A quality detection method for cast-in-situ slab with laminated slab based on visual identification

Abstract

The application belongs to the technical field of image recognition, and discloses a quality detection method for cast-in-situ plate strip of laminated slab based on visual recognition, which comprises the following steps: collecting sequence images of the cast-in-situ plate strip, and decomposing the sequence images to obtain gray sequence images and saturation sequence images; calculating frame-by-frame state observation of each spatial grid, and determining the state attribution of each spatial grid frame by frame to obtain a state label sequence and a state transition sequence; extracting each process characteristic component of each spatial grid based on the state label sequence and the state transition sequence, and combining the process characteristic components into a characteristic vector; obtaining a current vibrating position, and performing spatial adaptive standardization on the characteristic vector to obtain a standardized characteristic vector; and judging the under-vibration of each spatial grid based on the standardized characteristic vector, and performing connected domain analysis on the under-vibration grid to obtain an under-vibration area. The application realizes real-time perception of the compactness state of concrete during the vibrating process, and can identify the under-vibration position in time before the initial setting of the concrete.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a method for quality inspection of cast-in-place composite slab strips based on visual recognition. Background Technology

[0002] Composite slabs are precast components widely used in prefabricated buildings. The upper cast-in-place slab strip is poured and formed on-site, and vibration is used to ensure the concrete is fully compacted to guarantee the structural load-bearing capacity. The quality of vibration directly affects the compaction of the cast-in-place slab strip. Under-vibrated areas, due to internal pores and aggregate gaps, result in reduced load-bearing capacity and deteriorated durability, and are one of the main causes of quality defects in cast-in-place structures.

[0003] Existing methods for controlling the quality of concrete compaction primarily rely on the specifications in construction codes regarding the spacing, depth, and duration of vibrator insertion. These methods are performed by construction workers based on experience, lacking direct means of sensing the actual compaction state of the concrete. Some projects assess compaction quality through surface observation or ultrasonic testing after vibration, but these methods are all implemented immediately after vibration. By the time under-vibration defects are discovered, the concrete has already begun to set or is close to setting, leaving extremely limited opportunities for on-site re-vibration and making effective remediation difficult to guarantee. Therefore, how to perceive the compaction state of concrete in different areas in real time during vibration and promptly identify under-vibration locations is a pressing issue that needs to be addressed. Summary of the Invention

[0004] To overcome the aforementioned problems in the prior art, this invention proposes a visual recognition-based method for quality inspection of cast-in-place composite slab strips, which addresses the aforementioned issues.

[0005] This invention provides the following technical solution: A method for quality inspection of cast-in-place composite slab strips based on visual recognition includes: A sequence of images of the cast-in-place slab strip during the vibration process was acquired, and the sequence images were decomposed into grayscale sequence images and saturation sequence images. Based on the grayscale sequence image and the saturation sequence image, calculate the frame-by-frame state observation of each spatial grid. Based on the frame-by-frame state observation, the state attribution of each spatial grid is determined frame by frame, and the state label sequence and state transition sequence of each spatial grid during the vibration process are obtained. Based on the state label sequence and state transition sequence, extract the process feature components of each spatial grid, and combine the process feature components into a feature vector; Obtain the current vibration position, and based on the distance from each spatial grid to the current vibration position, perform spatial adaptive standardization on the feature vector to obtain a standardized feature vector; Based on the standardized feature vectors, under-vibration judgment is performed on each spatial grid, and the under-vibration region is obtained by performing connected component analysis on the under-vibration grid.

[0006] Preferably, the step of calculating the frame-by-frame state observations of each spatial grid based on the grayscale sequence image and the saturation sequence image includes: For each spatial grid in each frame of a grayscale image sequence, calculate the mean brightness within the pixel range covered by the grid to obtain the local mean brightness; calculate the variance of the pixel brightness values ​​to obtain the local texture energy; calculate the absolute value of the difference between the mean brightness of the grid and the corresponding grid in the previous frame to obtain the inter-frame brightness variation amplitude. For each spatial grid in each frame of the saturation sequence image, calculate the average saturation value within the pixel range covered by that grid to obtain the local average saturation value; The mean local brightness, local texture energy, inter-frame brightness variation amplitude, and mean local saturation are used as the frame-by-frame state observations of the spatial grid in that frame.

