A method and device for detecting quality problems of fine decoration air cracks and penetration based on multi-source image data fusion

Abstract

The application discloses a kind of based on multi-source image data fusion's fine decoration empty crack seepage quality problem detection method and device, solve how to carry out comprehensive, high-precision, intelligent detection to the technical problem of fine decoration surface empty crack seepage quality problem. Including:1) receiving respectively from visible light acquisition equipment, thermal infrared imaging equipment and three-dimensional laser scanning equipment the visible light image data, thermal infrared image data and three-dimensional laser point cloud data of target fine decoration surface;2) feature extraction is carried out to visible light image data, thermal infrared image data and three-dimensional laser point cloud data;3) get point cloud projection coordinates, then, first two-dimensional feature and second two-dimensional feature are respectively associated with corresponding three-dimensional feature Fusion, generate the fusion point cloud data containing multi-source feature;4) based on fusion point cloud data, the classification identification and spatial orientation of defect are carried out, and detection result is obtained. Realize comprehensive, high-precision detection to the empty crack seepage quality problem of fine decoration surface.

Description

Technical Field

[0001] This invention relates to the fields of image data processing and building engineering quality inspection technology, specifically to a method and device for detecting quality problems such as voids and seepage in fine decoration based on multi-source image data fusion. Background Technology

[0002] The "hollowing, cracking, and seepage" (hollowing, cracking, and seepage) quality problems commonly found on the surfaces of finished residential buildings (such as walls and ceilings) have always been a focus of homeowner complaints and a difficult point in project acceptance. These quality problems not only affect the aesthetics and functionality of the house, but in severe cases, they can even threaten the safety of residents.

[0003] Traditional project acceptance methods mainly rely on manual inspection, such as identifying hollow areas by tapping walls with a sounding hammer or visually inspecting cracks and water stains. This method suffers from problems such as low efficiency, high subjectivity, high rate of missed detections, and difficulty in data retention and traceability.

[0004] Although some technologies have been introduced to detect flatness using 3D laser scanners or to detect leaks using thermal infrared imagers, they are still in the stage of single-point technology application, with low detection accuracy and efficiency, and are difficult to comprehensively detect "cavitation and seepage" problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for detecting quality problems of voids and seepage on the surface of fine decoration based on multi-source image data fusion, and to solve the technical problem of how to conduct comprehensive, high-precision and intelligent detection of quality problems of voids and seepage on the surface of fine decoration.

[0006] The first aspect provides a method for detecting quality problems such as voids and seepage in fine-decorated interior decoration based on multi-source image data fusion, including the following steps:

[0007] 1) Receive visible light image data, thermal infrared image data, and three-dimensional laser point cloud data of the target finished surface, respectively acquired from visible light acquisition equipment, thermal infrared imaging equipment, and three-dimensional laser scanning equipment;

[0008] 2) Use a pre-trained first deep learning network to extract features from the visible light image data to obtain first two-dimensional features representing surface texture and color information; use a pre-trained second deep learning network to extract features from the thermal infrared image data to obtain second two-dimensional features representing surface temperature anomalies; use a pre-trained third deep learning network to extract features from the three-dimensional laser point cloud data to obtain three-dimensional features representing spatial geometric information.

[0009] 3) The 3D laser point cloud data is projected onto the pixel planes of the visible light image data and the thermal infrared image data using a mapping matrix to obtain point cloud projection coordinates. Then, based on the point cloud projection coordinates, the first 2D feature and the second 2D feature are associated and fused with their corresponding 3D features to generate fused point cloud data containing multi-source features. The mapping matrix is ​​constructed based on the extrinsic parameter transformation relationship between the 3D spatial coordinate system of the 3D laser scanning device and the camera coordinate system of the visible light acquisition device and the camera coordinate system of the thermal infrared imaging device, as well as the intrinsic parameter matrices of the visible light acquisition device and the thermal infrared imaging device, to achieve a complete projection transformation from 3D world coordinates to 2D pixel coordinates.

[0010] 4) Based on the fused point cloud data, defects are classified, identified, and spatially located to obtain detection results: For hollow areas, temperature anomaly regions in the thermal infrared image data features are identified, and combined with the spatial convex deformation information corresponding to these regions in the three-dimensional geometric features, a joint judgment is made; for cracks, texture morphology in the visible light image data features is identified, and the location of the crack is determined by combining the spatial coordinates of the mapped three-dimensional laser point cloud data; for leaks, color differences in the visible light image data features are identified, and the leak area is determined by combining the spatial coordinates of the mapped three-dimensional laser point cloud data.

[0011] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, further: the first deep learning network and the second deep learning network both adopt Faster R-CNN network; the third deep learning network adopts PointNet++ network.

[0012] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, the following further describes the construction process of the mapping matrix: First, the rotation matrix R and translation vector t between the three-dimensional laser scanning device, the visible light acquisition device, and the thermal infrared imaging device are obtained through calibration, forming the extrinsic parameter matrix [R|t]; then, the intrinsic parameter matrix K of each image acquisition device in the visible light acquisition device and the thermal infrared imaging device is combined to realize the projection transformation from the three-dimensional point cloud coordinates P_world to the two-dimensional pixel coordinates p_pixel; the calculation formula of the projection transformation is: s·p_pixel=K×[R|t]×P_world, where s is the scale factor; finally, distortion correction parameters are applied to eliminate the influence of lens distortion, and the final pixel-level alignment is achieved through the translation and scaling of pixel coordinates.

[0013] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of cracks and seepage in fine-decorated interior decoration based on multi-source image data fusion, the following further describes the correlation fusion using the PointFusion strategy, specifically including: establishing a correspondence between three-dimensional point cloud coordinates and two-dimensional image pixel coordinates through the mapping matrix; for each three-dimensional point in the fused point cloud data, retrieving its corresponding pixel position in the visible light image and thermal infrared image using the correspondence; assigning the first two-dimensional feature and the second two-dimensional feature at the retrieved pixel position to the three-dimensional point through a feature interpolation method; and using a multilayer perceptron to aggregate the geometric features, texture features, color features, and temperature features attached to each three-dimensional point to generate a global feature vector containing multi-source information.

