Vision and infrared combined precast beam appearance defect detection method

Abstract

The invention discloses a visual combined infrared precast beam appearance defect detection method, and relates to the technical field of civil engineering structure detection and nondestructive testing, and the method comprises the following steps: establishing an illumination time sequence observation layer, collecting multi-dimensional optical parameters aiming at the surface of a precast beam, and generating a dynamic reflection fingerprint baseline; and performing coherent phase decomposition based on the dynamic reflection fingerprint baseline, separating the specular reflection signal from the material texture signal, and calibrating light spot track and intensity evolution data in the unified baseline. According to the method, an illumination time sequence observation layer is constructed to generate reflection fingerprints, coherent phase decomposition and anti-fact playback are combined, reflection interference is stripped, and crack boundaries are recovered; time coordinates are reconstructed through double-mirror-image anchor points, and crack evolution is accurately recovered; phase conjugate projection and light field traction are combined, a dynamic threshold optical fence is established, the exposure rhythm is controlled in a closed-loop mode, and precise detection of the defects of the precast beam is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of civil engineering structure detection and nondestructive testing, and particularly relates to a visual and infrared combined precast beam appearance defect detection method. BACKGROUND

[0002] The visual and infrared combined precast beam appearance defect detection refers to, in quality detection of a precast beam, not only visible light images of a beam body surface are collected by using ordinary optical camera equipment to identify cracks, corner drop, honeycomb and other intuitive appearance defects, but also infrared imaging means is introduced to reveal internal hollowing, micro-cracks, delamination and other potential defects which are difficult to be directly observed by naked eyes. The two detection methods complement each other: visual detection can quickly find dominant appearance problems, infrared detection can reveal hidden or early defects, and the combination of the two can realize comprehensive, accurate and nondestructive monitoring of the appearance and near-surface quality of the precast beam, thereby improving the reliability and intelligent level of detection.

[0003] The prior art has the following disadvantages: In the visual detection process of the prior art, it is often disturbed by external moving light sources. When uneven reflection spots form dynamic focusing on the surface of the precast beam, it will have a significant impact on image acquisition. Since the strong reflection covers the area where the cracks are located, the local crack propagation process is shielded, and the detection system cannot accurately extract the true defect features, thereby causing the defect curve to appear distorted. Further, this distortion will directly affect the judgment basis for subsequent maintenance and reinforcement, causing critical hidden dangers to be missed, and there is a serious structural safety risk.

[0004] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a visual and infrared combined precast beam appearance defect detection method to solve the problems in the background.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a visual and infrared combined precast beam appearance defect detection method, comprising the following steps: Establish a light timing observation layer, collect multi-dimensional optical parameters for the surface of the precast beam, and generate a dynamic reflection fingerprint baseline; Based on the dynamic reflection fingerprint baseline, perform coherent phase decomposition to separate the specular reflection signal and the material texture signal, and calibrate the spot trajectory and intensity evolution data in the unified baseline; According to the spot trajectory and intensity evolution data, the counterfactual replay chain is triggered, the micro-exposure image sequence is replaced, the shadow image sequence is generated to identify the crack boundary dynamic drift characteristics in the cross-frame time sequence; For the dynamic drift characteristics of the crack boundary, a double-mirror anchor point with a time-domain symmetric structure is constructed within the image observation window, a unified time coordinate is reconstructed, and the shadow image sequence is reordered according to the crack drift vector to restore the real morphology of the crack; Based on the restored real morphology of the crack, a counter-illumination mask is generated to actively eliminate the highlight reflection signal of the phase conjugate projection system, and a corrected reflectivity distribution image is output; Based on the corrected reflectivity distribution image, time reversal light field traction is implemented in the steady-state observation window, inverse phase micro-patterns are injected, and a dynamic threshold optical fence is constructed by combining a programmable polarized light diaphragm and a shadow shading array to suppress residual spot interference and achieve exposure rhythm closed-loop control, thereby completing the dynamic and accurate detection of the appearance defects of the prefabricated beam.

[0007] Preferably, the dynamic reflection fingerprint baseline generation step is as follows: Place the prefabricated beam in a closed optical observation space, and cover the inner wall of the observation space with anti-reflection material to shield environmental stray light interference; A plurality of light source units are arranged on the top and both sides of the optical observation space, including linearly polarized light sources, circularly polarized light sources, and wide-band adjustable light sources. The light source units are arranged in an angle-progressive manner towards the surface of the prefabricated beam and are sequentially lit to form a controllable time sequence irradiation process; During the light source irradiation process, high-speed imaging devices arranged at multiple angles are used to collect image frames under different irradiation states, and record the corresponding illumination time information, light source angle information, light source type information, and polarization state information to form a time sequence observation image data set; The incident angle, polarization angle, brightness value, and brightness gradient value of the irradiation response region in the image frame are extracted by grid-by-grid, the irradiation response trajectory atlas is generated in time sequence, and the dynamic reflection fingerprint baseline is constructed.

[0008] Preferably, the spot trajectory and intensity evolution data calibration process is as follows: The brightness variation, polarization angle response, and incident angle position variation of the grid region on the surface of the prefabricated beam in each image frame are extracted, and compared with the dynamic reflection fingerprint baseline to identify the region with high brightness, radial diffusion of gradient, and weak amplitude change of polarization angle as the specular reflection interference point; Taking the specular reflection interference point as a reference, the reflection intensity data in the continuous image frames before and after are extracted, the brightness change rate and response time between adjacent frames are analyzed, the high-energy phase interference region with short response time and large brightness fluctuation is identified, and a buffer ring band execution signal stripping transition process is constructed; Trajectory extraction is performed on the specular reflection region in the image sequence, and the spatial position, brightness, polarization angle and incident angle are recorded to generate the trajectory and calibrate the life cycle parameters; The specular reflection trajectory is compared with the dynamic reflection fingerprint baseline, a correction vector is generated according to the average optical parameters of the non-interference region, the trajectory is corrected in reverse and output to the subsequent process.

