Image Recognition-Based Method and System for Identifying Defects in Building Insulation Exterior Walls

Abstract

This invention belongs to the field of image recognition technology. It discloses a method and system for identifying defects in building insulation exterior walls based on image recognition. The method includes: acquiring infrared thermal images of the exterior wall and registering them with seam grid constraint information; extracting local abnormal connected components and dividing them into candidate and normal sections after seam crossing processing; binding the candidate sections to the nearest spatially located normal sections to form candidate pairs, and forming normal pairs from adjacent normal sections; acquiring infrared image sequences to construct a spatiotemporal response matrix, extracting multi-dimensional features and calculating the pairing feature difference vector; calculating statistical deviation based on the image feature difference vector of the normal pairing set, determining anomalies in the candidate sections, and outputting defect detection results. This invention introduces seam topological constraints and micro-environment dynamic thermal response feature comparison, effectively overcoming outdoor environmental baseline drift, eliminating false defect interference caused by surface stains, and improving the accuracy of defect identification.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a method and system for identifying defects in building insulation exterior walls based on image recognition. Background Technology

[0002] Building insulated exterior walls typically employ a composite structural system consisting of insulation boards and plaster layers. Hollow areas are the most common and dangerous type of defect, manifesting as localized gaps between the plaster layer and the insulation board. In severe cases, this can lead to large-area exterior wall detachment, posing significant safety hazards. Hollow defects have no visual characteristics in visible light images; current detection methods primarily rely on infrared thermal imaging technology. By acquiring infrared thermal images of the exterior wall, areas with abnormal temperatures are extracted as candidate defects for assessment.

[0003] Several problems remain in actual inspection. First, environmental factors such as airflow disturbances, residual water, and dust on the exterior wall surface can also create localized thermal anomalies in infrared thermograms. These anomalies, mixed with those caused by actual hollow defects, result in a large number of candidate areas and a high false alarm rate. Inspectors must manually verify each candidate area, which is extremely inefficient in large-area exterior wall inspection scenarios. Second, for candidate areas that pass the initial screening, existing methods rely on the temperature amplitude and morphological characteristics of a single thermogram for judgment. However, actual hollow defects and some surface interferences appear highly similar in a single image, making it difficult to reliably distinguish them based solely on static image features, leading to insufficient accuracy in the final defect identification. Summary of the Invention

[0004] To overcome the aforementioned problems in the prior art, this invention proposes a method and system for identifying defects in building insulation exterior walls based on image recognition, which is used to solve the above problems.

[0005] This invention provides the following technical solution: Image recognition-based methods for identifying defects in building insulation exterior walls include: Infrared thermal images of the exterior wall to be inspected are collected, and the layout information of the exterior wall insulation panels is registered to the coordinate system of the infrared thermal image to obtain the joint grid constraint information. Local contrast anomaly detection is performed on the infrared thermal image to obtain a set of thermal anomaly connected components; The set of thermal anomaly connected domains is structured based on the seam mesh constraint information to obtain a set of candidate plates and a set of normal plates. For each candidate segment in the candidate segment set, bind the two spatially closest normal segments from the normal segment set, and use the closer one as the comparison segment to form a candidate comparison pair with the candidate segment. Then, use the two normal segments to form a normal pair with each other to obtain a candidate comparison pair set and a normal pair set. Infrared image sequences of the candidate comparison pairing set and the normal pairing set are collected. Spatiotemporal response matrices are constructed for the blocks in each pairing. Image features are extracted and the image feature difference vector of each pairing is calculated. Based on the image feature difference vector of the normal pairing set, anomaly determination is performed on the image feature difference vector of each pair in the candidate comparison pairing set, and the candidate blocks with the determination result of being abnormal are output as defect detection results.

[0006] Preferably, the step of registering the layout information of the external wall insulation panels to the infrared thermal image coordinate system includes: A visible light camera was used to capture images of the same area of ​​the exterior wall under the same conditions as an infrared camera, resulting in a visible light image corresponding to the infrared thermal image. The visible light image is registered with the layout information of the external wall insulation panel to establish the transformation relationship between the visible light image coordinate system and the external wall insulation panel layout information coordinate system; By utilizing the same-viewpoint correspondence between the infrared thermal image and the visible light image, the transformation relationship is transferred to the coordinate system of the infrared thermal image, so that the layout information of the external wall insulation panel is aligned with the infrared thermal image, and the splicing mesh constraint information is obtained.

[0007] Preferably, the step of detecting local contrast anomalies in the infrared thermal image includes: The infrared thermal image is subjected to a large-scale low-pass filter, with the filter kernel size being larger than the corresponding size of a single insulation board in the image, to obtain a low-frequency background image; The local deviation map is obtained by taking the absolute value of the difference between the infrared thermal image and the low-frequency background image. The pixel values ​​of the entire image of the local deviation map are statistically analyzed, and pixels with local deviation values ​​higher than the segmentation threshold are extracted as thermal anomaly pixels using a preset multiple of the standard deviation of the main peak as the segmentation threshold. Connectivity analysis is performed on thermally anomalous pixels, and connected components with areas lower than a preset minimum area threshold are filtered out to obtain a set of thermally anomalous connected components.