[0007] Preferably, the step of determining the state affiliation of each spatial grid frame by frame based on the frame-by-frame state observations includes: Five states are predefined: initial stockpiling state, vibration activation state, bubble escape state, slurry upwelling state, and surface stable state. According to the preset five state priority order, the state assignment of each spatial grid in each frame is determined by the threshold combination conditions of local brightness average, local texture energy, inter-frame brightness change amplitude and local saturation average. If the conditions are met, the grid is assigned to the corresponding state and the subsequent determination stops. If none of the priorities are met, the current frame state is set to the previous frame state. The state numbers are assigned in the order of initial stockpiling state, vibration activation state, bubble escape state, slurry upflow state and surface stable state. The state numbers of all frames are arranged in chronological order to obtain the state label sequence. The state label pairs of adjacent frames in the state label sequence are recorded in chronological order to obtain the state transition sequence.

[0008] Preferably, the process characteristic components include: state proportion component, bubble escape event density component, slurry upwelling rate component, state transition structure component, and neighborhood cooperation component.

[0009] Preferably, the extraction of process feature components for each spatial grid based on the state label sequence and state transition sequence includes: For each spatial grid: The proportion of each state in the statistical state label sequence to the total number of vibration frames is used to obtain the state proportion component. In the state label sequence, the ratio of the number of occurrences of inter-frame brightness local maxima events to the total number of frames in the frame segment belonging to the bubble escape state is counted. If the bubble escape state never appears, the value is assigned to zero, and the bubble escape event density component is obtained. In the state label sequence, a linear regression is performed on the local texture energy time series within the frame segment belonging to the slurry upflow state. The absolute value of the regression slope is taken. If the slurry upflow state has never appeared, it is assigned a value of zero to obtain the slurry upflow rate component. Traverse the adjacent frame state number pairs in the state transition sequence, and record them as forward transition, reverse transition and stay in place according to the order of the number pairs. Calculate the proportion of the number of the three types of records to the total number of records in the state transition sequence to obtain the forward advancement ratio, reverse retreat ratio and stay in place ratio, and combine them to obtain the state transition structure components. The neighborhood cooperative component is obtained by statistically analyzing the mean and variance of the surface stable state proportion in the state proportion component of each spatial grid within the preset neighborhood centered on the grid, and the mean and variance of the reverse retreat ratio in the state transition structure component.

[0010] Preferably, the spatial adaptive standardization of the feature vector based on the distance from each spatial grid to the current vibration position includes: Feature vector samples from multiple spatial grids under fully vibrated conditions were pre-collected and grouped according to the distance from each spatial grid to the vibration position. The expected median curves of the bubble escape event density component and the slurry upwelling rate component as a function of distance were fitted respectively. The expected mean curves of the state proportion component, the three types of ratios in the state transition structure component, and the mean of the neighborhood cooperation component as a function of distance were fitted respectively to obtain the feature reference curve. For each spatial grid, the corresponding expected value is retrieved from the characteristic reference curve based on its distance from the current vibration position. The density component of the bubble escape event and the slurry upwelling rate component are divided by the corresponding expected median, and the expected mean is subtracted from the mean of each state proportion component, each transition ratio in the state transition structure component, and each mean in the neighborhood cooperation component to obtain the standardized feature vector.

[0011] Preferably, the undervibration judgment of each spatial grid based on the standardized feature vector includes: The standardized feature vectors of each spatial grid are input into a pre-trained undervibration discriminant classifier to obtain a binary classification result of whether each spatial grid is undervibrated. Spatial grids that are judged to be undervibrated are marked as undervibration grids.