[0014] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, it further includes: 5) taking the detection results and the corresponding construction material information and construction process data of the target fine decoration surface as input features, and inputting them into a pre-trained random forest regression model to intelligently score and evaluate the fine decoration construction process of the target fine decoration surface.

[0015] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, further: the construction material information and the construction process data specifically include: cement brand, cement grade, sand ratio, curing days, interface agent brand, and number of putty applications; the specific process of intelligent scoring and evaluation includes: encoding and converting categorical variables, and standardizing numerical variables; using Bootstrap sampling to generate multiple subset datasets to train multiple decision trees in the random forest model; and optimizing the hyperparameters of the random forest model, such as the number of decision trees, maximum depth, and minimum number of leaf node samples, through grid search and cross-validation.

[0016] As an optimization and / or instantiation of the aforementioned method for detecting surface defects (cracks and seepage) in finished interior decoration based on multi-source image data fusion, further: the visible light acquisition device, thermal infrared imaging device, and 3D laser scanning device are integrated and installed on a mobile automatic inspection robot platform. The automatic inspection robot platform includes an autonomous navigation module and a lifting gimbal. The lifting gimbal is used to adjust the height and pitch angle of each device during the acquisition process to adapt to the inspection needs of finished interior decoration surfaces with different floor heights. Furthermore, the acquisition process is based on a weighted greedy algorithm combined with GIS spatial analysis. The path planning strategy plans the movement path of the automated inspection robot platform, specifically including: dividing the indoor space to be inspected into a grid, with the grid center as the candidate station of the automated inspection robot platform; constructing a weighting function E=ω0×C-ω1×A-ω2×D, where C is the coverage gain of the candidate station to the unscanned wall area, A is the incident angle, D is the scanning distance, and ω0, ω1, and ω2 are the corresponding weight coefficients; calculating the weight value E of each candidate station, and selecting the set of stations with the optimal weight value and coverage meeting the preset threshold as the acquisition path points to complete the path planning.

[0017] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, further: between step 1) and step 2), a data preprocessing step is also included: performing bilateral filtering on the visible light image data and thermal infrared image data to remove noise and retain edge information; performing statistical outlier removal processing on the three-dimensional laser point cloud data, specifically: for each point in the three-dimensional laser point cloud data, calculating its average distance to its k nearest neighbors, where k ranges from 10 to 50; calculating the average distance of all points, and calculating its mean and standard deviation; setting a threshold of mean + λ × std, where λ is a statistical confidence coefficient and ranges from 1.5 to 3.0, and removing points with an average distance exceeding the threshold as outlier noise points.

[0018] As an optimization and / or instantiation of the above-mentioned method for detecting defects in interior decoration based on multi-source image data fusion, further: In step 4), the joint determination of voids specifically includes: setting a temperature threshold T_th and a deformation threshold D_th; calculating the average temperature difference ΔT in the abnormal area of ​​the thermal infrared image features; if ΔT > T_th, it is determined to be a suspected thermal anomaly area; calculating the local curvature of the point cloud data corresponding to the area or the normal deviation relative to the fitted plane; if the geometric deviation exceeds D_th, it is determined to be a suspected deformation area; if and only if the same location simultaneously meets the conditions of a suspected thermal anomaly area and a suspected deformation area, it is confirmed as a void defect.

[0019] As an optimization and / or instantiation of the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, further: in step 1), the receiving process includes a data synchronization mechanism: the exposure of the three sensor devices—visible light acquisition device, thermal infrared imaging device, and three-dimensional laser scanning device—is synchronously started by a hardware trigger signal, and the timestamps of each sensor device are recorded; for sensor devices with different response speeds, the acquisition delay is compensated by a timestamp alignment algorithm to ensure the consistency of multi-source data in the time dimension; during the acquisition process, the three sensor devices are kept stationary at each station.

[0020] Secondly, a detection device for quality problems of voids and seepage in fine decoration based on multi-source image data fusion is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the detection method for quality problems of voids and seepage in fine decoration based on multi-source image data fusion described in the first aspect.

[0021] By receiving multi-source data from visible light acquisition devices, thermal infrared imaging devices, and 3D laser scanning devices, deep learning networks are used to extract features representing surface texture and color information, surface temperature anomalies, and spatial geometric information. Based on a mapping matrix, the 3D laser point cloud data is projected onto a pixel plane to obtain point cloud projection coordinates, thereby generating fused point cloud data containing multi-source features. Through the correlation and fusion of multi-source heterogeneous data, the limitations of single-method detection are overcome. Based on the fused point cloud data, defect classification, identification, and spatial localization can be achieved. Specifically, by combining temperature anomaly areas and spatial convex deformation information, hollow areas are jointly determined, and cracks and leakage areas are determined by combining texture morphology, color differences, and spatial coordinates, thus achieving comprehensive and high-precision detection of hollow, cracked, and leaking quality problems on fine-decorated surfaces.

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Additional aspects and advantages provided by the present invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice. Attached Figure Description

[0023] The accompanying drawings, which form part of this specification, are used to aid in understanding the invention. The contents provided in the drawings and their related descriptions in this specification can be used to explain the invention, but do not constitute an undue limitation of the invention.

[0024] Figure 1 In this embodiment of the invention, a visible light image of the target finished surface (wall) is acquired by a visible light acquisition device.

[0025] Figure 2In this embodiment of the invention, a thermal infrared image of the target finished surface (wall) is acquired by a thermal infrared imaging device.

[0026] Figure 3 This is a point cloud diagram of the target fine-finished surface (wall) in an embodiment of the present invention.

[0027] Figure 4 This is a three-dimensional feature map obtained in an embodiment of the present invention.

[0028] Figure 5 This is a fused point cloud map generated in an embodiment of the present invention.

[0029] Figure 6 This is the burst point map (detection result) obtained in an embodiment of the present invention.

[0030] Figure 7 This is a schematic diagram of the structure of the automatic inspection robot platform used in an embodiment of the present invention.

[0031] Figure 8 This is a three-dimensional schematic diagram of the automated inspection robot platform used in the method of this embodiment of the invention.

[0032] Figure 9 This is a diagram of the software operation interface for planning the movement path of the automatic inspection robot platform in an embodiment of the present invention.