[0008] Preferably, the step of structuring the set of thermal anomaly connected components based on the seam mesh constraint information includes: Based on the joint lines in the joint grid constraint information, the exterior wall is divided into several insulation panels, each of which is enclosed by adjacent joint lines. For each connected domain in the set of thermal anomaly connected domains, determine whether there is a seam line within the area enclosed by the connected domain. If there is a seam line, exclude the connected domain and include the insulation plate where the connected domain does not have a seam line in the candidate plate set. Insulation plates that do not belong to the candidate plate set are included in the normal plate set.

[0009] Preferably, the method for constructing the spatiotemporal response matrix includes: For each pixel within the region, extract its grayscale temporal vector across all frames of the infrared image sequence. Arrange the grayscale temporal vectors of all pixels within the region into rows by pixels and columns by frames to obtain the spatiotemporal response matrix of the region.

[0010] Preferably, the image features include spatiotemporal structure concentration features, temporal autocorrelation decay features, and intra-frame grayscale space complexity features.

[0011] Preferably, the method for extracting the spatiotemporal structure concentration feature includes: performing singular value decomposition on the spatiotemporal response matrix, extracting each singular value, calculating the ratio of the square of the largest singular value to the sum of the squares of all singular values, and obtaining the spatiotemporal structure concentration of the spatiotemporal response matrix; The method for extracting the temporal autocorrelation decay feature includes: taking the average value of the spatiotemporal response matrix by column to obtain the gray-scale average time series; calculating the autocorrelation coefficient of the gray-scale average time series under different lag steps, determining the number of lag frames corresponding to the first drop of the autocorrelation coefficient to a preset decay threshold, and obtaining the autocorrelation half-decay step number of the spatiotemporal response matrix; The method for extracting the intra-frame grayscale space complexity features includes: for each frame in the infrared image sequence, traversing all window positions in the region with a sliding window of a preset side length, calculating the local standard deviation of the pixel grayscale values ​​in each window, and taking the mean for all frames and all window positions to obtain the intra-frame grayscale space complexity of the region.

[0012] Preferably, calculating the image feature difference vector for each pair includes: The absolute value of the difference in spatiotemporal structural concentration between the two plates in the pair is calculated to obtain the spatiotemporal structural concentration difference. The autocorrelation decay difference is obtained by calculating the absolute value of the difference in the autocorrelation half-decay steps between the two plates in the pair. Calculate the absolute value of the difference in intra-frame grayscale spatial complexity between the two plates in the pairing to obtain the intra-frame spatial complexity difference. The difference in spatiotemporal structure concentration, the difference in autocorrelation attenuation, and the difference in intra-frame spatial complexity are used to form an image feature difference vector.

[0013] Preferably, the step of determining anomalies in the image feature difference vectors of each pair in the candidate comparison pair set, based on the image feature difference vectors of the normal pair set, includes: Based on the image feature difference vector of each normal pair in the normal pair set, the statistical parameters of the normal difference distribution are calculated; For each pair of image feature difference vectors in the candidate comparison pairing set, the statistical deviation of the difference vector relative to the normal difference distribution is calculated based on the statistical parameters. When the statistical deviation is not lower than the preset high deviation threshold, the candidate segment is judged as abnormal; when the statistical deviation is not lower than the preset low deviation threshold and is lower than the preset high deviation threshold, the candidate segment is marked as suspected abnormal; when the statistical deviation is lower than the preset low deviation threshold, the candidate segment is excluded.

[0014] This invention also provides a building insulation exterior wall defect identification system based on image recognition, used to implement a building insulation exterior wall defect identification method based on image recognition, including: The image acquisition and registration module is used to acquire infrared thermal images of the exterior wall to be inspected and register the layout information of the exterior wall insulation panels to the infrared thermal image coordinate system to obtain the splice grid constraint information. The local anomaly detection module is used to perform local contrast anomaly detection on the infrared thermal image to obtain a set of thermal anomaly connected components. The structured processing module is used to perform structured processing on the set of thermal anomaly connected domains according to the splice mesh constraint information to obtain a set of candidate plates and a set of normal plates. The pairing construction module is used to bind the two spatially closest normal plates from the normal plate set to each candidate plate in the candidate plate set, and use the closer one as the comparison plate to form a candidate comparison pair with the candidate plate. The two normal plates form a normal pair with each other, thus obtaining a candidate comparison pair set and a normal pair set. The feature extraction module is used to acquire infrared image sequences of the candidate comparison pairing set and the normal pairing set, construct spatiotemporal response matrices for each pairing, extract image features, and calculate the image feature difference vector for each pairing. The anomaly detection module is used to determine the anomalies of the image feature difference vectors of each pair in the candidate comparison pair set based on the image feature difference vectors of the normal pair set, and output the candidate blocks with the anomaly determination result as the defect detection result.