[0012] Preferably, the training method for the undervibration discriminant classifier includes: Collect sequential images of multiple cast-in-place slab strip vibration processes and corresponding internal density detection records from historical construction sites. Based on the internal density detection records, label each spatial grid with binary classification labels of under-vibration or non-under-vibration to construct a training dataset. Normalized feature vectors are extracted from the sequence images of each spatial grid in the training dataset and used as training samples; Using the training samples and the corresponding binary labels as training targets, an undervibration discriminant classifier is trained.

[0013] This invention provides a method for quality inspection of cast-in-place composite slab strips based on visual recognition, which has the following beneficial effects: This invention continuously acquires sequential images of the cast-in-place slab surface during vibration, decomposing the physical process of concrete compaction into five visually distinguishable evolutionary states: initial material accumulation state, vibration activation state, air bubble escape state, slurry uplift state, and surface stability state. It then performs frame-by-frame state assignment for each spatial grid, accumulating state label sequences and state transition sequences. From these, it extracts multi-dimensional process features reflecting the integrity of the compaction process, the activity level of air bubble escape, the rate of slurry uplift, and the dynamics of state transitions, achieving real-time perception of the concrete compaction state in each area during vibration. This invention directly judges the compaction state based on the dynamic visual changes of the concrete surface, without relying on the operator's experience. It can promptly identify under-vibration locations during vibration, allowing construction personnel to implement supplementary vibration in under-vibrated areas before the concrete initially sets. This solves the problem of existing methods that can only passively detect under-vibration defects after vibration is completed, resulting in a very limited window for remediation.

[0014] This invention performs spatial adaptive standardization of feature vectors based on the distance of each spatial grid to the current vibration location. Using the expected value of measured features at each distance under full vibration conditions as a benchmark, the feature components of each grid are converted into the degree of deviation relative to the normal compaction level at the same distance, eliminating the systematic differences in feature values ​​caused by the attenuation of vibration energy with distance. This makes the feature vectors of spatial grids at different distances from the vibration location comparable, enabling the use of a unified judgment standard for under-vibration detection across the entire map. It avoids misjudging normally compacted grids that are far from the vibration location as under-vibration, improving the spatial uniformity and accuracy of under-vibration detection. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for quality inspection of cast-in-place composite slab strips based on visual recognition according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0017] Please see Figure 1 In this embodiment, a method for quality inspection of cast-in-place composite slab strips based on visual recognition includes: S1. Collect sequential images of the cast-in-place slab strip during the vibration process, and decompose the sequential images into grayscale sequential images and saturation sequential images. In this embodiment, the camera is usually fixedly installed directly above the cast-in-place slab strip, with the lens looking down to cover the entire cast-in-place slab strip pouring area, and continuously acquiring color sequence images at a fixed frame rate, with the acquisition period covering the complete vibration process from the insertion to the removal of the vibrator.

[0018] It should be noted that during the compaction of concrete, the surface undergoes an observable sequence of physical morphological changes: before vibration, aggregates are randomly piled up, resulting in a rough, uneven surface with a dark color; after vibration begins, the concrete flows as a whole, causing continuous movement and disturbance on the surface; subsequently, air bubbles escape upwards from the pores and rupture on the surface, producing brief flashes of brightness in localized areas; then, cement paste migrates upwards and covers the aggregates, gradually covering the surface with a grayish-white paste, increasing overall brightness and homogenizing the color; finally, the paste surface stabilizes, and the surface morphology no longer changes. This is reflected in three quantifiable dimensions on the image: brightness distribution reflects the surface lightness and darkness and the state of paste coverage; texture structure reflects the surface roughness and the extent of aggregate exposure; and color saturation reflects the relative coverage ratio of paste and aggregate—the cement paste has low saturation, the aggregate has relatively high saturation, and the saturation decreases as the paste rises.

[0019] The HSV color space decouples luminance and saturation into independent channels, making it easier to extract the two types of information separately. Therefore, for each frame of color image, the RGB three channels are weighted and summed according to the standard luminance weight to obtain a grayscale image. At the same time, the color image is converted to the HSV color space to extract the saturation channel to obtain a saturation image. The grayscale sequence image and the saturation sequence image are then arranged according to the frame order.