[0009] Preferably, the crack boundary dynamic drift characteristic generation step is as follows: According to the spot trajectory and its time sequence calibration result, the spot interference region in the image frame sequence is identified, and whether the region overlaps with the expected path of the crack is judged; In the identified interference frame, a plurality of frames of undisturbed images before and after are selected as a reference window, crack boundary points in the target region are extracted, and a crack contour time sequence evolution trajectory is established; Based on the boundary dynamic drift model, the missing area in the interference frame is reconstructed in the predicted boundary form, and the crack real form is restored through interpolation layer and texture filling; The time continuity of the crack boundary in the shadow image sequence is checked, a boundary point mapping matrix is constructed, and the dynamic drift path feature is identified; Uniform image fusion processing is performed on all shadow frames, and the gray scale jump is eliminated and the texture consistency is maintained through gray scale normalization and boundary transition fusion.

[0010] Preferably, the crack real form restoration process is as follows: The time evolution trajectory of the crack boundary is identified in the shadow image sequence, the starting and ending key frames with time symmetry are selected as double-mirror anchor points, and the center coordinates, contour curvature and brightness features of the crack boundary in the corresponding frames are extracted; Based on the trajectory path between the double-mirror anchor points, a time scale is divided, a unified time coordinate axis is established, and the crack boundary in each image frame is mapped to the corresponding time position, completing the time sequence alignment; According to the unified time coordinate, the image frames are reordered, and a crack boundary surface model is constructed by resampling and three-dimensional stacking, realizing the spatio-temporal geometric restoration of the crack real form.

[0011] Preferably, in the construction process of the crack boundary surface model, the boundary cross-section line of the adjacent time scale points is connected by space interpolation, and the crack propagation direction and bifurcation position are judged based on the boundary curvature and brightness change trend.

[0012] Preferably, the reflectivity distribution image output step is as follows: The spatial projection position of the crack boundary in each image is extracted, a time-synchronized retro-illumination mask is constructed, and the brightness abnormal region is calibrated as an intervention target; According to the spatial position and reflection characteristics of each target region in the anti-illumination mask, a phase conjugate reverse wave front is generated by spatial light modulation and projected to the corresponding position to physically suppress the high-brightness reflection signal. Perform gray difference analysis on the suppressed image region, reconstruct the true reflectivity distribution map, and spatially map it with the crack geometry model to complete the reflectivity correction.

[0013] Preferably, based on the corrected reflectivity image, time reversal light field traction is implemented in the steady-state observation window, inverse phase micro-patterns are injected, and a dynamic threshold optical fence is constructed by linking a polarized light shutter and a shading array to suppress residual spot interference as follows: Calibrate the residual spot trajectory in the steady-state observation window and construct a light field inversion path for guiding reverse energy injection; Inject inverse phase micro-patterns along the light field inversion path to form interference regions opposite to the original reflection direction to destroy energy superposition; According to the image brightness and polarization fluctuation, dynamically adjust the polarized light shutter and shading array to construct a dynamic threshold optical fence with time synchronization; Extract the brightness and noise indicators of each region of the image to construct an exposure rhythm standard curve to achieve closed-loop control of the exposure state.

[0014] In the above technical solutions, the present application provides technical effects and advantages: The present application constructs a light timing observation layer, dynamically generates a reflection fingerprint baseline, so that the system can identify and model the reflection behavior under changing light source conditions; through coherent phase decomposition and counterfactual playback chain, the reflection interference and material texture information are effectively stripped, and the real crack boundary is preserved; further, with the help of double mirror anchor points and crack drift vectors, the dynamic evolution process of the crack in the cross-frame image is accurately restored; combined with phase conjugate projection and anti-illumination mask control, active energy dissipation and physical compensation of high-brightness reflection regions are realized; finally, time reversal light field traction and dynamic optical fence regulation mechanism are introduced to construct a closed-loop feedback chain for time sequence exposure and energy regulation. The whole method forms a full-process closed-loop technical path from physical interference perception to visual correction to exposure rhythm regulation, solves the defect information shielding and misjudgment problem caused by dynamic spot interference in traditional visual detection, and ensures the complete presentation and accurate identification of cracks, hollows, delamination and other defects. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments or prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0016] Figure 1 A method flow chart of a prefabricated beam appearance defect detection method combining vision and infrared. DETAILED DESCRIPTION

[0017] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive concept to those skilled in the art.

[0018] The present application provides a prefabricated beam appearance defect detection method combining vision and infrared as shown in Figure 1 The prefabricated beam appearance defect detection method combining vision and infrared includes the following steps: Establish a light timing observation layer, collect multi-dimensional optical parameters for the surface of the prefabricated beam, and generate a dynamic reflection fingerprint baseline; To identify the defect characteristics of the prefabricated beam under complex lighting conditions, a light timing observation layer should be established, optical parameters should be accurately collected, and a dynamic reflection fingerprint baseline should be generated. The specific implementation process is as follows: When constructing the light timing observation layer, the target prefabricated beam is placed in a closed optical observation space, the inner wall of which is coated with full-black anti-reflection material to shield the interference of environmental stray light sources. Nine groups of light source units are arranged at the top and both sides of the space, each group consisting of a linearly polarized light source, a circularly polarized light source, and a wide-band adjustable LED light source, respectively used to project incident light of different polarization directions and different wavebands. Each group of light sources is arranged towards the prefabricated beam by a fixed support in a 15-degree progressive manner, forming 27 different angle incident paths. All light sources are sequentially lit by a time control device, with a duration of 0.5 seconds each time, and an interval of no more than 0.2 seconds between any two times, thereby forming a time-resolved illumination sequence. The execution order of this light timing is fixed, repeated, and controllable, ensuring that the illumination conditions of each angle and waveband can be completely reproduced within the entire observation period. In order to stably record the light response, seven groups of high-speed imaging devices are fixed using a three-axis stabilizer during the light projection process, respectively located at the front, left and right, and above the beam body at an angle of 45 degrees, covering a total range of 270 degrees. Each imaging device uses a full-frame full-exposure mode to collect images, with an image resolution of no less than 8 million pixels and an exposure time of 5 milliseconds uniformly set to ensure that the image brightness remains consistent between frames, avoiding post-interference errors. At the same time, each frame of image automatically records the light time stamp, light source angle, light source type, and corresponding polarization state, which is used as a subsequent light-image sequence matching index.