[0015] This invention provides a method and system for identifying defects in building insulation exterior walls based on image recognition, which has the following beneficial effects: This invention registers the layout information of external wall insulation panels to the coordinate system of an infrared thermographic image, establishes joint mesh constraint information on the infrared thermographic image, and uses whether the connected domains cross the joint lines as a structured screening criterion. False anomalies that spread across joints within the thermal anomaly connected domains are eliminated, and the insulation panels containing the connected domains that do not cross joints are retained as candidate panels. Since the thermal anomalies of genuine hollow defects are constrained by the physical boundaries of the insulation board and do not cross joints, while thermal anomalies caused by surface interference such as airflow disturbances, water stains, and dust blockage are not constrained by joints and often cross multiple panels, the above screening mechanism can effectively eliminate a large number of false anomalies originating from surface interference before entering detailed analysis. This solves the problems of high false alarm rates and large workload of manual verification in existing methods for thermal anomaly candidate regions.

[0016] Based on this, the present invention binds the nearest normal plate in space to each candidate plate as a comparison plate. By acquiring infrared image sequences and constructing a spatiotemporal response matrix, three image features are jointly extracted from the time and spatial dimensions: spatiotemporal structure concentration, temporal autocorrelation decay, and intra-frame grayscale spatial complexity. The image feature difference vector between the candidate comparison pair and the normal pair is calculated, and the statistical deviation is calculated based on the distribution of the difference vector of the normal pair for anomaly determination. Since the difference in thermal response behavior between the hollow plate and the normal plate due to the difference in heat capacity has a stable structural manifestation in the temporal image features, and the feature difference vector between the normal plates is concentrated near zero, the above-mentioned determination method based on statistical deviation can reliably distinguish between real hollow defects and residual surface interference under natural thermal excitation conditions without the need for special thermal excitation equipment. This solves the problem that existing methods rely on the static features of single-frame thermal images and are difficult to accurately identify defects, thus improving the accuracy of hollow defect identification in thermal insulation exterior walls. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the image recognition-based method for identifying defects in building insulation exterior walls according to the present invention. Figure 2 This is a schematic diagram of the module of the image recognition-based building insulation exterior wall defect identification system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0019] Please see Figure 1In this embodiment, the image recognition-based method for identifying defects in building insulation exterior walls includes: S1. Collect infrared thermal images of the exterior wall to be inspected, and register the layout information of the exterior wall insulation panels to the infrared thermal image coordinate system to obtain the splice grid constraint information; In this embodiment, an infrared thermal imager is used to photograph the exterior wall to be inspected, resulting in an infrared thermal image. The grayscale value of each pixel in the infrared thermal image corresponds to the infrared radiation intensity at that location, reflecting the surface thermal state at that location.

[0020] The process of registering the layout information of the external wall insulation panels to the infrared thermal image coordinate system includes: A visible light camera was used to capture images of the same area of ​​the exterior wall under the same conditions as an infrared camera, resulting in a visible light image corresponding to the infrared thermal image. The visible light image is registered with the layout information of the external wall insulation panel to establish the transformation relationship between the visible light image coordinate system and the external wall insulation panel layout information coordinate system; By utilizing the same-viewpoint correspondence between the infrared thermal image and the visible light image, the transformation relationship is transferred to the coordinate system of the infrared thermal image, so that the layout information of the external wall insulation panel is aligned with the infrared thermal image, and the splicing mesh constraint information is obtained.

[0021] In this embodiment, the layout information of the external wall insulation panels can be derived from digital versions of architectural construction drawings, elevation layout diagrams exported from BIM models, or manually drawn schematic diagrams of insulation panel distribution after on-site surveys. The core information includes the arrangement, dimensions, and joint orientation of the insulation panels on the external wall facade. Taking common EPS insulation boards as an example, a single insulation board typically measures 600mm × 1200mm and is laid with horizontal staggered joints. This determines the horizontal joint spacing to be 600mm, the vertical joint spacing to be 1200mm, and the orientation angles to be both horizontal and vertical.

[0022] When registering visible light images with the layout information of exterior wall insulation panels, geometric feature points such as window frame corners and the endpoints of exterior wall decorative lines in the visible light image can be extracted and correlated with the corresponding structural feature points in the layout information. A random sampling consistency algorithm is used to eliminate mismatched point pairs, the affine transformation matrix is ​​estimated, and a transformation relationship from the visible light image coordinate system to the layout information coordinate system is established. Since the infrared thermal image and the visible light image are acquired from the same viewing angle, the same physical point in both images corresponds to the same pixel position. Therefore, the above transformation relationship can be directly transferred to the infrared thermal image coordinate system, so that the seam lines in the layout information are finally aligned with the infrared thermal image, obtaining the seam mesh constraint information. The seam mesh constraint information is stored in the form of a grid overlaid on the infrared thermal image, recording the position of each seam line in the pixel coordinate system of the infrared thermal image.

