive thresholding: Adaptive thresholding , , using the threshold to filter probability points, where are the global mean and standard deviation of the image, ; DBSCAN clustering: By setting the neighborhood radius And the minimum number of points MinPts, cluster the detection points, and retain the clusters with points greater than MinPts as valid results.

[0029] The present invention also provides a millimeter wave image-based dry-hanging stone curtain wall metal hanger position detection system for implementing the above-mentioned detection method, the system comprising: An image acquisition module, used to obtain millimeter wave images of dry-hanging stone curtain walls; The point target detection neural network module uses the U-Net architecture combined with the MobileNet module to input millimeter wave images and output key point location probability maps; Point loss function module, including positioning error calculation unit, heat map regression loss calculation unit and L2 regularization unit, is used to optimize network training; Advanced feature extraction and matching module, integrating spatial transformer network (STN) and contrast loss calculation unit to correct key point regions and enhance feature similarity; Sub-pixel positioning module, which realizes sub-pixel coordinate calculation based on Gaussian fitting method; The post-processing optimization module includes a heat map adaptive thresholding unit and a DBSCAN clustering unit to remove noise and false positives.

[0030] The output of the point target detection neural network module is: ; in, It is a lightweight convolutional network (MobileNet). This is the enhanced millimeter wave image.

[0031] The contrast loss (ContrastiveLoss) function of the contrast loss calculation unit is: ; in, is the eigenvector distance, For labels, is the marginal value, which is used to measure the similarity of feature vectors.

[0032] The sub-pixel positioning module fits the local heat map area through a two-dimensional Gaussian distribution model. The Gaussian distribution expression is: ; in , The error function E of the local heat map area is obtained by solving it using the least squares method; is the amplitude, , is the center of the Gaussian distribution, i.e., the sub-pixel coordinate, 、 is the standard deviation.

[0033] The DBSCAN clustering unit uses the Euclidean distance to determine the similarity between points. When the number of points in the neighborhood is ≥MinPts, they are grouped into the same cluster, isolated noise points are removed, and valid detection clusters are retained; are the two points detected, is the neighborhood radius, and MinPts is the minimum number of points required to form a dense area.

[0034] Working principle: The present invention proposes a method for detecting the position of metal hangers on a dry-hanging stone curtain wall based on millimeter wave images, which mainly includes the following steps: 1. Design a lightweight point object detection neural network to directly locate the highly reflective bright spots of metal pendants, avoiding the redundant information of traditional bounding box methods.

[0035] 2. Design a point-wise loss function to better train the point-wise network to the best possible outcome.

[0036] 3. Advanced feature extraction and matching: Use the spatial transformer network (STN) and contrastive loss function to automatically correct the position and size of the key point area, and enhance the robustness and accuracy of the detection results by comparing the similarity of feature vectors.

[0037] 4. Accurate positioning: Sub-pixel positioning technology is used to further improve the detection accuracy of the metal pendant position.

[0038] 5. Post-processing optimization: Cluster and filter the detection results to remove noise and false positives to ensure the reliability of the detection results.

[0039] (2) Technical solution 1. Point object detection neural network: MetalPointNet is based on the U-Net architecture and combined with the MobileNet module to enhance feature extraction capabilities. Suppose the input image is , the output is the probability map of the key point location ,in Represents image is the probability of the key point. The network output is as follows ; It is a lightweight convolutional network (MobileNet).

[0040] 2. Point-wise loss function design: 1) Positioning error: Key point positioning error : MAE is used as the basic loss, the formula is as follows ; in and Represent the real and predicted key point coordinates respectively, is the sample size.

[0041] 2) Heatmap regression loss : In order to enhance the positioning accuracy of key points, a heatmap regression mechanism is introduced, and its loss is defined as: ; in, and Represent the true and predicted heat map distributions, and are the height and width of the heatmap respectively 3) Regularization term: In order to avoid overfitting, L2 regularization term is added

[0042] ; is the model weight set, is the regularization coefficient.

[0043] 4) Definition of total loss function: ; parameter Adjusted according to experiments to achieve the best performance, used to balance the heatmap regression loss and localization loss.

[0044] 3. Advanced feature extraction and matching: Use Spatial Transformer Networks (STN) to automatically adjust the position and scale of the region of interest to improve the accuracy of key point detection. Let the transformation matrix be , which is calculated as follows: ; in is the scaling factor, is the translation amount.

[0045] For feature matching, the contrast loss (ContrastiveLoss) based on the Siamese network structure is used to measure the similarity between two feature vectors: ; in is the distance between two eigenvectors, is the label (same or different), is the marginal value.

[0046] 4. Sub-pixel positioning solution The goal of sub-pixel positioning is to accurately locate the highly reflective bright spots of metal pendants at a sub-pixel level (e.g., 0.1 pixel). The present invention uses a Gaussian fitting method to achieve sub-pixel positioning.