[0019] During the acquisition of multidimensional optical parameters, a regular coordinate grid was divided on the surface of the precast beam, with each grid area being 10 square millimeters. The reflection response under each illumination condition was analyzed point by point. The incident angle was obtained based on the calculation of the three-dimensional spatial coordinate difference between the preset light source spatial position and the center point of the grid. The angle between the incident light direction and the surface normal was accurately restored using a three-dimensional coordinate transformation formula. The polarization angle was calculated by analyzing the trend of reflection brightness change in the same grid area under different polarized light sources. The corresponding Malus's law response curve was restored using nonlinear fitting, and the direction of maximum polarization response was extracted as the target polarization angle value. The brightness gradient was constructed based on the grayscale change rate of each pixel in the grayscale image under adjacent frames. By comparing the average grayscale change of the current frame and the reference frame in each grid, the spatial distribution map of the brightness gradient value was obtained. All optical parameters were labeled under a unified time index and archived into a three-dimensional tabular dataset according to the grid number, including the illumination frame number, pixel coordinates, incident angle, polarization angle, brightness value, and gradient value, ensuring that each parameter value corresponds to a specific location and temporal state.

[0020] In constructing the dynamic reflection fingerprint baseline, all the collected optical parameters were serialized chronologically, and a time-trajectory map of illumination response was established using a grid system. Each trajectory node consists of a quintuple of timestamp, incident angle, polarization angle, brightness intensity, and brightness gradient. Interpolation was used to ensure the continuity and smoothness of the trajectory curve along the time axis. At the data structure level, each grid corresponds to an independent illumination response trajectory file, containing all optical response data arranged chronologically. Subsequently, the optical parameters in multiple trajectory files were normalized, and abnormal data caused by lens distortion, illegal incidence, or image edge effects were removed. Based on differences in brightness variation patterns, incident angle ranges, and polarization response amplitudes, all trajectories were clustered into eight basic reflection behavior curves, and a standard curve with typical characteristics was generated for each type of reflection behavior. Finally, based on the deviation between all grid trajectories and their respective standard curves, a complete dynamic reflection fingerprint dataset was constructed as a reference baseline for comparing actual reflection behavior in subsequent image frames.

[0021] After establishing the dynamic reflectivity fingerprint baseline, image frames are acquired in real time and compared frame by frame to identify and label potential light spot interference. Specifically, the incident angle, polarization angle, and brightness response value of each grid region in the current frame image are extracted and compared with the expected values ​​of that grid at the corresponding time point in the dynamic fingerprint baseline, calculating the offsets for each parameter. If the offset of any parameter in the current frame exceeds three times the standard deviation of that parameter in the dynamic fingerprint baseline, it is identified as an illumination disturbance area. Subsequently, based on the offset evolution trend of each grid within the disturbance area over several consecutive frames, the movement trajectory of the light spot in the image space is tracked, and its persistence and coverage intensity in the time series are recorded. To avoid misjudgment, the maximum spacing between trajectory breakpoints is limited to within two observation periods, and the interference trajectory is required to drift in the same direction for at least three consecutive periods to be confirmed as a valid light spot trajectory. This method not only accurately identifies the spatial location of bright reflectivity spots but also derives their evolutionary patterns, providing precise data support for subsequent image reconstruction and reflectivity correction.

[0022] Coherent phase decomposition is performed based on the dynamic reflection fingerprint baseline to separate the specular reflection signal from the material texture signal, and the spot trajectory and intensity evolution data are calibrated in the unified baseline. To effectively isolate specular reflection interference and identify the evolution of reflection trajectories, coherent phase decomposition is required based on the established dynamic reflection fingerprint baseline to clearly separate specular reflection signals from material surface texture reflection signals. The specific implementation steps are as follows: For each frame of illuminated image, all grid areas on the surface of the precast beam are analyzed using historical optical response data corresponding to the dynamic reflection fingerprint baseline. The brightness variations, polarization angle response trends, and spatial positional changes of the angle between the incident angle and the normal are analyzed for each area in the current frame. Specifically, based on the illumination response curve of the area under multi-angle illumination in the previous period, the average pixel brightness of that area in the current frame is compared. If the current brightness value is higher than 25% of the historical average, and the brightness gradient exhibits a radial diffusion distribution centered on the pixel focal point, while the polarization angle response of that area does not exceed a five-degree amplitude variation range, then the illumination intensity of that area is determined to be abnormal and not caused by intrinsic material reflection. Accordingly, this area is marked as a specular reflection interference point. Compared to conventional techniques that rely solely on grayscale intensity, this method comprehensively uses multi-parameter behavioral comparison and performs highly reliable identification based on measured physical response differences, effectively avoiding misidentification of bright areas of material texture as light spots.

[0023] To further separate specular reflection signals from material surface texture reflection signals, coherent phase analysis based on temporal continuity is required. In this process, using the identified specular reflection points as reference points, the response data sequence of these pixels in nine consecutive frames under fixed light source angles and polarization states is retrieved, and their reflection intensity change trajectory curve is plotted. In the specular reflection interference region, this curve exhibits a distinct peak structure, with a significant rate of brightness change between consecutive frames. The brightness difference between two adjacent sampling points in the curve is greater than 30%, and the waveform forms and decays rapidly within three frames. In contrast, the material texture region shows a smooth and stable response curve with slow brightness changes and continuous polarization angle changes conforming to the established material scattering laws. Based on this characteristic, a reflection phase change rate threshold is set as the identification standard. Any region where the brightness peak span exceeds the set threshold within a continuous temporal sequence and the response time is less than three frames is identified as a high-energy phase interference region of specular reflection. After such regions are separated from the material texture signal, adjacent signals are prone to mixing at their edges. To avoid error propagation, a five-pixel-wide buffer ring is constructed around each separated region. Its brightness variation trend is extracted and linearly interpolated to achieve signal transition. Unlike conventional edge filtering methods, this processing strategy is based on temporal behavior, giving edge interpolation a physical basis and ensuring the integrity of the material texture signal.