[0023] S2. Perform local contrast anomaly detection on the infrared thermal image to obtain a set of thermal anomaly connected components; The local contrast anomaly detection of the infrared thermal image includes: The infrared thermal image is subjected to a large-scale low-pass filter, with the filter kernel size being larger than the corresponding size of a single insulation board in the image, to obtain a low-frequency background image; The local deviation map is obtained by taking the absolute value of the difference between the infrared thermal image and the low-frequency background image. The pixel values ​​of the entire image of the local deviation map are statistically analyzed, and pixels with local deviation values ​​higher than the segmentation threshold are extracted as thermal anomaly pixels using a preset multiple of the standard deviation of the main peak as the segmentation threshold. Connectivity analysis is performed on thermally anomalous pixels, and connected components with areas lower than a preset minimum area threshold are filtered out to obtain a set of thermally anomalous connected components.

[0024] In this embodiment, since a normal wall surface typically presents as a low-frequency grayscale distribution that changes slowly in space in an infrared thermal image, while thermal anomalies caused by defects such as hollowness appear as local abrupt changes superimposed on this low-frequency background, the local abrupt changes can be separated from the overall temperature trend by subtracting the original image from its low-frequency background.

[0025] The filter kernel size for large-scale low-pass filtering should be larger than the pixel size of a single insulation board in the image. Taking a shooting distance of 5 meters, an infrared thermal imager resolution of 640×480, and a field of view of 25° as an example, a single 600mm×1200mm insulation board occupies approximately 80×160 pixels in the image. Therefore, the filter kernel size can be set to 1.5 to 2 times the pixel size of the insulation board, that is, approximately 120×240 pixels, to ensure that the low-frequency background image does not contain local changes in the scale of a single board.

[0026] The absolute value of the difference between the infrared thermal image and the low-frequency background image is used to obtain a local deviation map. The purpose of taking the absolute value is to simultaneously capture both high and low temperature anomalies, because at different stages of thermal excitation, the hollow area may exhibit either a faster temperature rise (higher temperature) or a faster temperature drop (lower temperature) compared to the normal area, both of which are valid abnormal signals.

[0027] When performing histogram analysis on the pixel values ​​of the entire image in the local deviation map, the normal wall area constitutes the vast majority of the image, and its local deviation values ​​are concentrated near zero, forming the main peak of the histogram. Using 2 to 3 times the standard deviation of the main peak as a segmentation threshold, pixels that significantly deviate from the normal background can be extracted as thermal anomaly pixels. The threshold factor can be adjusted according to the noise level of the site; in noisy environments, the factor can be appropriately increased to reduce false detections.

[0028] After performing connected component analysis on the thermal anomaly pixels, connected components with excessively small areas are filtered out. The minimum area threshold can be set according to the detection requirements, with an actual area of ​​0.01 square meters as the lower limit, corresponding to approximately 130 pixels for the above shooting parameters. Connected components smaller than this area are considered noise and filtered out, resulting in a set of thermal anomaly connected components.

[0029] S3. The set of thermal anomaly connected domains is structured according to the splice mesh constraint information to obtain the candidate plate set and the normal plate set; The step of structuring the set of thermal anomaly connected components based on the seam mesh constraint information includes: Based on the joint lines in the joint grid constraint information, the exterior wall is divided into several insulation panels, each of which is enclosed by adjacent joint lines. For each connected domain in the set of thermal anomaly connected domains, determine whether there is a seam line within the area enclosed by the connected domain. If there is a seam line, exclude the connected domain and include the insulation plate where the connected domain does not have a seam line in the candidate plate set. Insulation plates that do not belong to the candidate plate set are included in the normal plate set.

[0030] In this embodiment, the specific implementation of the seam crossing judgment is as follows: For each connected component in the thermal anomaly connected component set, extract all pixel coordinates within the area enclosed by its contour pixel list, query the seam mesh constraint information one by one, and determine whether a seam line passes through the area. If the pixel coordinates of any seam line fall within the area enclosed by the connected component, the connected component is considered to cross the seam line and is excluded. This is because the seams of the insulated exterior wall are the physical boundaries between adjacent insulation boards. Thermal anomalies caused by defects are usually confined to the range of a single insulation board and will not spread across the seam; while thermal anomalies caused by surface interference such as airflow disturbances and water stains are not constrained by the seams and often cross multiple boards, which can be excluded by the seam crossing judgment.

[0031] Connected components retained based on seam crossing criteria are included in the candidate component set within the insulation plate. If multiple non-seam-crossing thermal anomaly connected components exist within an insulation plate, that plate is counted only once. Insulation plates containing no thermal anomaly connected components are included in the normal component set as a reference for subsequent feature comparisons.

[0032] S4. For each candidate segment in the candidate segment set, bind the two spatially closest normal segments from the normal segment set, use the closer one as the comparison segment to form a candidate comparison pair with the candidate segment, and use the two normal segments to form a normal pair with each other to obtain a candidate comparison pair set and a normal pair set. In this embodiment, spatial distance is measured by the Euclidean distance between the centroids of the candidate plates and the centroids of the normal plates in the pixel coordinate system of the infrared thermal image. For each candidate plate in the candidate plate set, the normal plates are sorted from closest to furthest in the normal plate set. The closest normal plate is taken as the comparison plate, and the second closest normal plate is taken as another normal plate. Both participate in subsequent processing.