[0020] The image is spatially meshed by dividing it into several rectangular grids with a fixed pixel size, for example, each grid covering 16×16 pixels. The selection of the grid size needs to balance spatial resolution and computational stability: if the grid is too small, the number of pixels in each grid will be insufficient, resulting in a large variance in the statistics; if the grid is too large, the spatial resolution will decrease, and it will be impossible to distinguish the density differences between adjacent regions.

[0021] S2. Based on the grayscale sequence image and the saturation sequence image, calculate the frame-by-frame state observation of each spatial grid. The calculation of the frame-by-frame state observations of each spatial grid based on the grayscale sequence image and the saturation sequence image includes: For each spatial grid in each frame of a grayscale image sequence, calculate the mean brightness within the pixel range covered by the grid to obtain the local mean brightness; calculate the variance of the pixel brightness values ​​to obtain the local texture energy; calculate the absolute value of the difference between the mean brightness of the grid and the corresponding grid in the previous frame to obtain the inter-frame brightness variation amplitude. For each spatial grid in each frame of the saturation sequence image, calculate the average saturation value within the pixel range covered by that grid to obtain the local average saturation value; The mean local brightness, local texture energy, inter-frame brightness variation amplitude, and mean local saturation are used as the frame-by-frame state observations of the spatial grid in that frame.

[0022] In this embodiment, local texture energy is characterized by the variance of pixel brightness values ​​within the grid. When aggregate is piled up, the surface is rough and uneven, resulting in a discrete distribution of pixel brightness and a large variance. After the slurry is applied, the surface becomes smoother, the brightness distribution is more uniform, and the variance is smaller. Therefore, the variance can effectively reflect the dynamic changes in the roughness of the concrete surface. The inter-frame brightness variation amplitude is calculated as the absolute value of the difference between the mean brightness values ​​of the same grid in the current frame and the previous frame. This quantity reflects the overall brightness variation intensity of the grid between adjacent frames. Concrete flow and air bubble escape caused by the vibrator will have a significant response to this quantity. For the first frame image, since there is no previous frame, the inter-frame brightness variation amplitude is assigned a value of zero.

[0023] S3. Based on the frame-by-frame state observation, the state attribution of each spatial grid is determined frame by frame to obtain the state label sequence and state transition sequence of each spatial grid during the vibration process. The frame-by-frame determination of state attribution for each spatial grid based on the frame-by-frame state observations includes: Five states are predefined: initial stockpiling state, vibration activation state, bubble escape state, slurry upwelling state, and surface stable state. The judgment priority is determined from high to low according to the strictness of each state judgment condition. The state assignment of each spatial grid is determined in each frame by the threshold combination conditions of local brightness average, local texture energy, inter-frame brightness change amplitude and local saturation average. If the condition is met, the grid is assigned to the corresponding state and the subsequent judgment stops. If none of the priorities are met, the current frame state is set to the previous frame state. The state numbers are assigned in the order of initial stockpiling state, vibration activation state, bubble escape state, slurry upflow state and surface stable state. The state numbers of all frames are arranged in chronological order to obtain the state label sequence. The state label pairs of adjacent frames in the state label sequence are recorded in chronological order to obtain the state transition sequence.

[0024] In this embodiment, the initial material accumulation state corresponds to the state when the concrete has just been poured but has not yet been vibrated. The accumulation of aggregates makes the surface rough and uneven, with high local texture energy and low overall brightness. Furthermore, the lack of vibration results in low inter-frame variation. Therefore, the determination conditions are that the texture energy is higher than the texture high threshold, the average brightness is lower than the first brightness threshold, and the inter-frame variation is lower than the variation low threshold.