[0024] After phase decomposition and spatial stripping of the specular reflection signal, its spatiotemporal evolution in the image sequence needs to be fully recorded and its trajectory calibrated. To this end, a set of pixels representing all specular reflection regions is extracted from each frame, recording their position coordinates, peak brightness, polarization response value, and incident angle. This data is then compiled into a time-ordered set of reflection events. For regions in consecutive frames that overlap or are spatially adjacent and have similar optical parameters, their spatial centroid offset and brightness change rate are calculated. If the difference is within a set range, they are determined to be the time-continuous trajectory of the same reflective spot. This trajectory is then connected throughout the entire image sequence to form a continuous spot movement path. For each path, its lifecycle information is further extracted, including the first occurrence frame number, maximum brightness frame number, total number of frames in the trajectory, peak brightness, average brightness change slope, and maximum polarization response offset point. All of this information is appended to the trajectory data as time tags and compared bidirectionally with the previously established dynamic reflection fingerprint baseline to identify whether the trajectory deviates from the expected reflection behavior range at each stage. Interference intensity markers are added at the locations where deviations occur. This approach enables dynamic feature modeling of specular reflection behavior, providing precise boundaries for image post-processing and laying the data foundation for dynamic compensation.

[0025] To improve the spatial accuracy and temporal consistency of the light spot trajectory, it is necessary to correct the fitting error between the specular reflection trajectory and the dynamic reflection fingerprint baseline. This operation uses a stable region adjacent to the trajectory path that is not affected by light spot interference as a control group. The average brightness value, average polarization angle, and average incident angle of this region are extracted from the current frame and the fingerprint baseline, and the difference vector of these three parameters is calculated. This vector is used as a correction factor and projected back onto the coordinates and response value of each node on the trajectory path, recalibrating its temporal position and spatial location. If multiple nodes with abrupt shifts exist in a trajectory, they are smoothed by fitting curves to ensure a natural trajectory transition and avoid jumps. The corrected trajectory will be used as the final reflection behavior identification data input into subsequent processing flows to ensure the consistency of the entire reflection recognition process in terms of temporal and physical response.

[0026] Based on the light spot trajectory and intensity evolution data, a counterfactual playback chain is triggered to replace the micro-exposure image sequence and generate a shadow image sequence to identify the dynamic drift characteristics of crack boundaries in cross-frame time series. To address the issue of missing crack boundary identification caused by strong reflective light spots covering images, a counterfactual playback chain needs to be constructed based on the calibrated light spot trajectories and their intensity evolution data. This chain performs temporal replacement processing on small exposed areas within specific frames, thereby generating a shadow image sequence for defect reconstruction. The process includes the following steps: Based on the previous spot trajectory calibration results, all frames exhibiting reflection interference behavior in the image frame sequence were selected. In practice, each spot trajectory was sorted chronologically, and the spatial coverage area of ​​the trajectory in each frame was marked to further determine whether this area overlapped with the expected crack evolution path. To improve the accuracy of the judgment, each frame image was divided into grids, with each grid area set to 10 square millimeters, and mapped to the image resolution. In each interference frame, the grid number covering the spot was matched with the crack boundary crossing trajectory in the preceding and following reference frames. If more than 50% of the grids overlapped, and the pixel brightness value of the area in the current frame was more than 50% higher than the brightness value at the same location in the previous frame, it was considered a "crack boundary obscured" state. All frames meeting this condition and their corresponding grid numbers were used as key target areas for counterfactual playback processing.

[0027] To compensate for missing image information in key areas, multiple undisturbed frames of images preceding and following this area in the time series are selected as a reference window to construct the data foundation for morphological reconstruction. This reference window typically includes three preceding and following frames, totaling six frames, covering the evolution path of the crack boundary along the time axis. In each reference frame, the contour of the crack boundary at the corresponding location in the target area is extracted using gray-level edge difference, combined with the distribution of polarization angle changes in the image, to determine the spatial extension direction of the boundary points. All boundary points are arranged in the order of image frames, and the displacement vector and deflection angle of corresponding boundary points between each pair of adjacent frames are calculated to form the temporal evolution trajectory of the crack contour. Averaging and sliding processing is applied to all trajectory data to extract the overall boundary drift trend, forming a dynamic boundary drift model based on real observations. This model will be used to predict the possible distribution location of the crack boundary in the target interference frame.

[0028] Based on the aforementioned boundary dynamic drift model, a counterfactual playback operation is performed to reconstruct the missing regions in the interference frame as fitted content in a verifiable manner. Specifically, using the crack boundary in the previous frame as the initial reference state, and combining the displacement prediction value from the trajectory model at the current moment, the boundary contour is spatially translated and its angle fine-tuned to generate the expected boundary shape at the corresponding time point of the interference frame. After the boundary is generated, a brightness interpolation layer is constructed within its internal region according to the grayscale value change trend in the preceding and following reference frames. Furthermore, texture patches are extracted from neighboring frames based on the material surface reflection characteristics for spatial filling, ensuring that the reconstructed area is consistent with the overall texture of the image. The final generated content is the counterfactual image structure of the occluded region in the interference frame, replacing the pixel regions in the original image that are distorted by strong reflections. All replaced frames will form a shadow image sequence.