[0033] The selection of the nearest normal plate in space as the comparison plate is to ensure that the candidate plate and the comparison plate are under similar external environmental conditions, including similar solar incidence angles, similar shading conditions, and similar external wall orientations, so that they receive the most similar external thermal excitation. This eliminates the influence of environmental factors in the subsequent characteristic difference calculation, so that the difference only reflects the difference in their internal structures.

[0034] Candidate plates are paired with nearby normal plates to form candidate contrast pairs, and two normal plates form normal pairs. The purpose of normal pairs is to provide the distribution of natural feature differences between two normal plates under the same environmental conditions, serving as a benchmark for judging whether candidate contrast pairs are abnormal.

[0035] S5. Collect infrared image sequences of the candidate comparison pairing set and the normal pairing set, construct spatiotemporal response matrices for each pairing, extract image features, and calculate the image feature difference vector for each pairing; In this embodiment, multiple frames of infrared images are continuously acquired at a preset sampling interval for the plate areas involved in the candidate comparison pairing set and the normal pairing set. The sampling interval can be set to 30 seconds to 2 minutes, and the acquisition time is not less than 20 minutes, resulting in an infrared image sequence of not less than 10 frames. After acquisition, inter-frame registration is performed on each frame. Taking the first frame as the reference frame, the transformation relationship from each frame to the reference frame is estimated using fixed thermal feature points on the wall, ensuring that the same pixel position in each frame corresponds to the same physical point, thus obtaining the registered infrared image sequence.

[0036] The method for constructing the spatiotemporal response matrix includes: For each pixel within the region, extract its grayscale temporal vector across all frames of the infrared image sequence. Arrange the grayscale temporal vectors of all pixels within the region into rows by pixels and columns by frames to obtain the spatiotemporal response matrix of the region.

[0037] In this embodiment, taking a candidate plate as an example, assuming that the plate corresponds to an 80×120 pixel area in the infrared thermal image, and a total of 15 frames of images were acquired, then the spatiotemporal response matrix of the plate is a matrix of 9600 rows × 15 columns. Each row corresponds to the gray-level temporal vector of a pixel within the plate in the 15 frames, and each column corresponds to the gray-level value of all pixels within the plate in a certain frame. The spatiotemporal response matrices of the comparison plate and the normal plate are constructed in the same way.

[0038] The image features include spatiotemporal structure concentration features, temporal autocorrelation decay features, and intra-frame grayscale space complexity features.

[0039] The method for extracting the spatiotemporal structure concentration feature includes: performing singular value decomposition on the spatiotemporal response matrix, extracting each singular value, calculating the ratio of the square of the largest singular value to the sum of the squares of all singular values, and obtaining the spatiotemporal structure concentration of the spatiotemporal response matrix. The method for extracting the temporal autocorrelation decay feature includes: taking the average value of the spatiotemporal response matrix by column to obtain the gray-scale average time series; calculating the autocorrelation coefficient of the gray-scale average time series under different lag steps, determining the number of lag frames corresponding to the first drop of the autocorrelation coefficient to a preset decay threshold, and obtaining the autocorrelation half-decay step number of the spatiotemporal response matrix; The method for extracting the intra-frame grayscale space complexity features includes: for each frame in the infrared image sequence, traversing all window positions in the region with a sliding window of a preset side length, calculating the local standard deviation of the pixel grayscale values ​​in each window, and taking the mean for all frames and all window positions to obtain the intra-frame grayscale space complexity of the region.

[0040] In this embodiment, regarding the spatiotemporal structure concentration characteristics, after performing singular value decomposition on the spatiotemporal response matrix, if the square of the largest singular value accounts for a high proportion of the sum of the squares of all singular values, it indicates that the matrix can be approximately represented by a rank-1 matrix. This means that the temporal change patterns of all pixels within the region are highly consistent, exhibiting a unified response morphology spatially. The hollow region has a small heat capacity, and under external thermal excitation, the center heats up the fastest, followed by the edges, forming a spatially regular heat mound morphology. Temporally, the change direction of each pixel is consistent, the spatiotemporal response matrix is ​​approximately low-rank, and the spatiotemporal structure concentration is high. In the normal plate, the heat capacity of each pixel is similar, the temporal change is uniform and slow, the singular values ​​of the spatiotemporal response matrix are relatively uniform, and the spatiotemporal structure concentration is low.

[0041] Regarding the temporal autocorrelation decay characteristics, the mean grayscale time series of the segment is obtained by taking the average of each column of the spatiotemporal response matrix, reflecting the overall grayscale change trend of the segment over time. The autocorrelation coefficient measures the similarity between the time series and its own delayed version; the faster the autocorrelation coefficient decays, the more rapid the time series changes and the more agile the thermal response. The preset decay threshold can be set to 0.5, that is, the number of lag frames corresponding to the first time the autocorrelation coefficient drops to 0.5 is calculated as the autocorrelation half-decay step. Hollow segments have small thermal capacity, and the mean grayscale time series fluctuates rapidly with external stimuli, resulting in a small autocorrelation half-decay step; normal segments have large thermal capacity, and the time series changes slowly, resulting in a large autocorrelation half-decay step.