[0033] Figure 10 This is a hardware architecture diagram of a high-quality interior decoration crack and seepage detection system based on multi-source image data fusion, according to an embodiment of the present invention.

[0034] Figure 11 This is a diagram of the software interface used in an embodiment of the invention to intelligently score and evaluate the finishing process of the target finished surface.

[0035] The following are labeled in the figure: Visible light acquisition device 111; Thermal infrared imaging device 112; Three-dimensional laser scanning device 113. Detailed Implementation

[0036] The present invention will now be clearly and completely described in conjunction with the accompanying drawings. Those skilled in the art will be able to implement the present invention based on these descriptions. Before describing the present invention in conjunction with the accompanying drawings, it should be particularly noted that:

[0037] The technical solutions and features provided in the various sections, including the following description, can be combined with each other without conflict. Furthermore, where possible, these technical solutions, features, and related combinations can be given specific technical subject matter and protected by relevant patents.

[0038] The embodiments of the present invention described below are generally only some embodiments and not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of patent protection.

[0039] The terms "comprising," "including," "having," and any variations thereof in this specification, the corresponding claims, and related sections are intended to cover non-exclusive inclusion. Other related terms and units can be reasonably interpreted based on the relevant content provided in this specification.

[0040] Figure 10 This is a hardware architecture diagram of a high-quality interior decoration crack and seepage detection system based on multi-source image data fusion, according to an embodiment of the present invention. Figure 10 As shown, the high-quality interior decoration crack and seepage detection system based on multi-source image data fusion in this embodiment mainly includes two parts: an automatic inspection robot platform and a user workstation (such as a laptop computer); the two communicate with each other via wireless signal transmission. The user workstation is essentially a high-quality interior decoration crack and seepage detection device based on multi-source image data fusion. The automatic inspection robot platform is the core carrier for data acquisition required by this invention.

[0041] Figure 7 This is a schematic diagram of the automatic inspection robot platform used in the method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion, as described in an embodiment of the present invention. Figure 8 This is a 3D schematic diagram of the automated inspection robot platform used in the method for detecting quality problems such as voids and seepage in fine-decorated interior decoration based on multi-source image data fusion, as described in this embodiment of the invention. Figures 7-8 As shown, the automated inspection robot platform integrates and is equipped with three key sensor devices.

[0042] The first is the visible light acquisition device 111, used to acquire high-definition surface texture and color information (visible light image data) of the target finished surface (such as walls and ceilings), such as... Figure 1 The image shown is a visible light image.

[0043] Secondly, there is a thermal infrared imaging device 112, used to collect thermal infrared image data of the target's finished surface, which can reveal temperature anomalies that are not visible to the naked eye, such as... Figure 2 The thermal infrared image shown.

[0044] Thirdly, a three-dimensional laser scanning device 113 is used to acquire high-precision three-dimensional spatial geometric information of the target fine-finished surface and generate dense three-dimensional laser point cloud data, such as... Figure 3 The point cloud diagram shown.

[0045] To adapt to the inspection needs of finished surfaces at different heights and angles, the automated inspection robot platform is equipped with an autonomous navigation module and a lifting gimbal. The lifting gimbal is used to flexibly adjust the height and pitch angle of the three types of sensor devices during the data acquisition process, ensuring comprehensive coverage of the acquisition field of view.

[0046] More importantly, this physical architecture, which integrates three sensor devices onto a single movable and flexibly adjustable gimbal, is crucial for constructing the mapping matrix in subsequent methods for detecting cracks and seepage in finished interior decoration based on multi-source image data fusion. The mapping matrix is ​​key to multi-source data fusion; it is constructed based on the extrinsic parameter transformation relationships between the 3D spatial coordinate system of the 3D laser scanning device and the camera coordinate systems of the visible light acquisition device and the thermal infrared imaging device, as well as the intrinsic parameter matrices of the visible light acquisition device and the thermal infrared imaging device, achieving a complete projection transformation from 3D world coordinates to 2D pixel coordinates. Since these three sensor devices are physically rigidly fixed to the same gimbal base, their relative positions and angles (i.e., relative extrinsic parameters) remain constant. This means that regardless of how the automated inspection robot platform moves, or how the gimbal adjusts its height and pitch angle to adapt to different detection scenarios, the coordinate transformation relationships between the devices remain stable. This physical integration structure greatly simplifies the calculation of complex mapping matrices in dynamic environments, eliminates calibration errors caused by independent movement of equipment, and ensures that the mapping matrix can always accurately project and align the spatial coordinates of the 3D laser point cloud data onto the pixel plane of the visible light image data and the thermal infrared image data, thus laying the foundation for achieving high-precision and high-reliability multi-source feature association fusion.

[0047] It should be noted that although the above embodiments employ an automated inspection robot platform integrating three sensor devices, the method for detecting quality problems of voids and seepage in finished decoration based on multi-source image data fusion of the present invention is not limited to this data acquisition method. In other embodiments, the automated inspection robot platform can also be replaced by a handheld data acquisition device. This handheld data acquisition device also integrates a visible light acquisition device, a thermal infrared imaging device, and a three-dimensional laser scanning device. These three sensor devices are rigidly fixed on the mechanical frame of the handheld device, and their relative positions and angles (i.e., relative extrinsic parameters) remain constant. The operator can hold the device and, by adjusting the handheld posture and moving position, acquire data from the target finished decoration surface, obtaining visible light image data, thermal infrared image data, and three-dimensional laser point cloud data. Because the three sensor devices are physically rigidly fixed, the coordinate transformation relationship between them remains stable. No matter how the operator moves or adjusts the posture of the handheld device, the mapping matrix can always accurately project and align the spatial coordinates of the three-dimensional laser point cloud data onto the pixel plane of the visible light image data and the thermal infrared image data, thereby achieving high-precision multi-source data fusion. This handheld solution is more lightweight and flexible than the automated inspection robot platform, making it suitable for refined inspection needs in small areas or specific regions. However, it still ensures the accuracy of the mapping matrix construction and the reliability of multi-source feature association and fusion, enabling effective detection of quality issues such as cracks and seepage in finished interior decoration. Therefore, the difference between using an automated inspection robot platform and a handheld data acquisition device lies only in the data acquisition method and does not affect the implementation of the core technical solution of this invention—the multi-source image data fusion detection method based on the mapping matrix.