[0029] After constructing the shadow image sequence, the continuity of the reconstructed crack boundaries in all replacement frames needs to be verified on the time axis to identify the dynamic drift characteristics of the crack boundaries during cross-frame evolution. To this end, the coordinates of all boundary points of the crack boundaries in each shadow frame are extracted to construct an inter-frame boundary point mapping matrix. For each boundary point, its closest corresponding point in adjacent shadow frames is found, and its spatial displacement vector and angular change amplitude are calculated to draw the boundary drift path. After trajectory fitting of all drift paths, the velocity change curve, maximum drift amplitude, continuity interruption point, and direction reversal time of each path are extracted. If a crack boundary point exhibits a sudden jump of more than seven pixels in three consecutive frames, or two direction reversals within five frames, the point is determined to have undergone abrupt crack behavior, indicating a potential risk of expansion. This data analysis method based on replacement image reconstruction combined with spatial-temporal linkage can reveal the dynamic crack evolution process that was not identified in the original image due to reflection occlusion, possessing high reliability and timeliness.

[0030] To ensure that the shadow image sequence is consistent with the original image sequence in terms of grayscale levels, texture consistency, and edge transitions, a unified image fusion process must be performed on all reconstructed image frames. During the fusion process, the average grayscale value of each frame in both the original and shadow image sequences is first extracted. Then, the brightness of the shadow frames is normalized based on the grayscale offset to ensure that their overall brightness values ​​are consistent with the original sequence. Within the boundary band of each reconstructed region, a five-pixel-wide transition zone is defined. Edge pixels from the original image are linearly fused with those from the shadow image to eliminate potential grayscale jumps and texture discontinuities at their boundaries. The fused image is then subjected to edge sharpening to ensure that the crack outline remains clear. The resulting shadow image sequence will be consistent with the original image in visual characteristics and data structure, possessing cross-frame consistency and high availability, and can be directly input into the next stage of crack morphology restoration and anti-illumination mask generation.

[0031] To address the dynamic drift characteristics of crack boundaries, a double mirror anchor point with a temporal symmetry structure is constructed within the image observation window to reconstruct a unified time coordinate. The shadow image sequence is then reordered based on the crack drift vector to restore the true crack morphology. To accurately capture the dynamic drift behavior of crack boundaries caused by reflection interference in a time series and to restore the true geometric morphology of the crack, based on the shadow image sequence, it is necessary to construct a dual mirror anchor point with a time-domain symmetric structure, reconstruct a unified time coordinate, and perform time reordering on the image sequence. The specific implementation process is as follows: Based on the cross-frame drift trajectory of crack boundaries in the shadow image sequence, the spatial position change of each crack boundary line segment in consecutive image frames is identified, and two key points with temporal symmetry are selected as double mirror anchor points in its trajectory evolution. In the specific implementation, the boundary coordinates of each frame in the shadow image sequence are first extracted using a gridded method. The extracted region includes the pixel position, contour length, local curvature change, and brightness gradient of the crack boundary line. Subsequently, the boundary lines extracted from adjacent image frames are spatially compared, and the average moving speed and direction of the crack boundary on the time axis are calculated. When a stable direction of movement is found in the crack boundary line across three consecutive frames, and the overall boundary center point coordinate change shows a unidirectional trend, accompanied by an increase in boundary length or an intensification of curvature change, the trajectory segment is determined to have typical dynamic drift characteristics. A pair of anchor points are established in the starting and ending frames of this trajectory segment. These anchor points are symmetrical to each other in time and spatially record the boundary geometric information, center coordinate position, extension angle, and brightness distribution characteristics, respectively. These mirrored anchor points not only represent the start and end of the crack boundary drift process, but also mark the turning point of crack propagation behavior in time, providing a starting and convergence benchmark for the subsequent construction of a unified time coordinate.

[0032] Using established mirror anchor points as start and end points, a complete spatial trajectory path of the crack boundary within the observation time window is constructed, and a unified time coordinate axis is reconstructed accordingly to standardize the temporal relationship between different frames. In this process, multiple time intervals are first divided between the mirror anchor points. Each time interval is interpolated based on the image frame number, further subdividing each time interval into a fixed number of time scale points. The expected boundary center position, boundary morphology change rate, and brightness change value are recorded at each scale point. Then, the actually observed crack boundary information is extracted from each shadow image frame and mapped to the scale point position on the time coordinate axis. If there is a deviation between the actual observation time and the target scale point in a frame, the boundary information of that frame is temporally shifted along the trajectory path direction to ensure alignment of the boundary data on the time axis. This alignment process is not a simple adjustment of the frame order, but rather a precise correction of the position of each frame in the time domain by combining the spatial drift direction and velocity, resulting in a continuous and smooth behavior curve for the crack evolution trajectory on the time axis. The reconstructed unified time coordinate axis reflects the actual expansion process of the crack within the observation window, and is not affected by jitter in the original image acquisition time, inconsistent frame rate, or exposure distortion, thus possessing high physical consistency and behavioral interpretability.

[0033] Based on the reconstructed time axis, all image frames in the shadow image sequence are reordered according to their new positions on the time axis, and geometric restoration of the true crack morphology is performed accordingly. In this step, firstly, spatial contour resampling is performed on the crack boundary in each frame to ensure that the boundary line has a uniform spatial sampling density in the reordered frame sequence. Subsequently, the reordered image data are stacked into a three-dimensional dataset, where the X and Y axes represent the spatial position of the image, and the Z axis represents the position on the unified time axis. The crack boundary forms a continuous surface in this three-dimensional data structure. To restore the crack geometry, cross-sections of the boundary line are extracted from each time scale point within the three-dimensional dataset, and spatial points between adjacent cross-sections are interpolated to construct an equidistant surface model of the crack morphology evolving over time. This model not only reveals the movement path of the crack in the image sequence but also accurately reproduces the crack's behavior characteristics such as extension, bending, bifurcation, or closure in space. Meanwhile, by analyzing the changes in boundary line length, boundary curvature, and brightness decay trend per unit time in the surface model, the crack propagation rate, development direction, and stress release state can be deduced.