[0042] Regarding the intra-frame grayscale space complexity feature, the preset side length can be approximately one-tenth of the pixel size of the insulation board. Taking the above parameters as an example, an 8×8 pixel sliding window can be used. For each frame of the image, the sliding window is used to traverse all positions within the plate area, and the local standard deviation of the pixel grayscale value within each window is calculated. The average value is then taken for all frames and all window positions. This feature reflects the degree of non-uniformity of grayscale distribution within a single frame of the plate. Hollow plates have uneven grayscale distribution due to the formation of heat mounds, resulting in high intra-frame grayscale space complexity; normal plates have a flat grayscale distribution, resulting in low intra-frame grayscale space complexity.

[0043] The calculation of the image feature difference vector for each pair includes: The absolute value of the difference in spatiotemporal structural concentration between the two plates in the pair is calculated to obtain the spatiotemporal structural concentration difference. The autocorrelation decay difference is obtained by calculating the absolute value of the difference in the autocorrelation half-decay steps between the two plates in the pair. Calculate the absolute value of the difference in intra-frame grayscale spatial complexity between the two plates in the pairing to obtain the intra-frame spatial complexity difference. The difference in spatiotemporal structure concentration, the difference in autocorrelation attenuation, and the difference in intra-frame spatial complexity are used to form an image feature difference vector.

[0044] In this embodiment, taking candidate comparison pairing as an example, the absolute differences between the candidate and comparison plates in three features are calculated to form a three-dimensional image feature difference vector. If the candidate plate has hollow areas, the difference in spatiotemporal structure concentration is significantly greater than zero (the spatiotemporal structure of the candidate plate is highly concentrated while that of the comparison plate is relatively dispersed), the difference in autocorrelation decay is significantly greater than zero (the autocorrelation half-decay step of the candidate plate is small while that of the comparison plate is large), and the difference in intra-frame spatial complexity is significantly greater than zero (the gray level inside the candidate plate is uneven while that of the comparison plate is flat). All three components are relatively large, and the difference vector deviates significantly from the zero vector. For normal pairing, the two normal plates have similar thermal properties, and the differences in the three features are all close to zero, with the difference vector concentrated near the zero vector.

[0045] S6. Using the image feature difference vector of the normal pairing set as a benchmark, perform anomaly determination on the image feature difference vector of each pair in the candidate comparison pairing set, and output the candidate blocks with the anomaly determination result as the defect detection result.

[0046] The step of using the image feature difference vector of the normal pairing set as a benchmark to determine anomalies in the image feature difference vector of each pair in the candidate comparison pairing set includes: Based on the image feature difference vector of each normal pair in the normal pair set, the statistical parameters of the normal difference distribution are calculated; For each pair of image feature difference vectors in the candidate comparison pairing set, the statistical deviation of the difference vector relative to the normal difference distribution is calculated based on the statistical parameters. When the statistical deviation is not lower than the preset high deviation threshold, the candidate segment is judged as abnormal; when the statistical deviation is not lower than the preset low deviation threshold and is lower than the preset high deviation threshold, the candidate segment is marked as suspected abnormal; when the statistical deviation is lower than the preset low deviation threshold, the candidate segment is excluded.

[0047] In this embodiment, there are multiple ways to calculate statistical parameters and statistical deviations, such as Mahalanobis distance based on mean vectors and covariance matrices, negative log-likelihood based on kernel density estimation, and discriminative distance based on single-class support vector machines. Taking mean vectors and covariance matrices as an example, based on the image feature difference vectors of all normal pairs in the normal pairing set, a three-dimensional mean vector and a 3×3 covariance matrix are calculated, reflecting the central location of the normal pairing difference vectors and the dispersion of each dimension and the correlation between dimensions, respectively. Since the features of normal plates are similar, the components of the mean vector are close to zero, and the covariance matrix reflects the scale of the natural differences between normal plates. The corresponding statistical deviation is calculated using Mahalanobis distance. Mahalanobis distance normalizes each dimension using the covariance matrix, eliminating the influence of different feature dimensions and considering the correlation between dimensions, making it an effective means of measuring the degree of deviation of multidimensional vectors from the reference distribution. The larger the Mahalanobis distance between the image feature difference vector of the candidate contrast pair and the normal difference distribution, the more significant the difference in thermal response between the candidate plate and the contrast plate, the more it deviates from the natural difference level between normal plates, and the more likely there is a hollow defect.

[0048] The preset high deviation threshold and preset low deviation threshold can be determined based on historical detection data statistics, or they can be set using the significance level of the chi-square distribution. Taking Mahalanobis distance as an example, the square of Mahalanobis distance in the three-dimensional feature space approximately follows a three-degree-of-freedom chi-square distribution. Taking the chi-square critical value corresponding to a significance level of 0.01 as the high deviation threshold and the chi-square critical value corresponding to a significance level of 0.05 as the low deviation threshold, statistical anomaly detection can be achieved while controlling the false alarm rate.