[0025] The vibration activation state corresponds to the stage when the vibrator starts to vibrate and the concrete begins to flow. The obvious overall movement of the surface causes the brightness change amplitude between frames to increase significantly. At the same time, the texture structure continues to change during vibration. Therefore, the judgment condition is that the amplitude of the change between frames is higher than the high threshold of the amplitude of change and the absolute value of the difference in texture energy between adjacent frames is higher than the texture fluctuation threshold.

[0026] The bubble escape state corresponds to the stage where bubbles escape upwards from within pores and burst on the surface. The bursting of the bubble generates a brief local brightness maxima. Such events occur frequently within consecutive frames centered on the current location. Therefore, the determination criteria, based on the vibration activation state criteria, include an additional requirement that the frequency of local brightness maxima exceeds a bubble frequency threshold. The determination criteria for the bubble escape state and the vibration activation state are inclusive, with the bubble escape state criteria being more stringent. Therefore, its determination priority is set to the highest to ensure that the bubble escape stage is not mistakenly classified as the vibration activation state.

[0027] The upward surge of the slurry corresponds to the stage where the cement slurry migrates upward and covers the surface. The slurry coverage causes the surface texture energy to continuously decrease, the overall brightness to increase, and the saturation to decrease. Therefore, the judgment condition is that the decrease in texture energy relative to the previous frame is higher than the texture decrease threshold, the average brightness is higher than the second brightness threshold, and the average saturation is lower than the saturation threshold.

[0028] The surface stable state corresponds to the stage where the surface is uniform and stable after compaction is completed. The slurry no longer flows, resulting in extremely low inter-frame variation. The slurry coverage results in high brightness, low texture energy, and low saturation. The judgment conditions correspond to this.

[0029] It should be noted that the priority of the five states needs to be preset to ensure that a unique assignment result can be obtained when the observation of a certain frame simultaneously meets multiple state determination conditions. In this embodiment, the determination priority is arranged from high to low according to the strictness of each state determination condition: the determination condition of the bubble escape state is the most stringent, requiring the simultaneous satisfaction of three conditions: inter-frame variation amplitude, texture fluctuation, and frequency of local brightness maxima, and has the highest priority; the vibration activation state requires the simultaneous satisfaction of two conditions: inter-frame variation amplitude and texture fluctuation, and has the second highest priority; the slurry surge state requires the simultaneous satisfaction of three conditions: texture decrease, average brightness, and average saturation, but the observations involved are different from the previous two, and has the third highest priority; the surface stable state requires the simultaneous satisfaction of four conditions: low inter-frame variation amplitude, high average brightness, low texture energy, and low average saturation, all of which are determinations of a low-activity state, and has the fourth highest priority; the determination condition of the initial material accumulation state requires the satisfaction of high texture energy, low average brightness, and low inter-frame variation amplitude, and has the lowest priority. If none of the five state determination conditions in the current frame are met, the current frame state is set to the previous frame state to maintain the temporal continuity of the state label sequence. This priority arrangement ensures the uniqueness of the determination result when conditions overlap.

[0030] Each threshold needs to be pre-calibrated based on actual construction site conditions. The calibration method involves collecting several frames of images of a known compacted area before construction or during trial vibration, statistically analyzing the typical distribution range of the four observations for each state, and using the boundary values ​​of the typical ranges for each state as the initial thresholds. These thresholds are then fine-tuned during subsequent construction based on the reasonableness of the judgment results. For example, the high texture threshold can be initially set to the 75th percentile of the texture energy of the entire image in the grayscale sequence, and the high variation threshold can be initially set to 3 to 5 times the variation amplitude between frames of the static concrete surface. It should be noted that the state numbers are assigned 0, 1, 2, 3, and 4 sequentially according to the initial material accumulation state, vibration activation state, bubble escape state, slurry uplift state, and surface stable state. The state label sequence is the sequence obtained by arranging the state numbers of each frame for each grid in chronological order during the vibration period, reflecting the complete evolution process of the grid from initial material accumulation to surface stability. The state transition sequence is obtained by recording the number pairs of adjacent frames in the state label sequence in chronological order, where each element represents the transition from the current frame state to the next frame state.