[0034] Based on the restored true morphology of the crack, an anti-illumination mask is generated, the phase conjugate projection system is controlled to actively eliminate the high-brightness reflection signal, and a corrected reflectivity distribution image is output. To further compensate for the loss of regional information caused by strong reflection interference in the image, based on the reconstructed true geometry of the crack, a precise back-illumination mask needs to be constructed. This mask is then used to actively suppress the interference region optically through phase conjugate projection, ultimately outputting complete image data with corrected reflectivity. This process includes the following operations: Based on the contour projection region of the crack's true geometry in each image frame, a backlighting mask with spatial boundary accuracy and temporal synchronization is constructed to constrain the scope of subsequent projection operations. Specifically, the positional parameters of the crack boundary line in the two-dimensional image coordinate system are first extracted from each frame, including the boundary line's start and end coordinates, total length, local curvature variation, boundary normal vector, and the corresponding grayscale gradient intensity distribution. Then, symmetrical regions are constructed on both sides of the boundary line, with the expansion range on each side determined by the boundary grayscale gradient descent rate. The initial boundary is set at twice the average crack width by default, but is adaptively adjusted based on the distribution pattern of high-brightness reflective areas if necessary. Within this region, the brightness values ​​of all pixels are differentially analyzed against the average brightness of surrounding non-cracked areas. If the brightness value is more than 30% higher than the surrounding average and exhibits temporal consistency of strong reflection peaks in at least three adjacent frames, the pixel is included in the intervention target mask. All mask points are matched one-to-one with their corresponding frame indices on the time axis, ultimately forming a three-dimensional backlighting mask data volume with complete spatial coordinates, time sequence numbers, and grayscale peak records. This mask differs from traditional static occlusion masks. It is not generated temporarily based on image pixel features, but is jointly constructed by the actual physical boundary of the crack, the historical trajectory matching of light interference, and the time synchronization logic of reflection behavior. It has temporal stability, spatial boundary continuity, and structural feature recognition.

[0035] Based on the intervention area defined by the anti-illumination mask, active phase-conjugate optical projection is performed to physically suppress reflected interference signals and achieve real-time energy dissipation control of bright areas in the image. During operation, for each pixel in the target area of ​​the mask, the brightness value, polarization response direction, incident angle, and reflection angle at the moment of high brightness in the original image frame are obtained, and the corresponding optical propagation path is constructed. Then, by setting the spatial light modulation unit of the laser, a reverse wavefront precisely opposite to this reflection path is generated. This wavefront has the characteristics of being of the same frequency, amplitude, and path as the original highly reflected wave, but with opposite phase. During projection, it is ensured that the reverse beam and the original reflected light form coherent interference in the target area, thereby producing a destructive effect when the two beams are superimposed in space, effectively reducing the reflection intensity in this area. To avoid the accumulation of phase errors, each projection must be controlled within 50% of the image acquisition time interval, and projection preparation must be completed 0.2 milliseconds before each frame acquisition. For image frames with multiple coexisting interference regions, multi-point parallel elimination processing is achieved through sequential partition activation. After each projection, the energy distribution of the result is detected to ensure that the extinction rate reaches more than 80% of the initial reflection intensity. This operation differs from filtering shields or polarization shielding methods. The core lies in using spatial light modulation to construct a precise reverse wavefront to achieve physical energy reduction, rather than visual masking. It has active response, adjustable control, and high-resolution elimination capabilities, and is particularly suitable for precise interference removal in irregular reflective areas on the surface of precast beams.

[0036] After completing the anti-illumination intervention, the entire image is subjected to reflectivity correction to construct a true reflectivity distribution map, which is used to restore the intrinsic reflectivity characteristics of the material and eliminate the influence of light spot disturbance. During implementation, all areas in the image that have undergone anti-illumination processing are first traversed, and the difference in grayscale values ​​before and after the intervention is compared. The interference suppression rate is calculated per point, which is the percentage of the original grayscale minus the grayscale after intervention divided by the original grayscale. Then, the expected reflectivity standard curve is reconstructed by combining the historical brightness values ​​of the area in non-interference frames. Points with abrupt reflectivity changes are subjected to three-frame smoothing interpolation to eliminate high-frequency variations and ensure that reflectivity changes are consistent with the crack boundary morphology. To further enhance the correspondence between the image's reflectivity features and the real physical structure, each corrected pixel is bound to the spatial coordinates of its corresponding crack model. Material feature markers, crack boundary line projection layers, and strong reflectivity boundary layers are added to the reflectivity distribution map, enabling direct pixel-to-structure mapping in subsequent analysis. The final output image not only shows the complete, continuous, and unobstructed crack structure, but also accurately presents the reflective response capability of each location in the grayscale space, providing basic physical parameter support for structural health assessment and defect classification and identification.

[0037] Based on the corrected reflectivity distribution image, time-reversed light field traction is implemented within the steady-state observation window, inverse phase micro-patterns are injected, and a dynamic threshold optical fence is constructed in conjunction with a programmable polarization aperture and a shadow shading array to suppress residual light spot interference and achieve closed-loop control of exposure rhythm, thereby completing the dynamic and accurate detection of appearance defects of precast beams. To comprehensively suppress residual reflection interference and ensure the illumination stability and temporal consistency of image acquisition, time-reversed light field traction is performed within the steady-state observation window based on the corrected reflectivity distribution image. This is combined with inverse phase micropatterns, polarization stops, and a light-blocking array to construct a dynamic threshold optical fence, achieving closed-loop control of the exposure rhythm and completing high-precision closed-loop regulation for defect detection. The specific implementation process is as follows: Based on the reflectivity distribution image after backlighting correction, the residual light spot region is identified, and a reverse path for light field traction is planned within the steady-state observation window. To this end, regions with abnormally high reflectivity are first extracted from all frames in the image sequence. The brightness continuity, center point stability, and local gradient changes of these regions in adjacent frames are compared to select a set of pixels that repeat in at least three frames and whose reflectivity peak exceeds the surrounding average by 30%. These pixels are then arranged chronologically to form a residual light spot trajectory point series, and their initial energy input point and light propagation direction are inferred based on their spatial drift direction. According to this propagation direction, a light field inversion path is constructed within the image observation window, ending at the center of the interference source and starting at the incident point. This path not only provides a coordinate basis for subsequent reverse light energy control but also ensures energy concentration and path accuracy during the reverse projection process, avoiding interference or projection shifts.