[0048] Combination Figure 7 and Figure 8 This allows for a clearer understanding of the structure and working principle of the autonomous navigation module and lifting platform in the automatic inspection robot platform, which ensures comprehensive and flexible inspection of the finished surfaces.

[0049] The automated inspection robot platform rests on a four-wheeled mobile chassis. This chassis serves as the physical platform for the autonomous navigation module's movement. Various environmental sensors (such as LiDAR and inertial sensors) are integrated on the chassis to build a real-time map of the surrounding environment (or utilize existing maps) and determine the automated inspection robot platform's position within the map. Based on a pre-set inspection task or path planning algorithm (such as a path planning strategy combining a weighted greedy algorithm with GIS spatial analysis, which will be discussed later), the autonomous navigation module calculates the target path and outputs control commands to the chassis's motor drive system. This controls the rotation speed and steering of the four wheels, enabling the automated inspection robot platform to move forward, backward, and turn autonomously within the indoor space to reach the designated inspection station.

[0050] A gimbal is a composite mechanical device mounted on a mobile chassis. It mainly consists of a lifting mechanism and a gimbal head, and is used to mount and adjust the attitude of three types of sensor equipment (visible light acquisition equipment, thermal infrared imaging equipment, and 3D laser scanning equipment).

[0051] To achieve efficient and automated data acquisition, the high-quality interior decoration crack and seepage detection system based on multi-source image data fusion uses a path planning strategy that combines a weighted greedy algorithm running on the user's workstation with GIS spatial analysis to plan the movement path of the automatic inspection robot platform.

[0052] Figure 9 This is a diagram of the software operation interface for planning the movement path of the automated inspection robot platform in an embodiment of the present invention. Figure 9 As shown, before data collection, the indoor space to be inspected is first divided into a grid, with the center of the grid serving as the candidate station for the automatic inspection robot platform.

[0053] The core of mobile path planning is to construct a weighted function E=ω0×C-ω1×A-ω2×D, where C is the coverage gain of the candidate site for the unscanned wall area, A is the incident angle, D is the scanning distance, and ω0, ω1, and ω2 are the corresponding weight coefficients, set according to the actual scenario requirements. During mobile path planning, the weight value E of each candidate site is calculated, and the set of sites with the optimal weight value and coverage meeting a preset threshold is selected as the collection path points, thus completing the path planning.

[0054] Coverage gain C is calculated using GIS viewshed analysis technology. For each candidate site, the user workstation software simulates the field of view of sensor devices (especially 3D laser scanning devices) and calculates the effective wall area that the site can cover that has not yet been covered by previous sites. The larger this value, the more new information the site can provide, and the higher its weight.

[0055] The incident angle A refers to the angular deviation between the sensor's line of sight and the normal direction of the target finished surface. In optical imaging and laser scanning, direct shooting (i.e., an incident angle close to 0 degrees) yields the clearest texture and the most accurate point cloud data. As the incident angle increases, image distortion and point cloud noise increase. Therefore, this term acts as a penalty (with a negative sign) to guide the automated inspection robot platform to select a position directly facing the wall for data acquisition.

[0056] The scanning distance D refers to the straight-line distance from the candidate site to the target surface. Too large a distance leads to decreased point cloud density and image resolution, making it difficult to identify minute cracks; too small a distance results in an excessively small scan area, increasing the total number of sites and reducing efficiency. By adjusting the weighting coefficient ω2, the scanning distance can be controlled within the optimal depth of field.

[0057] In actual execution, the weighted greedy algorithm is an iterative optimization process: First, calculate the E value among all candidate sites and select the site with the largest E value as the first collection point; then, update the state of the covered area, recalculate the coverage gain C of the remaining candidate sites (the covered area is no longer included in the gain), and select the site with the largest E value again; repeat this step until the cumulative coverage reaches a preset threshold (e.g., above 95%), thereby generating an ordered collection path composed of several optimal sites.

[0058] When the automated inspection robot platform collects data according to the planned path and / or after the data collection is completed, the user workstation receives a massive amount of multi-source image data and performs subsequent data processing (a method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion).

[0059] The method for detecting quality problems of hollow cracks and seepage in fine-decorated interior decoration based on multi-source image data fusion according to embodiments of the present invention includes the following steps:

[0060] 1) Receive visible light image data, thermal infrared image data, and three-dimensional laser point cloud data of the target finished surface, respectively acquired from visible light acquisition equipment, thermal infrared imaging equipment, and three-dimensional laser scanning equipment.

[0061] The automated inspection robot platform moves to each inspection station according to a pre-planned path (based on a weighted greedy algorithm and GIS spatial analysis). Once it reaches the designated location, the lifting platform locks its posture, and a hardware trigger signal synchronously activates three sensor devices for parallel data acquisition. Each sensor device records its acquisition timestamp simultaneously. To ensure consistency of multi-source data across time, a timestamp alignment algorithm is used to compensate for differences in response latency between different sensor devices. At each acquisition station, the automated inspection robot platform remains completely stationary to avoid motion blur caused by minor platform vibrations or movements affecting data quality.

[0062] Visible light acquisition equipment (specifically a Sony Alpha 6000L) captures RGB images of the target finished surfaces (walls, ceilings, etc.). These visible light image data mainly record the surface texture details, color changes, and visible physical forms.

[0063] Thermal infrared imaging equipment (specifically, a P20Max thermal imager) captures thermal infrared images (thermal infrared image data) of the target's finished surface. Due to differences in the heat capacity of different materials, or the presence of voids (air insulation layer) or moisture (evaporation and heat absorption) within the target's finished surface, uneven surface temperature distribution can occur. For example... Figure 2 As shown, thermal infrared images can reveal temperature anomalies that are invisible to the naked eye, which is an important basis for judging voids and leaks.