[0031] S4. Extract the process feature components of each spatial grid based on the state label sequence and state transition sequence, and combine the process feature components into a feature vector; The process characteristic components include: state proportion component, bubble escape event density component, slurry upwelling rate component, state transition structure component, and neighborhood cooperation component.

[0032] The extraction of process feature components for each spatial grid based on the state label sequence and state transition sequence includes: For each spatial grid: The proportion of each state in the statistical state label sequence to the total number of vibration frames is used to obtain the state proportion component. In the state label sequence, the ratio of the number of occurrences of inter-frame brightness local maxima events to the total number of frames in the frame segment belonging to the bubble escape state is counted. If the bubble escape state never appears, the value is assigned to zero, and the bubble escape event density component is obtained. In the state label sequence, a linear regression is performed on the local texture energy time series within the frame segment belonging to the slurry upflow state. The absolute value of the regression slope is taken. If the slurry upflow state has never appeared, it is assigned a value of zero to obtain the slurry upflow rate component. Traverse the adjacent frame state number pairs in the state transition sequence, and record them as forward transition, reverse transition and stay in place according to the order of the number pairs. Calculate the proportion of the number of the three types of records to the total number of records in the state transition sequence to obtain the forward advancement ratio, reverse retreat ratio and stay in place ratio, and combine them to obtain the state transition structure components. The neighborhood cooperative component is obtained by statistically analyzing the mean and variance of the surface stable state proportion in the state proportion component of each spatial grid within the preset neighborhood centered on the grid, and the mean and variance of the reverse retreat ratio in the state transition structure component.

[0033] In this embodiment, the state proportion component is the ratio of the number of frames in each of the five states to the total number of vibration frames, and the sum of the five sub-components is 1. A fully compacted mesh should have a significantly higher proportion of surface stable states after the compaction process is completed, while an under-vibrated mesh has a higher proportion of initial stockpile states and a lower proportion of surface stable states because the compaction process is not fully advanced.

[0034] The bubble escape event density component reflects the activity level of the bubble escape process. Within a frame segment belonging to the bubble escape state, a local maximum brightness value occurring between adjacent frames within the neighborhood centered on the current grid (e.g., a 3×3 grid range) is counted as a bubble escape event. The ratio of the total number of events to the total number of frames in that segment is calculated. In under-vibrated regions, due to insufficient skeletal loosening and poor pore connectivity, bubbles have difficulty escaping upwards, resulting in a lower event density; in fully vibrated regions, bubble escape is frequent, resulting in a higher event density.

[0035] The upwelling rate component reflects the rate of decrease in texture energy during the upwelling phase. Within a frame segment belonging to the upwelling state, as the upwelling continuously covers the surface, the texture energy decreases continuously over time. A linear regression is performed with the frame number as the independent variable and the local texture energy as the dependent variable. The regression slope is negative, and its absolute value is taken as the upwelling rate. In well-vibrated areas, the upwelling process is sufficient, the texture energy decreases rapidly, and the rate component is relatively large; in under-vibrated areas, the amount of slurry is insufficient or the upwelling is incomplete, and the rate component is relatively small.

[0036] The state transition structural components are obtained by traversing the state transition sequence and statistically analyzing it. In this example, the state numbers are incremented. For each adjacent frame number pair in the state transition sequence, if the number of the next frame is greater than the number of the previous frame, it is recorded as a forward transition, corresponding to the advancement of the concrete compaction process; if the number of the next frame is less than the number of the previous frame, it is recorded as a reverse transition, corresponding to a state regression; if the numbers are the same, it is recorded as a state in place, corresponding to the state being maintained in the current frame. The proportions of the three types of records to the total number of records in the state transition sequence are statistically analyzed to obtain three sub-components: forward advancement ratio, reverse regression ratio, and in-place stay ratio. The state of a fully compacted grid advances unidirectionally along the physical evolution direction, with a high forward advancement ratio, a reverse regression ratio close to zero, and an in-place stay ratio reflecting the duration of each state. The state of an under-vibrated grid is insufficient to maintain continuous state advancement due to insufficient vibration energy, and the state repeatedly switches between the initial material accumulation state and the vibration-activated state, with a high reverse regression ratio and a low forward advancement ratio. This characteristic is the core basis for distinguishing between under-vibrated and compacted states in the state transition structural components.