[0047] 1) Gaussian fitting model Assume that the high-reflectivity bright spots of metal pendants follow a two-dimensional Gaussian distribution in the heat map: ; in: is the amplitude, is the center of the Gaussian distribution (i.e., sub-pixel coordinates), is the standard deviation.

[0048] 2) Derivation of sub-pixel distribution center Assume that the local heat map area is , the fitting objective is to minimize the following error function: ; Solve for Gaussian distribution parameters using the least squares method: ; The solution is: ; 3) Sub-pixel positioning step a) Extract local area: Centered on the pixel-level detection results, extract The local heat map area.

[0049] b) Fitting Gaussian distribution: Use the least squares method to fit the Gaussian distribution parameters of the local area. .

[0050] c) Calculate sub-pixel coordinates: The Gaussian center obtained by fitting As sub-pixel coordinates 5. Post-processing methods The goal of post-processing is to optimize the thermal map and coordinates output by the model, remove noise points and false detection points, and ensure the reliability of the detection results.

[0051] 1) Heatmap adaptive thresholding Threshold the heat map output by the model and retain the probability value greater than the threshold The points: ; in: is the position in the heat map The probability value of is the adaptive threshold

[0052] in, and are the global mean and standard deviation of the image, is the empirical coefficient (taken as 3 in this invention) 2) Cluster screening Use DBSCAN clustering algorithm to cluster the detected points and remove isolated noise points. With MinPts, we can identify high-density areas in the dataset and treat these areas as different clusters. Points that fail to meet the conditions for becoming core points and do not belong to any cluster are considered noise.

[0053] Define the Euclidean distance between two points ,like And the number of points in the neighborhood , they are classified into the same cluster.

[0054] in: are the two points detected. Yes The coordinates of is the neighborhood radius. MinPts is the minimum number of points required to form a dense area.

[0055] Specific steps: a) Initialize all points to be unvisited.

[0056] b) Randomly select an unvisited point to start with and find its All points in the neighborhood.

[0057] c) If the number of points in the neighborhood is less than , then mark the point as noise; otherwise, create a new cluster and add all points in the neighborhood to this cluster.

[0058] d) Repeat the above process for each point in the new cluster until no new points can be added to any cluster.

[0059] e) When all points have been visited, the algorithm ends.

[0060] After clustering, the clusters with points greater than MinPts are retained as the final metal pendant detection results.

[0061] 3) Visualization of candidate boxes To help the human eye quickly locate metal pendants, a fixed-size red candidate box is added to the center of each valid cluster to visually identify the metal pendant's location. In the final detection result image, the center of the metal pendant is surrounded by the red candidate box, while the noise points are not framed.

[0062] Beneficial effects: The millimeter-wave image-based metal hanger position detection method for dry-hanging stone curtain walls proposed in this paper effectively addresses the shortcomings of existing methods in terms of bounding box redundancy, insufficient detection accuracy, and computational overhead through targeted technical improvements. The specific technical effects are as follows: 1. Detection efficiency is significantly improved By designing a lightweight point object detection network (MetalPointNet), combined with the MobileNet module and U-Net architecture, the model complexity is significantly reduced. Experiments show that compared with the traditional YOLOv5 model: Parameters reduced by 66%: The model parameters are reduced from 7.2M in YOLOv5 to 2.4M; Inference speed increased by 3.2 times: Single-frame image processing time reduced from 25ms to 7.8ms (NVIDIA Jetson TX2 platform), meeting real-time detection needs; Reduced computing resource usage: The floating-point operations (FLOPs) are reduced to 0.8G, making it suitable for edge computing device deployment.

[0063] 2. Detection accuracy is significantly improved To address the low resolution and high noise characteristics of millimeter-wave images, the following technologies are used to improve sub-pixel positioning accuracy and robustness: Sub-pixel positioning error reduction: After using Gaussian fitting technology, the positioning error is reduced from ±1.5 pixels of the traditional method to ±0.1 pixels (error reduction of 93%); Improved false positive rate and missed detection rate: In the measured data set, the false positive rate (FPR) was reduced from 8.7% to 1.2%, and the missed detection rate (FNR) was reduced from 6.5% to 0.9%; Heatmap regression and contrastive loss optimization: The key point detection accuracy (AP@0.5) reaches 98.4%, an increase of 21% over Faster R-CNN.