[0064] A 3D laser scanning device (specifically, the SLAM200 laser radar scanner from Shenzhen Pegasus Robotics Co., Ltd.) emits a laser beam and receives the echo. Distance is measured using either the Time-of-Flight (ToF) method or the phase method, generating dense 3D laser point cloud data. This 3D laser point cloud data accurately records the spatial geometric coordinates (x, y, z) of the target's finished surface, reflecting the flatness, perpendicularity, and minute unevenness of the surface.

[0065] The aforementioned visible light image data, thermal infrared image data, and three-dimensional laser point cloud data are transmitted to the user workstation (a high-quality interior decoration crack and seepage quality problem detection device based on multi-source image data fusion) via the onboard communication module of the automatic inspection robot platform through a wireless network (such as Wi-Fi 6 or 5G) or a wired interface, awaiting further processing.

[0066] 2) Use a pre-trained first deep learning network to extract features from the visible light image data to obtain first two-dimensional features representing surface texture and color information; use a pre-trained second deep learning network to extract features from the thermal infrared image data to obtain second two-dimensional features representing surface temperature anomalies; use a pre-trained third deep learning network to extract features from the three-dimensional laser point cloud data to obtain three-dimensional features representing spatial geometric information.

[0067] Both the first deep learning network (for visible light) and the second deep learning network (for thermal infrared) employ the Faster R-CNN architecture backbone (such as ResNet or VGG). The network extracts feature maps through layers of convolutional and pooling layers. For visible light images, the extracted first two-dimensional features mainly include edge gradients, texture patterns (such as thin stripes in cracks), and color anomalies (such as discoloration caused by leakage). For thermal infrared images, the extracted second two-dimensional features primarily focus on areas of drastic temperature gradient changes, i.e., potential thermal anomaly patches.

[0068] The third deep learning network uses PointNet++. Since point cloud data is unordered, PointNet++ first extracts the geometric features of local regions (such as local normal vector changes and curvature) through hierarchical sampling and grouping, and then aggregates them into global features using a multilayer perceptron (MLP) and max pooling. The resulting 3D features can accurately characterize the unevenness, flatness deviations, and spatial geometric abrupt changes of the wall surface.

[0069] In addition, between step 1) and step 2), a data preprocessing step is included: bilateral filtering is performed on the visible light image data and thermal infrared image data to remove noise and retain edge information; statistical outlier removal is performed on the three-dimensional laser point cloud data, specifically: for each point in the three-dimensional laser point cloud data, the average distance to its k nearest neighbors is calculated, where k ranges from 10 to 50; the average distance of all points is calculated, and its mean and standard deviation are calculated; a threshold is set as mean + λ × std, where λ is a statistical confidence coefficient and ranges from 1.5 to 3.0, and points with an average distance exceeding the threshold are removed as outlier noise points.

[0070] Because the acquisition environment may contain variations in lighting, dust interference, or electronic noise from the equipment itself, the raw data must be cleaned. First, image denoising is performed. For visible light and thermal infrared images, a bilateral filter is used. Unlike ordinary Gaussian filtering, bilateral filtering considers both the spatial proximity of pixels and the similarity of pixel values ​​(color or temperature). This means that it performs smooth denoising in flat areas, while in edge areas (such as crack edges or color block boundaries), due to the large differences in pixel values, the filtering weight is reduced, thus preserving crucial edge information and preventing crack features from being blurred.

[0071] Next, point cloud denoising is performed using a statistical outlier removal algorithm. For 3D point clouds, dust in the air or measurement errors can create "flying points" (outliers). The processing logic is as follows: for each point P in the point cloud, calculate the average distance d from it to its k nearest neighbors, where k is typically between 10 and 50, adjusted according to the point cloud density. Calculate the mean and standard deviation (std) of the average distances of all points. Set a threshold of mean + λ × std, where λ is typically between 1.5 and 3.0, determined based on the scene noise level. Points with an average distance d greater than this threshold are considered outliers and removed. This step significantly improves the accuracy of subsequent geometric feature extraction.

[0072] In the data preprocessing step, temperature standardization based on radiometric calibration was implemented for thermal infrared images. Because the temperature measurement accuracy of thermal infrared imaging equipment is affected by ambient temperature, humidity, and the emissivity of the measured surface, a surface at the same temperature may exhibit different grayscale values ​​under different environmental conditions. To eliminate this systematic bias, environmental parameters (via airborne temperature and humidity sensors) were recorded synchronously at each acquisition station, and a temperature-grayscale value mapping model was established based on Planck's radiation law to convert the pixel grayscale values ​​of the thermal infrared images into absolute temperature values ​​(unit: degrees Celsius), achieving comparability of thermal infrared data across stations and time periods. For visible light images, automatic white balance correction based on the grayscale world assumption was implemented to eliminate color shifts caused by differences in color temperature from different light sources, ensuring that the same material exhibits consistent color characteristics across different stations.

[0073] 3) The 3D laser point cloud data is projected onto the pixel planes of the visible light image data and the thermal infrared image data using a mapping matrix to obtain point cloud projection coordinates. Then, based on the point cloud projection coordinates, the first 2D feature and the second 2D feature are associated and fused with their corresponding 3D features to generate fused point cloud data containing multi-source features. The mapping matrix is ​​constructed based on the extrinsic parameter transformation relationship between the 3D spatial coordinate system of the 3D laser scanning device and the camera coordinate system of the visible light acquisition device and the camera coordinate system of the thermal infrared imaging device, as well as the intrinsic parameter matrices of the visible light acquisition device and the thermal infrared imaging device, to achieve a complete projection transformation from 3D world coordinates to 2D pixel coordinates.

[0074] This is the core step, solving the "alignment" and "fusion" problems of multi-source image data. First, a mapping matrix is ​​constructed. Since the three sensor devices are rigidly integrated on the same lifting platform, their relative positions (rotation and translation) are fixed. The construction process of the mapping matrix includes: first, obtaining the rotation matrix R and translation vector t between the 3D laser scanning device, the visible light acquisition device, and the thermal infrared imaging device through calibration, forming the extrinsic parameter matrix [R|t]; then, combining the intrinsic parameter matrices K of each image acquisition device in the visible light acquisition device and the thermal infrared imaging device, realizing the projection transformation from 3D point cloud coordinates P_world to 2D pixel coordinates p_pixel; the calculation formula for the projection transformation is: s·p_pixel=K×[R|t]×P_world, where s is the scale factor; finally, distortion correction parameters are applied to eliminate the influence of lens distortion, and the final pixel-level alignment is achieved through the translation and scaling of pixel coordinates.