[0037] The neighborhood co-operation component utilizes information from spatially adjacent grids to aid in judgment. Under-vibration regions are spatially contiguous; therefore, the compaction index of each grid within a pre-defined neighborhood (e.g., a 5×5 grid range) centered on that grid should exhibit a consistently low characteristic. If the compaction index of a certain grid is abnormal while that of all neighboring grids is normal, it is more likely noise interference than genuine under-vibration, and the variance component plays a supporting role in this identification. Specifically, it statistically analyzes the mean and variance of the proportion of surface stable states in each grid within the neighborhood, as well as the mean and variance of the reverse regression ratio, resulting in four sub-components.

[0038] The five sub-components of the state proportion component, the bubble escape event density component, the slurry upwelling rate component, the three sub-components of the state transition structure component, and the four sub-components of the neighborhood cooperation component are concatenated in a fixed order to obtain the feature vector of each spatial grid.

[0039] S5. Obtain the current vibration position, and based on the distance from each spatial grid to the current vibration position, perform spatial adaptive standardization on the feature vector to obtain a standardized feature vector; In this embodiment, since the vibrator is a point energy source, the grids closer to the vibration position receive stronger vibration energy. Even if the concrete is uniformly dense, the density of air bubble escape events and the rate of slurry rise in the near-end grids will naturally be higher than those in the far-end grids, and the proportion of stable surface states will also be higher. If this distance effect is not eliminated, the classifier will misclassify the far-end grids that are normally dense as under-vibrated. Therefore, it is necessary to use the distance of each grid to the vibration position as a benchmark to convert each component in the feature vector into the degree of deviation relative to the level of full vibration at the same distance.

[0040] The current vibration position is the grid where the vibrator is currently operating, which can be obtained using existing target tracking algorithms. For example, the vibrator in a color sequence image can be detected and tracked. Initial frame detection can be performed using the differences in color and shape between the vibrator and the concrete surface. Then, a target tracking algorithm based on correlation filtering or Kalman filtering can be used to continuously locate the image coordinates of the vibrator in subsequent frames. These coordinates are then mapped to the corresponding spatial grid, thus obtaining the current vibration position grid. The Euclidean distance from the center point of each spatial grid to the center point of the current vibration position grid is multiplied by the actual physical size of the individual grid to obtain the physical distance from each grid to the current vibration position, which serves as the basis for subsequent distance standardization.

[0041] The spatial adaptive standardization of the feature vector based on the distance from each spatial grid to the current vibration position includes: Feature vector samples from multiple spatial grids under fully vibrated conditions were pre-collected and grouped according to the distance from each spatial grid to the vibration position. The expected median curves of the bubble escape event density component and the slurry upwelling rate component as a function of distance were fitted respectively. The expected mean curves of the state proportion component, the three types of ratios in the state transition structure component, and the mean of the neighborhood cooperation component as a function of distance were fitted respectively to obtain the feature reference curve. For each spatial grid, the corresponding expected value is retrieved from the characteristic reference curve based on its distance from the current vibration position. The density component of the bubble escape event and the slurry upwelling rate component are divided by the corresponding expected median, and the expected mean is subtracted from the mean of each state proportion component, each transition ratio in the state transition structure component, and each mean in the neighborhood cooperation component to obtain the standardized feature vector.

[0042] In this embodiment, when collecting feature vector samples of multiple spatial grids under fully vibrated conditions in advance, the samples should be collected evenly according to the distance. That is, the number of samples at each distance level should be roughly the same to avoid the reference statistics being biased towards the near end due to too many samples at the near end, which would result in incomplete elimination of the distance effect.