[0038] Along the light field traction path, inverse-phase micro-patterns, opposite to the original reflection direction, are injected point by point to generate a controllable coherent interference region within the interference area, disrupting the original energy superposition behavior of the light spot. In implementation, a high-resolution spatial light control device encodes the inverse-phase light patterns frame by frame. The pattern shape is determined by the shape of the interference area identified in the previous stage, and the area of ​​each pattern unit is controlled to be less than 0.5 square millimeters to ensure spatial accuracy and interference effect. The phase angle of each injection point is determined by the historical reflectivity fluctuation curve of that point and undergoes a 180-degree phase reversal. To achieve dynamic tracking and real-time response, the projection time interval of each pattern is controlled within 2 milliseconds and synchronously calibrated according to the time axis of each frame's image acquisition, ensuring temporal consistency between the projected signal and image recording. All inverse-phase micro-patterns spatially constitute a "light energy stripping track" opposite to the interference light path and temporally form a "light interference scene" opposite to the time of interference generation, thereby achieving dynamic dissipation of residual reflection energy at the physical level. Unlike existing methods that use static photomasks or average filtering to reduce brightness, this inverse phase micropattern not only has spatial resolution control capabilities but also temporal coupling properties, enabling it to accurately lock the residual energy loop of each light spot and implement targeted intervention.

[0039] After the temporal injection of the inverse phase light field is completed, to prevent exposure fluctuations caused by new interference or environmental changes, a polarizing stop and shadow shading array with dynamic response capabilities need to be introduced into the image imaging channel to form a set of optical exposure limiting threshold structures that change at any time. In implementation, the entire image is first divided into observation units with an average area of ​​no more than 1 square centimeter based on the statistical distribution of crack boundaries, structural corners, and high-reflectivity textures in the current image frame. The brightness value, brightness fluctuation frequency, and polarization angle change trend of the current frame are extracted from each unit. If a unit experiences a brightness fluctuation exceeding 25% and a polarization direction change exceeding 15 degrees in two consecutive frames, it is determined to be a potential light interference trigger zone. The corresponding polarizing stop adjustment operation is immediately initiated to change the light transmission direction and polarization angle of that area to reduce incident energy. Miniature shading blades are deployed at the same location and quickly inserted or withdrawn by a high-speed driver according to the image frame synchronization signal, forming a fine shading structure. The coordinated action of the polarizing stop and the shading array completes all adjustment actions within the image frame rate, ensuring that the light field state at each exposure moment is within a physically controllable threshold range. Unlike existing methods that automatically adjust the aperture or use fixed light-blocking plates, this method performs local control based on the real-time reflection response of each image region, forming a highly dynamic light suppression mechanism that is region-independent and time-synchronized.

[0040] After completing the light field traction, micro-pattern injection, and grating construction operations, an exposure rhythm closed-loop model is established for the image acquisition process to ensure that the brightness, reflectivity, boundary sharpness, and background noise of each frame remain within the preset standard range. During operation, the average brightness, boundary grayscale gradient, and noise perturbation frequency of all crack main path regions, unstructured background regions, and optical boundary transition zones in the image are extracted to construct a set of standard curves for exposure control. Subsequently, the expected illumination state before each frame acquisition is compared in real time with the image state after acquisition. When any indicator deviates from the standard curve by more than 5%, the system adjusts the exposure time, inverse phase injection intensity, or shading response time before the next frame to achieve automatic adjustment. This closed-loop control method does not rely on a fixed illumination model or employ a constant ambient light scheme; instead, it uses the actual physical response of the image as feedback, giving the illumination control "self-interpretation, self-feedback, and self-correction" capabilities, greatly improving the robustness and long-term stability of dynamic defect identification.

[0041] This invention constructs an illumination time-series observation layer to dynamically generate a reflection fingerprint baseline, enabling the system to identify and model reflection behavior under constantly changing light source conditions. Through coherent phase decomposition and counterfactual playback chains, it effectively removes reflection interference and material texture information, preserving the true boundaries of cracks. Furthermore, by utilizing dual-mirror anchor points and crack drift vectors, it accurately reconstructs the dynamic evolution process of cracks across frames. Combining phase conjugate projection and anti-illumination mask control, it achieves active energy dissipation and physical compensation for high-brightness reflection areas. Finally, it introduces a time-reversal light field traction and dynamic optical fence control mechanism to construct a closed-loop feedback chain for time-series exposure and energy regulation. The entire method, from optical modeling and image processing to projection intervention, forms a closed-loop technical path from physical interference perception to visual correction and exposure rhythm control. This solves the problem of defect information occlusion and misjudgment caused by dynamic light spot interference in traditional visual inspection, ensuring the complete presentation and accurate identification of defects such as cracks, hollow areas, and delamination.

[0042] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for detecting appearance defects in precast beams using a combination of visual and infrared technologies, characterized in that, Includes the following steps: Establish an illumination time-series observation layer, collect multi-dimensional optical parameters for the surface of precast beams, and generate a dynamic reflection fingerprint baseline; Coherent phase decomposition is performed based on the dynamic reflection fingerprint baseline to separate the specular reflection signal from the material texture signal, and the spot trajectory and intensity evolution data are calibrated in the unified baseline. Based on the light spot trajectory and intensity evolution data, a counterfactual playback chain is triggered to replace the micro-exposure image sequence and generate a shadow image sequence to identify the dynamic drift characteristics of crack boundaries in cross-frame time series. To address the dynamic drift characteristics of crack boundaries, a double mirror anchor point with a temporal symmetry structure is constructed within the image observation window to reconstruct a unified time coordinate. The shadow image sequence is then reordered based on the crack drift vector to restore the true crack morphology. Based on the restored true morphology of the crack, an anti-illumination mask is generated, the phase conjugate projection system is controlled to actively eliminate the high-brightness reflection signal, and a corrected reflectivity distribution image is output. Based on the corrected reflectivity distribution image, time-reversed light field traction is implemented within the steady-state observation window, inverse phase micro-patterns are injected, and a dynamic threshold optical fence is constructed in conjunction with a programmable polarization stop and a shadow shading array to suppress residual light spot interference and achieve closed-loop control of exposure rhythm.