[0075] Secondly, correlation and fusion are performed using the PointFusion strategy. Specifically, this includes: establishing a correspondence between 3D point cloud coordinates and 2D image pixel coordinates through the mapping matrix; for each 3D point in the fused point cloud data, retrieving its corresponding pixel position in the visible light image and thermal infrared image using the correspondence; assigning the first and second 2D features at the retrieved pixel position to the 3D point using a feature interpolation method; and aggregating the geometric features, texture features, color features, and temperature features attached to each 3D point using a multilayer perceptron to generate a global feature vector containing multi-source information.

[0076] 4) Based on the fused point cloud data, defects are classified, identified, and spatially located to obtain detection results: For hollow areas, temperature anomaly regions in the thermal infrared image data features are identified, and combined with the spatial convex deformation information corresponding to these regions in the three-dimensional geometric features, a joint judgment is made; for cracks, texture morphology in the visible light image data features is identified, and the location of the crack is determined by combining the spatial coordinates of the mapped three-dimensional laser point cloud data; for leaks, color differences in the visible light image data features are identified, and the leak area is determined by combining the spatial coordinates of the mapped three-dimensional laser point cloud data.

[0077] To further clarify the physical mechanism of the above classification and recognition, it is necessary to point out the limitations of single sensor data and the complementary advantages of multi-source data fusion. In practical engineering scenarios, single thermal infrared data is easily affected by ambient light, while single visible light images are difficult to quantify the spatial scale of defects. Therefore, the core of this method lies in constructing a multi-dimensional feature vector space: for hollow areas, the "thermal resistance effect" of thermal infrared data and the "geometric deformation" of laser point clouds are used to construct a dual constraint; for cracks and leaks, the high-resolution texture information of visible light is used to compensate for the sparsity of point clouds in subtle features, while the precise depth information of point clouds is used to endow two-dimensional image features with three-dimensional spatial attributes.

[0078] The joint determination of hollow areas is the most typical. First, a temperature threshold T_th and a deformation threshold D_th are set. The algorithm checks the thermal infrared component in the fused features and calculates the average temperature difference ΔT in the local area. If ΔT is greater than T_th, it is marked as "suspected thermal anomaly" (this could be a misjudgment due to direct sunlight). Simultaneously, the algorithm checks the corresponding geometric components of the area, calculating the local curvature or the normal deviation relative to the fitted plane. If this geometric deviation exceeds D_th, it is marked as "suspected deformation" (this could be due to uneven plastering). The determination logic is that the system ultimately confirms a "hollow area" only when the same spatial location simultaneously satisfies both "suspected thermal anomaly" and "suspected deformation." This logic utilizes the physical characteristics of hollow areas, which possess both an air insulation layer (temperature difference) and surface micro-bulges (deformation).

[0079] like Figure 6 As shown, the detection results are presented in the form of a colored point cloud, with different colors corresponding to different degrees of geometric deformation. The deformation scale on the right side of the image uses a two-color gradient mapping; warm colors indicate positive convex deformation, cool colors indicate negative concave or normal flat areas, and the intermediate transition area represents a range of slight fluctuations. Figure 6 The abnormal region at the "hollow" point (within the white dashed box) shows a significant deviation in its normal direction relative to the reference plane, indicating a marked convex deformation. Furthermore, the thermal infrared temperature component carried in the fused features of this region also exhibits abnormally high values ​​(see [link to original text]). Figure 5 The two indicators completely overlap in spatial location, meeting the dual criteria for hollowness, and are therefore automatically marked as "hollowness" defects.

[0080] For crack detection, the main reliance is on texture features (edge ​​detection, elongated connected regions) in visible light. However, by combining point cloud coordinates, the actual length, width, and precise 3D location of the crack within the room can be directly output, rather than just the pixel location on the image.

[0081] To address the spatial localization and geometric quantization of cracks, a 3D crack reconstruction algorithm based on point cloud skeleton extraction can be implemented. The specific process is as follows: First, extract all points classified as "cracks" from the fused point cloud, forming a crack point cloud subset. Second, perform a density-based clustering algorithm (DBSCAN) on this subset to aggregate spatially continuous crack points into independent crack objects, filtering out isolated noise points. Then, project the point cloud of each crack object onto a local fitting plane, and perform a thinning algorithm (such as the Zhang-Suen thinning algorithm) on the 2D plane to extract the crack centerline skeleton. Finally, back-project the 2D skeleton onto 3D space and perform interpolation smoothing along the wall surface to obtain a 3D curve representation of the crack. Based on this 3D curve, calculate the crack's geodesic length (considering the actual length of the wall surface undulations), average width (based on the vertical distance statistics between the crack point cloud and the skeleton), maximum crack depth (based on the point cloud normal deviation), crack direction angle, and its start and end positions in the architectural coordinate system. These quantitative parameters not only provide objective basis for quality assessment, but also support structural safety analysis, such as determining whether cracks extend along the load-bearing wall and whether there are dangerous patterns such as intersecting network cracks.

[0082] For leak detection, the specific area and location of the leak are determined by relying on color differences in visible light features (such as dark water stains) and combining them with point cloud coordinates.

[0083] For precise delineation of leakage areas, a segmentation algorithm based on color space clustering can be employed. Specifically, the visible light image is converted from the RGB color space to the LAB color space, which better matches the human eye's perception of color differences. In the LAB space, water stains caused by leakage typically manifest as a decrease in the L channel (brightness) and a shift in the A channel (red-green) and B channel (yellow-blue) in specific directions. A leakage color feature template is established, and the similarity between each pixel and the leakage template is measured using Mahalanobis distance to generate a leakage probability map. After adaptive threshold segmentation of the probability map, preliminary leakage candidate regions are obtained. Subsequently, morphological closing operations are performed to fill internal holes, and connected component analysis is used to extract independent leakage patches. For each patch, its boundary pixel coordinates are extracted and back-projected to a 3D point cloud space using a mapping matrix. The 3D point set corresponding to the boundary pixels is then found in the point cloud to construct the 3D envelope of the leakage region. Based on this envelope, the projected area (expanded area of ​​the wall), the depth distribution of leakage (the degree of depression of the point cloud relative to the dry area, reflecting the severity of leakage), the spatial location of the leakage area (height from the ground, distance from the corner of the wall) and its spatial relationship with building structural elements (such as window frames, pipe penetration points) are calculated, providing clues for tracing the source of leakage.