[0049] When the statistical deviation is not lower than the preset high deviation threshold, the candidate plate is determined to be abnormal, that is, it has a hollow defect, and its location, area and statistical deviation are output as the defect detection result; when the statistical deviation is not lower than the preset low deviation threshold but lower than the preset high deviation threshold, it is marked as a suspected abnormality and manual review is recommended; when the statistical deviation is lower than the preset low deviation threshold, the candidate plate is excluded and identified as a non-hollow defect.

[0050] As shown above, this embodiment registers the layout information of the external wall insulation panels to the infrared thermal image, introduces the physical topological constraints of the insulation panel seam grid, and uses cross-seam judgment to filter out false anomalies that are affected by surface interference and spread across the seams from the thermal anomaly connectivity domain, effectively reducing the false alarm base in the candidate area. Based on this, by binding candidate panels to adjacent normal panels to construct pairings, collecting infrared image sequences to construct a spatiotemporal response matrix, and extracting image features reflecting the thermal response characteristics of the panels from both temporal and spatial dimensions, statistical judgment is made using the feature difference distribution differences between candidate and normal pairings. This achieves the identification of hollow defects under natural thermal excitation conditions without the need for dedicated thermal excitation equipment, solving the problems of high false alarm rate and insufficient accuracy of single-frame analysis in existing methods. Example 2

[0051] Please see Figure 2 This invention provides a building insulation exterior wall defect identification system based on image recognition, used to implement a building insulation exterior wall defect identification method based on image recognition, including: The image acquisition and registration module is used to acquire infrared thermal images of the exterior wall to be inspected and register the layout information of the exterior wall insulation panels to the infrared thermal image coordinate system to obtain the splice grid constraint information. The local anomaly detection module is used to perform local contrast anomaly detection on the infrared thermal image to obtain a set of thermal anomaly connected components. The structured processing module is used to perform structured processing on the set of thermal anomaly connected domains according to the splice mesh constraint information to obtain a set of candidate plates and a set of normal plates. The pairing construction module is used to bind the two spatially closest normal plates from the normal plate set to each candidate plate in the candidate plate set, and use the closer one as the comparison plate to form a candidate comparison pair with the candidate plate. The two normal plates form a normal pair with each other, thus obtaining a candidate comparison pair set and a normal pair set. The feature extraction module is used to acquire infrared image sequences of the candidate comparison pairing set and the normal pairing set, construct spatiotemporal response matrices for each pairing, extract image features, and calculate the image feature difference vector for each pairing. The anomaly detection module is used to determine the anomalies of the image feature difference vectors of each pair in the candidate comparison pair set based on the image feature difference vectors of the normal pair set, and output the candidate blocks with the anomaly determination result as the defect detection result.

[0052] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0053] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0054] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying defects in building insulation exterior walls based on image recognition, characterized in that, include: Infrared thermal images of the exterior wall to be inspected are collected, and the layout information of the exterior wall insulation panels is registered to the coordinate system of the infrared thermal image to obtain the joint grid constraint information. Local contrast anomaly detection is performed on the infrared thermal image to obtain a set of thermal anomaly connected components; The set of thermal anomaly connected domains is structured based on the seam mesh constraint information to obtain a set of candidate plates and a set of normal plates. For each candidate segment in the candidate segment set, bind the two spatially closest normal segments from the normal segment set, and use the closer one as the comparison segment to form a candidate comparison pair with the candidate segment. Then, use the two normal segments to form a normal pair with each other to obtain a candidate comparison pair set and a normal pair set. Infrared image sequences of the candidate comparison pairing set and the normal pairing set are collected. Spatiotemporal response matrices are constructed for the blocks in each pairing. Image features are extracted and the image feature difference vector of each pairing is calculated. Based on the image feature difference vector of the normal pairing set, anomaly determination is performed on the image feature difference vector of each pair in the candidate comparison pairing set, and the candidate blocks with the determination result of being abnormal are output as defect detection results.

2. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 1, characterized in that, The process of registering the layout information of the external wall insulation panels to the infrared thermal image coordinate system includes: A visible light camera was used to capture images of the same area of ​​the exterior wall under the same conditions as an infrared camera, resulting in a visible light image corresponding to the infrared thermal image. The visible light image is registered with the layout information of the external wall insulation panel to establish the transformation relationship between the visible light image coordinate system and the external wall insulation panel layout information coordinate system; By utilizing the same-viewpoint correspondence between the infrared thermal image and the visible light image, the transformation relationship is transferred to the coordinate system of the infrared thermal image, so that the layout information of the external wall insulation panel is aligned with the infrared thermal image, and the splicing mesh constraint information is obtained.

3. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 2, characterized in that, The local contrast anomaly detection of the infrared thermal image includes: The infrared thermal image is subjected to a large-scale low-pass filter, with the filter kernel size being larger than the corresponding size of a single insulation board in the image, to obtain a low-frequency background image; The local deviation map is obtained by taking the absolute value of the difference between the infrared thermal image and the low-frequency background image. The pixel values ​​of the entire image of the local deviation map are statistically analyzed, and pixels with local deviation values ​​higher than the segmentation threshold are extracted as thermal anomaly pixels using a preset multiple of the standard deviation of the main peak as the segmentation threshold. Connectivity analysis is performed on thermally anomalous pixels, and connected components with areas lower than a preset minimum area threshold are filtered out to obtain a set of thermally anomalous connected components.

4. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 3, characterized in that, The step of structuring the set of thermal anomaly connected components based on the seam mesh constraint information includes: Based on the joint lines in the joint grid constraint information, the exterior wall is divided into several insulation panels, each of which is enclosed by adjacent joint lines. For each connected domain in the set of thermal anomaly connected domains, determine whether there is a seam line within the area enclosed by the connected domain. If there is a seam line, exclude the connected domain and include the insulation plate where the connected domain does not have a seam line in the candidate plate set. Insulation plates that do not belong to the candidate plate set are included in the normal plate set.

5. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 4, characterized in that, The method for constructing the spatiotemporal response matrix includes: For each pixel within the region, extract its grayscale temporal vector across all frames of the infrared image sequence. Arrange the grayscale temporal vectors of all pixels within the region into rows by pixels and columns by frames to obtain the spatiotemporal response matrix of the region.

6. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 5, characterized in that, The image features include spatiotemporal structure concentration features, temporal autocorrelation decay features, and intra-frame grayscale space complexity features.

7. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 6, characterized in that, The method for extracting the spatiotemporal structure concentration feature includes: performing singular value decomposition on the spatiotemporal response matrix, extracting each singular value, calculating the ratio of the square of the largest singular value to the sum of the squares of all singular values, and obtaining the spatiotemporal structure concentration of the spatiotemporal response matrix. The method for extracting the temporal autocorrelation decay feature includes: taking the average value of the spatiotemporal response matrix by column to obtain the gray-scale average time series; calculating the autocorrelation coefficient of the gray-scale average time series under different lag steps, determining the number of lag frames corresponding to the first drop of the autocorrelation coefficient to a preset decay threshold, and obtaining the autocorrelation half-decay step number of the spatiotemporal response matrix; The method for extracting the intra-frame grayscale space complexity features includes: for each frame in the infrared image sequence, traversing all window positions in the region with a sliding window of a preset side length, calculating the local standard deviation of the pixel grayscale values ​​in each window, and taking the mean for all frames and all window positions to obtain the intra-frame grayscale space complexity of the region.

8. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 7, characterized in that, The calculation of the image feature difference vector for each pair includes: The absolute value of the difference in spatiotemporal structural concentration between the two plates in the pair is calculated to obtain the spatiotemporal structural concentration difference. The autocorrelation decay difference is obtained by calculating the absolute value of the difference in the autocorrelation half-decay steps between the two plates in the pair. Calculate the absolute value of the difference in intra-frame grayscale spatial complexity between the two plates in the pairing to obtain the intra-frame spatial complexity difference. The difference in spatiotemporal structure concentration, the difference in autocorrelation attenuation, and the difference in intra-frame spatial complexity are used to form an image feature difference vector.

9. The method for identifying defects in building insulation exterior walls based on image recognition according to claim 8, characterized in that, The step of using the image feature difference vector of the normal pairing set as a benchmark to determine anomalies in the image feature difference vector of each pair in the candidate comparison pairing set includes: Based on the image feature difference vector of each normal pair in the normal pair set, the statistical parameters of the normal difference distribution are calculated; For each pair of image feature difference vectors in the candidate comparison pairing set, the statistical deviation of the difference vector relative to the normal difference distribution is calculated based on the statistical parameters. When the statistical deviation is not lower than the preset high deviation threshold, the candidate segment is judged as abnormal; when the statistical deviation is not lower than the preset low deviation threshold and is lower than the preset high deviation threshold, the candidate segment is marked as suspected abnormal; when the statistical deviation is lower than the preset low deviation threshold, the candidate segment is excluded.

10. A building insulation exterior wall defect identification system based on image recognition, used to implement the building insulation exterior wall defect identification method based on image recognition as described in any one of claims 1-9, characterized in that, include: The image acquisition and registration module is used to acquire infrared thermal images of the exterior wall to be inspected and register the layout information of the exterior wall insulation panels to the infrared thermal image coordinate system to obtain the splice grid constraint information. The local anomaly detection module is used to perform local contrast anomaly detection on the infrared thermal image to obtain a set of thermal anomaly connected components. The structured processing module is used to perform structured processing on the set of thermal anomaly connected domains according to the splice mesh constraint information to obtain a set of candidate plates and a set of normal plates. The pairing construction module is used to bind the two spatially closest normal plates from the normal plate set to each candidate plate in the candidate plate set, and use the closer one as the comparison plate to form a candidate comparison pair with the candidate plate. The two normal plates form a normal pair with each other, thus obtaining a candidate comparison pair set and a normal pair set. The feature extraction module is used to acquire infrared image sequences of the candidate comparison pairing set and the normal pairing set, construct spatiotemporal response matrices for each pairing, extract image features, and calculate the image feature difference vector for each pairing. The anomaly detection module is used to determine the anomalies of the image feature difference vectors of each pair in the candidate comparison pair set based on the image feature difference vectors of the normal pair set, and output the candidate blocks with the anomaly determination result as the defect detection result.

Images