[0064] 3. Enhanced algorithm robustness The spatial transformer network (STN) and post-processing optimization techniques effectively address common geometric deformations and noise interference in millimeter-wave images: Adaptive correction capability: STN can maintain key point detection accuracy (AP@0.5 ≥ 96%) under conditions of image rotation ±15° and scaling ±20%; Noise suppression effect: Based on DBSCAN clustering and adaptive thresholding, the noise false detection point removal rate is as high as 95% in low-quality images with a signal-to-noise ratio (SNR) of 10dB; Feature matching robustness: Contrastive Loss reduces the mismatch rate to below 3% in similar metal interference scenarios.

[0065] 4. Practical application value Improved engineering efficiency: In curtain wall inspection projects, the time for a single inspection is greatly reduced; Economic optimization: Supports deployment of low-computing-power devices, reducing equipment costs compared to high-computing-power platforms; Safety assurance: By locating the hidden dangers of metal hangers with high precision, we can effectively prevent curtain wall falling accidents and improve public safety.

[0066] 5. Technology comparison experiment In the curtain wall metal dataset, the comparison results with existing mainstream methods are as follows:

[0067] in conclusion: Through lightweight network design, point-wise loss function optimization, sub-pixel positioning, and post-processing techniques, this invention significantly surpasses existing methods in detection efficiency, accuracy, and robustness, providing efficient and reliable technical support for the engineering application of millimeter-wave imaging in curtain wall inspection. The technical solution of this invention has a wide range of applications and market prospects, including in construction, bridges, transportation, and other fields.

[0068] The present invention can achieve a high classification accuracy and can distinguish between a variety of different dry-hanging stone curtain wall metal pendants. At the same time, the technical solution uses millimeter wave images as input, avoiding the problem of traditional optical influence and improving the scope of application and reliability of detection. In addition, the present invention uses a deep learning algorithm for model training, which has high accuracy and robustness and can effectively improve recognition efficiency and work precision. Using the technical solution of the present invention can greatly reduce the cost of metal pendant detection and improve the economic benefits of the enterprise. After actual testing, the accuracy of the present invention in experimental scenarios can reach more than 90%, which has high practical value and application prospects.

[0069] Example 1: 1. Millimeter-wave image dataset construction (1) Data collection and annotation Acquisition equipment: Use millimeter-wave imaging radar to scan the building curtain wall and obtain images of highly reflective bright spots of metal hangers.

[0070] Labeling method: Manually label the center coordinates of the metal pendant in the image (not the bounding box) and generate a corresponding coordinate label file (such as CSV or JSON format). For example, if a metal pendant is located at pixel (120, 85) in the image, the label is `{"x":120,"y":85}`.

[0071] (2) Data enhancement and preprocessing Geometric transformation: Randomly flip the image horizontally / vertically, rotate it (±15°), and translate it (±10%) to simulate different shooting angles.

[0072] Noise injection: Gaussian noise (σ=0.01) and salt and pepper noise (density=0.05) are added to enhance the model's anti-interference ability.

[0073] Normalization: Normalize the pixel values to the range [0,1], the formula is:

[0074] in and are the mean and standard deviation of the training set.

[0075] (3) Dataset division The training set, validation set, and test set are divided into 7:2:1 ratios to ensure data distribution consistency.

[0076] 2. Point target detection model training (1) Network architecture configuration Design network based on MetalPointNet (U-Net+MobileNet lightweight module): Encoder: MobileNetV3 is used as the backbone network to extract multi-scale features.

[0077] Decoder: U-Net symmetrical structure, fusing shallow and deep features through upsampling and skip connections.

[0078] Output layer: 1-channel heat map (size consistent with input), which represents the probability distribution of key points after Sigmoid activation.

[0079] (2) Loss Function and Optimizer Total loss function: ; Optimizer: AdamW, initial learning rate , and adjust the learning rate with the cosine annealing strategy.

[0080] (3) Training process Input data: batch size (BatchSize) = 16, image size 512×512.

[0081] Training cycle: maximum 100 epochs, using the early stopping strategy. If the validation set loss does not decrease for 10 consecutive epochs, training is terminated.

[0082] Weight preservation: Save the model weights in each round, and finally select the weight with the smallest validation set loss as the best model (such as `best_model.pth`).

[0083] 3. Model Inference and Post-Processing (1) Heat map generation Input the millimeter wave image to be detected into the trained optimal weight network and output the heat map , where the highlighted area represents the probability distribution of the metal pendant location.

[0084] (2) Adaptive thresholding Calculate dynamic threshold based on global statistical characteristics of heat map (3) Clustering denoising Parameter setting: Neighborhood radius Pixels, minimum number of cluster points .

[0085] Clustering algorithm process: a) Perform density clustering on candidate points to eliminate isolated noise points (such as single points or small clusters of two points).

[0086] b) Retain the center point of the dense area after clustering as the final detection result.

[0087] (4) Visualization of candidate boxes A fixed-size red candidate box is added to the center of each valid cluster to visually identify the location of the metal pendant.