[0084] Furthermore, step 4) can further introduce neighborhood feature aggregation and spatial smoothing mechanisms. Specifically, during the classification, identification, and spatial localization of defects, the feature vector of each point is not evaluated in isolation. Instead, a K-nearest neighbor graph is constructed based on three-dimensional spatial distance. The texture, color, and geometric features of neighboring points are integrated into the center point through weighted aggregation, so that the feature vector contains local contextual information. Subsequently, a conditional random field model is used to globally optimize the preliminary classification results. By minimizing the energy function, spatially adjacent points with similar features are forced to maintain semantic label consistency. Finally, the smoothed defect points of the same type are clustered by Euclidean distance to generate independent defect objects. Physical indicators are directly calculated based on the three-dimensional point set. For example, the volume and projected area are calculated using the convex hull of the hollow point set, and the length of the geodesic line along the wall undulation is calculated using the skeleton extraction technology of the crack point set, thereby realizing a complete process from qualitative detection to quantitative evaluation.

[0085] 5) The detection results, along with the corresponding construction material information and construction process data of the target fine-finished surface, are used as input features and input into a pre-trained random forest regression model to intelligently score and evaluate the fine-finished construction process of the target fine-finished surface.

[0086] This step aims to uncover the underlying technological causes of quality problems, achieving a closed loop from "problem discovery" to "source tracing and evaluation." The construction material information and construction process data specifically include: cement brand, cement grade, sand ratio, curing days, interface agent brand, and number of putty applications. This non-image data is correlated with the detected defect data through a digital input system. The specific process of intelligent scoring and evaluation includes: encoding and transforming categorical variables, standardizing numerical variables; using Bootstrap sampling to generate multiple subset datasets to train multiple decision trees in the random forest model; and optimizing the hyperparameters of the random forest model—the number of decision trees, maximum depth, and minimum number of leaf node samples—through grid search and cross-validation.

[0087] Specifically, the training process of the random forest regression model utilizes a historical engineering database, which contains a large number of known combinations of construction parameters and their corresponding final quality acceptance scores. The random forest regression model learns the nonlinear mapping relationship between "construction parameters + defect distribution pattern" and "process score." For example, the model might learn that when "curing days" are insufficient and the "sand ratio" is too high, if large areas of micro-cracks appear in the inspection results, the corresponding process score will significantly decrease. In the prediction phase, the system inputs quantitative indicators such as the currently detected hollow rate, crack density, and leakage area, along with the actual recorded construction materials (e.g., cement grade) and process data (e.g., number of putty applications) for the room, into the model. Multiple decision trees in the model compute in parallel, ultimately outputting a comprehensive score. This score not only reflects the current quality status but also implicitly evaluates the standardization of the construction process, helping managers identify whether the problem stems from materials or improper human operation, thus providing a scientific basis for subsequent rectification and supplier evaluation.

[0088] like Figure 11 The software interface shown demonstrates the practical application results of intelligent scoring and evaluation. Figure 11 The project information is obscured using shadows. The interface presents the evaluation results of multiple construction units in a table format, with each row corresponding to a specific construction area unit. The table includes columns such as serial number, task name, solution name, number of problems, number of risks, number of improvements, score, and operation. The score column is visualized using a five-star system, intuitively reflecting the overall quality level of each construction unit. Based on the number, type, and severity of defects identified in the aforementioned inspection results, combined with the construction material information and process parameters entered in the unit, a quantitative score is automatically calculated using a random forest regression model. This achieves seamless integration from the underlying image inspection data to the high-level process evaluation results, providing project managers with an intuitive decision support tool that enables them to quickly identify weak points in the process and trace the responsible units, forming a closed loop of "inspection-evaluation-improvement" quality management.

[0089] The device for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion.

[0090] The aforementioned detection device for leaks and seepage in finished interior decoration, based on multi-source image data fusion, stores in its memory all pre-trained model files, including Faster R-CNN, PointNet++, and random forest regression models, as well as map data for path planning and calibration parameters for coordinate transformation. The processor, by calling these program codes, coordinates data reception, preprocessing, feature extraction, fusion projection, defect identification, and final process evaluation, encapsulating the complex algorithm flow into a one-click automated operation. Finally, the detection results and detailed quality analysis reports are intuitively displayed on the software interface.

[0091] The foregoing has described the relevant content of the present invention. Those skilled in the art will be able to implement the present invention based on these descriptions. All other embodiments obtained by those skilled in the art based on the foregoing content of this specification without inventive effort should fall within the scope of the present invention.