2. The method for detecting appearance defects in precast beams using a combination of visual and infrared imaging as described in claim 1, characterized in that, The steps for generating a dynamic reflective fingerprint baseline are as follows: The precast beam is placed in a closed optical observation space, and the inner wall of the observation space is covered with anti-reflective material to shield against environmental stray light interference. Multiple light source units are set at the top and sides of the optical observation space. The light source units include linear polarization light sources, circular polarization light sources and broadband adjustable light sources. The light source units are arranged facing the surface of the precast beam in an angular progressive manner and are lit in sequence to form a controllable time sequence illumination process. During the illumination process, high-speed imaging devices deployed at multiple angles acquire image frames under different illumination conditions, and record the illumination time information, light source angle information, light source type information and polarization state information corresponding to the image frames, forming a time-series observation image data set. By extracting the incident angle, polarization angle, brightness value and brightness gradient value of the illumination response region in the image frame grid by grid, an illumination response trajectory map is generated in time sequence, and a dynamic reflection fingerprint baseline is constructed.

3. The method for detecting appearance defects in precast beams using a combination of visual and infrared imaging as described in claim 2, characterized in that, The calibration process for light spot trajectory and intensity evolution data is as follows: The brightness change, polarization angle response, and incident angle position change of the grid area on the surface of the precast beam in each image frame are extracted and compared with the dynamic reflection fingerprint baseline. The areas with increased brightness, radial gradient diffusion, and weak polarization angle amplitude change are identified as specular reflection interference points. Using specular reflection interference points as a reference, the reflection intensity data in consecutive image frames is extracted, the brightness change rate and response time between adjacent frames are analyzed, high-energy phase interference regions with short response times and large brightness fluctuations are identified, and a buffer ring is constructed to perform signal stripping and transition processing. Trajectory extraction is performed on the specular reflection area in the image sequence, and the spatial position, brightness, polarization angle and incident angle are recorded to generate the trajectory and calibrate the life cycle parameters. The mirror reflection trajectory is compared with the dynamic reflection fingerprint baseline. A correction vector is generated based on the average optical parameters of the non-interference area. The trajectory is then reverse-corrected and output to the subsequent process.

4. The method for detecting appearance defects in precast beams using a combination of vision and infrared technology according to claim 3, characterized in that, The steps for generating dynamic drift characteristics of crack boundaries are as follows: Based on the light spot trajectory and its time sequence calibration results, identify the light spot interference area in the image frame sequence and determine whether the area overlaps with the expected crack path; In the identified interference frames, multiple frames of undisturbed images before and after are selected as reference windows to extract crack boundary points in the target area and establish the temporal evolution trajectory of the crack profile. Based on the boundary dynamic drift model, the missing regions in the interference frames are reconstructed with the predicted boundary morphology, and the real morphology of the crack is restored through interpolation layers and texture filling. Perform temporal continuity verification on crack boundaries in shadow image sequences, construct boundary point mapping matrices, and identify dynamic drift path features; A unified image fusion process is performed on all shadow frames, eliminating grayscale jumps and maintaining texture consistency through grayscale normalization and boundary transition fusion.

5. The method for detecting appearance defects in precast beams using a combination of visual and infrared imaging as described in claim 4, characterized in that, The process of restoring the true shape of the crack is as follows: In the shadow image sequence, the temporal evolution trajectory of the crack boundary is identified. The start and end key frames with temporal symmetry are selected as double mirror anchor points, and the center coordinates, contour curvature and brightness features of the crack boundary in the corresponding frame are extracted. Based on the trajectory path between the two mirror anchor points, a time scale is divided, a unified time coordinate axis is established, and the crack boundary in each image frame is mapped to the corresponding time position to complete the time alignment. The image frames are reordered based on a unified time coordinate, and a crack boundary surface model is constructed by resampling and 3D stacking.

6. The method for detecting appearance defects in precast beams using a combination of vision and infrared technology according to claim 5, characterized in that, In the process of constructing the crack boundary surface model, spatial interpolation is used to connect the boundary cross-sections of adjacent time scale points, and the crack propagation direction and bifurcation position are determined based on the trend of boundary curvature and brightness changes.

7. The method for detecting appearance defects in precast beams using a combination of vision and infrared technology according to claim 5, characterized in that, The steps for outputting the reflectance distribution image are as follows: Extract the spatial projection position of the crack boundary in each frame image, construct a time-synchronized backlighting mask, and mark the brightness anomalous areas as intervention targets; Based on the spatial location and reflection characteristics of each target area in the anti-illumination mask, a phase conjugate reverse wavefront is generated by spatial light modulation and projected onto the corresponding position to physically suppress the high-brightness reflection signal. Gray-level difference analysis is performed on the suppressed image region to reconstruct the true reflectance distribution map, and spatial mapping is performed with the crack geometry model to complete reflectance correction.

8. The method for detecting appearance defects in precast beams using a combination of visual and infrared imaging as described in claim 7, characterized in that, Based on the corrected reflectivity image, time-reversed light field traction is performed within the steady-state observation window, inverse phase micro-patterns are injected, and a dynamic threshold optical fence is constructed by linking the polarization stop and the shading array to suppress residual light spot interference. The steps are as follows: Within the steady-state observation window, the residual light spot trajectory is calibrated and the light field inversion path is constructed to guide the reverse energy injection; Inverse phase micro-patterns are injected along the optical field inversion path to form an interference region opposite to the original reflection direction in order to disrupt energy superposition. Dynamically adjust the polarization stop and the light-blocking array based on image brightness and polarization fluctuations to construct a dynamic threshold optical fence with temporal synchronization; Brightness and noise indices of each region of the image are extracted to construct a standard curve for exposure rhythm.

Images