[0088] 4. Implementation sequence Data preprocessing, model training, and inference must be performed strictly in steps and cannot be interchanged; In post-processing, the order of thresholding and clustering is fixed, and low-probability points are filtered first and then noise is removed.

[0089] Through the above-mentioned implementation, the present invention realizes efficient and accurate detection of metal pendants in millimeter wave images, avoids redundant calculations of traditional bounding box methods, and has strong anti-noise capabilities.

[0090] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made by using the contents of the present description and drawings under the inventive concept of the present invention, or direct / indirect application in other related technical fields are included in the protection scope of the present invention.

Claims

1. A millimeter wave image dry hanging stone curtain wall metal hanger position detection method, characterized in that: The method comprises the following steps: S1, design a lightweight point target detection neural network based on the U-Net architecture and combined with the MobileNet module. It inputs the millimeter wave image I and outputs a probability map P of the key point location, where P(i, j) represents the probability that the pixel (i, j) is a key point of the metal pendant. S2, design a point loss function, which includes positioning error loss, heatmap regression loss and L2 regularization term, and optimizes the training of the point target detection neural network through weighted summation; S3, which uses a spatial transformer network and contrastive loss function for advanced feature extraction and matching, automatically corrects the position and scale of key point regions, and enhances feature vector similarity; S4, uses Gaussian fitting method to achieve sub-pixel positioning, accurately locating the highly reflective bright spots of metal pendants below the pixel level; S5, post-processing optimization of the detection results, removing noise and false positives through heat map adaptive thresholding and DBSCAN clustering algorithm, retaining valid detection points and visualizing candidate boxes.

2. The detection method according to claim 1, characterized in that In step S2, the total loss function of the point-wise loss function is defined as: ； in, is the positioning error loss based on MAE, is the heatmap regression loss, is the L2 regularization term, Adjust parameters for your experiment.

3. The detection method according to claim 1, wherein In step S3, the transformation matrix T of the spatial transformation network is: ； in, 、 is the scaling factor, 、 is the translation amount, which is used to automatically adjust the position and scale of the region of interest.

4. The detection method according to claim 1, wherein In step S4, sub-pixel positioning includes: Centered on the pixel-level detection results, a 3×3 local heatmap region is extracted; Assuming that the local area obeys a two-dimensional Gaussian distribution, the Gaussian distribution parameters are fitted by the least squares method to solve the sub-pixel coordinates. ,in: ； is the probability value of the local heat map.

5. The detection method according to claim 1, wherein In step S5, post-processing optimization includes: Heatmap adaptive thresholding: Adaptive thresholding , , using the threshold to filter probability points, where are the global mean and standard deviation of the image, ; DBSCAN clustering: By setting the neighborhood radius And the minimum number of points MinPts, cluster the detection points, and retain the clusters with points greater than MinPts as valid results.

6. A millimeter wave image dry hanging stone curtain wall metal hanger position detection system, used to implement the method according to any one of claims 1 to 5, characterized in that: The system includes: An image acquisition module, used to obtain millimeter wave images of dry-hanging stone curtain walls; The point target detection neural network module uses the U-Net architecture combined with the MobileNet module to input millimeter wave images and output key point location probability maps; Point loss function module, including positioning error calculation unit, heat map regression loss calculation unit and L2 regularization unit, is used to optimize network training; Advanced feature extraction and matching module, integrating spatial transformer network (STN) and contrast loss calculation unit to correct key point regions and enhance feature similarity; Sub-pixel positioning module, which realizes sub-pixel coordinate calculation based on Gaussian fitting method; The post-processing optimization module includes a heat map adaptive thresholding unit and a DBSCAN clustering unit to remove noise and false positives.

7. The detection system according to claim 6, characterized in that The output of the point target detection neural network module is: ； in, For lightweight convolutional networks, This is the enhanced millimeter wave image.

8. The detection system according to claim 6, characterized in that The contrast loss function of the contrast loss calculation unit is: ； in, is the eigenvector distance, For labels, is the marginal value, which is used to measure the similarity of feature vectors.

9. The detection system according to claim 6, characterized in that: The sub-pixel positioning module fits the local heat map area through a two-dimensional Gaussian distribution model. The Gaussian distribution expression is: ； in , The error function E of the local heat map area is obtained by solving it using the least squares method; is the amplitude, , is the center of the Gaussian distribution, i.e., the sub-pixel coordinate, 、 is the standard deviation.

10. The detection system according to claim 6, characterized in that: The DBSCAN clustering unit uses the Euclidean distance to determine the similarity between points. When the number of points in the neighborhood is ≥MinPts, they are grouped into the same cluster, isolated noise points are removed, and valid detection clusters are retained; are the two points detected, is the neighborhood radius, and MinPts is the minimum number of points required to form a dense area.

Images