Claims

1. A method for detecting quality problems such as voids, cracks, and seepage in fine-finished interior decoration based on multi-source image data fusion, characterized in that: Includes the following steps: 1) Receive visible light image data, thermal infrared image data, and three-dimensional laser point cloud data of the target finished surface, respectively acquired from visible light acquisition equipment, thermal infrared imaging equipment, and three-dimensional laser scanning equipment; 2) Use a pre-trained first deep learning network to extract features from the visible light image data to obtain first two-dimensional features that characterize surface texture and color information; The thermal infrared image data is used to extract features by a pre-trained second deep learning network to obtain a second two-dimensional feature characterizing the surface temperature anomaly. The three-dimensional laser point cloud data is used to extract features by a pre-trained third deep learning network to obtain three-dimensional features that represent spatial geometric information. 3) The 3D laser point cloud data is projected onto the pixel planes of the visible light image data and the thermal infrared image data using a mapping matrix to obtain point cloud projection coordinates. Then, based on the point cloud projection coordinates, the first 2D features and the second 2D features are associated and fused with their corresponding 3D features to generate fused point cloud data containing multi-source features. The mapping matrix is ​​constructed based on the extrinsic parameter transformation relationship between the 3D spatial coordinate system of the 3D laser scanning device and the camera coordinate system of the visible light acquisition device and the camera coordinate system of the thermal infrared imaging device, as well as the intrinsic parameter matrices of the visible light acquisition device and the thermal infrared imaging device, respectively. A complete projection transformation from 3D world coordinates to 2D pixel coordinates is achieved. The correlation fusion adopts the PointFusion strategy, which specifically includes: establishing a correspondence between 3D point cloud coordinates and 2D image pixel coordinates through the mapping matrix; for each 3D point in the 3D laser point cloud data, retrieving its corresponding pixel position in the visible light image and thermal infrared image using the correspondence; assigning the first 2D feature and the second 2D feature at the retrieved pixel position to the 3D point through feature interpolation; and aggregating the geometric features, texture features, color features, and temperature features attached to each 3D point using a multilayer perceptron to generate a global feature vector containing multi-source information. 4) Based on the fused point cloud data, defects are classified, identified, and spatially located to obtain detection results: For hollow areas, abnormal temperature regions in the temperature features are identified, and the spatial convex deformation information corresponding to these regions in the geometric features is used for joint judgment; for cracks, texture morphology in the texture features is identified, and the location of the crack is determined by combining the spatial coordinates of the mapped 3D laser point cloud data; for leaks, color differences in the color features are identified, and the leak area is determined by combining the spatial coordinates of the mapped 3D laser point cloud data.

2. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: The first and second deep learning networks both use the Faster R-CNN network; the third deep learning network uses the PointNet++ network.

3. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: The construction process of the mapping matrix includes: first, obtaining the rotation matrix R and translation vector t between the 3D laser scanning device, the visible light acquisition device, and the thermal infrared imaging device through calibration, forming the extrinsic parameter matrix [R|t]; then, combining the intrinsic parameter matrices K of each image acquisition device in the visible light acquisition device and the thermal infrared imaging device, realizing the projection transformation from the 3D point cloud coordinates P_world to the 2D pixel coordinates p_pixel; the calculation formula of the projection transformation is: s·p_pixel=K×[R|t]×P_world, where s is the scale factor; finally, applying distortion correction parameters to eliminate the influence of lens distortion, and achieving the final pixel-level alignment through the translation and scaling of pixel coordinates.

4. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: It also includes: 5) taking the detection results and the corresponding construction material information and construction process data of the target fine decoration surface as input features, and inputting them into a pre-trained random forest regression model to intelligently score and evaluate the fine decoration construction process of the target fine decoration surface.

5. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 4, characterized in that: The construction material information and construction process data specifically include: cement brand, cement grade, sand ratio, curing days, interface agent brand, and number of putty applications; The specific process of intelligent scoring and evaluation includes: encoding and transforming categorical variables, and standardizing numerical variables; using Bootstrap sampling to generate multiple subset datasets to train multiple decision trees in the random forest model; and optimizing the hyperparameters of the random forest model, such as the number of decision trees, maximum depth, and minimum number of leaf node samples, through grid search and cross-validation.

6. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: The visible light acquisition device, thermal infrared imaging device, and three-dimensional laser scanning device are integrated and installed on a mobile automatic inspection robot platform. The automatic inspection robot platform includes an autonomous navigation module and a lifting gimbal. The lifting gimbal is used to adjust the height and pitch angle of each device during the acquisition process to adapt to the inspection needs of fine decoration surfaces with different floor heights. The data acquisition process is based on a path planning strategy that combines a weighted greedy algorithm with GIS spatial analysis to plan the movement path of the automatic inspection robot platform. Specifically, it includes: dividing the indoor space to be inspected into a grid, with the grid center as the candidate station of the automatic inspection robot platform; constructing a weighting function E=ω0×C-ω1×A-ω2×D, where C is the coverage gain of the candidate station on the unscanned wall area, A is the incident angle, D is the scanning distance, and ω0, ω1, and ω2 are the corresponding weight coefficients; calculating the weight value E of each candidate station, and selecting the set of stations with the optimal weight value and coverage meeting the preset threshold as the data acquisition path points to complete the path planning.

7. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: Between step 1) and step 2), a data preprocessing step is also included: bilateral filtering is performed on the visible light image data and thermal infrared image data to remove noise and retain edge information; statistical outlier removal is performed on the three-dimensional laser point cloud data, specifically: for each point in the three-dimensional laser point cloud data, the average distance to its k nearest neighbors is calculated, where k ranges from 10 to 50; the average distance of all points is calculated, and its mean and standard deviation are calculated; a threshold is set as mean + λ × std, where λ is a statistical confidence coefficient and ranges from 1.5 to 3.0, and points with an average distance exceeding the threshold are removed as outlier noise points.

8. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: In step 4), the joint determination of hollow areas specifically includes: setting a temperature threshold T_th and a deformation threshold D_th; calculating the average temperature difference ΔT in the local area; if ΔT > T_th, it is determined to be a suspected thermal anomaly area; calculating the local curvature or normal deviation relative to the fitted plane corresponding to the area; if the geometric deviation exceeds D_th, it is determined to be a suspected deformation area; and confirming a hollow defect only when the same location simultaneously meets the conditions of a suspected thermal anomaly area and a suspected deformation area.

9. The method for detecting quality problems of voids and seepage in fine-decorated interior decoration based on multi-source image data fusion as described in claim 1, characterized in that: In step 1), the receiving process includes a data synchronization mechanism: the exposure of the three sensor devices, namely the visible light acquisition device, the thermal infrared imaging device, and the three-dimensional laser scanning device, is synchronously started by a hardware trigger signal, and the timestamp of each sensor device is recorded; for sensor devices with different response speeds, the acquisition delay is compensated by a timestamp alignment algorithm to ensure the consistency of multi-source data in the time dimension; during the acquisition process, the three sensor devices are kept stationary at each station.

10. A device for detecting quality problems such as voids and seepage in fine-decorated interior decoration based on multi-source image data fusion, characterized in that: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for detecting quality problems of voids and seepage in fine decoration based on multi-source image data fusion as described in any one of claims 1-9.

Images