[0043] The density component of bubble escape events and the slurry upwelling rate component are standardized by dividing by the expected median because these two components exhibit a right-skewed distribution in fully vibrated samples. A small number of over-vibrated samples can significantly increase the mean. The median is robust to outliers and can more stably reflect the typical level of fully vibrated compaction. After standardization, the two components are benchmarked against 1; values ​​greater than 1 indicate a level higher than the level of fully vibrated compaction at the same distance, and values ​​less than 1 indicate a level lower than that level. The means of each state proportion component, each transition ratio of the state transition structure component, and each mean of the neighborhood cooperation component are standardized by subtracting the expected mean. These components are all ratio data with values ​​between 0 and 1, and their distribution is relatively symmetrical. The means can stably reflect the expected level of fully vibrated compaction. After subtraction to eliminate distance offset, the mean is set against 0; positive values ​​indicate higher than expected, and negative values ​​indicate lower than expected. The variance sub-component in the neighborhood cooperation component is not systematically affected by distance. Regardless of the distance from the vibration location, the neighborhood variance of fully compacted areas should be smaller, and that of under-vibrated areas should be larger. Therefore, it is not standardized, and the original values ​​are retained.

[0044] The standardized sub-components are concatenated with the neighborhood co-components and variance sub-components that retain the original values, in a fixed order according to the original eigenvectors, to obtain the standardized eigenvectors.

[0045] S6. Based on the standardized feature vector, determine the undervibration of each spatial grid, and perform connected component analysis on the undervibration grid to obtain the undervibration region.

[0046] In this embodiment, the method for judging under-vibration based on standardized feature vectors can employ various technical means according to actual needs. For example, a threshold-based judgment method can be used, setting an under-vibration judgment threshold for each standardized component. When the standardized value of the initial stockpile state ratio is higher than the set upper limit and the standardized value of the reverse retreat ratio is higher than the set upper limit, it is directly judged as under-vibration. Alternatively, a clustering method can be used to cluster the standardized feature vectors of each grid in the entire map, marking the cluster center with significantly lower compaction features as under-vibration areas. An anomaly detection method can also be used, using the distribution of standardized feature vectors of each grid under fully vibrated conditions as a benchmark, judging grids with deviations exceeding a preset threshold as under-vibration.

[0047] The undervibration judgment of each spatial grid based on the standardized feature vector includes: The standardized feature vectors of each spatial grid are input into a pre-trained undervibration discriminant classifier to obtain a binary classification result of whether each spatial grid is undervibrated. Spatial grids that are judged to be undervibrated are marked as undervibration grids.

[0048] The training method for the undervibration discriminant classifier includes: Collect sequential images of multiple cast-in-place slab strip vibration processes and corresponding internal density detection records from historical construction sites. Based on the internal density detection records, label each spatial grid with binary classification labels of under-vibration or non-under-vibration to construct a training dataset. Normalized feature vectors are extracted from the sequence images of each spatial grid in the training dataset and used as training samples; Using the training samples and the corresponding binary labels as training targets, an undervibration discriminant classifier is trained.

[0049] In this embodiment, a classifier is used for training and judgment. The classifier training dataset is constructed as follows: Historical vibration video recordings from construction sites are collected. The internal density of the cast-in-place slab strip after vibration is tested (e.g., using ultrasonic transmission or core drilling followed by porosity measurement). The test results are mapped to corresponding image space grids. Grids with porosity exceeding a preset limit (e.g., 5%) are labeled as under-vibrated, and the rest are labeled as not under-vibrated, thus constructing binary classification labels. The process of extracting standardized feature vectors from each grid in the training dataset is completely consistent with online inference, including state determination, feature component calculation, and spatial adaptive standardization steps, ensuring consistency between the distribution of training input and inference input. The classifier can employ methods such as Support Vector Machines or Random Forests. The input is a standardized feature vector, and the output is a binary classification result (under-vibrated or not under-vibrated) and the corresponding confidence score.

[0050] When performing connected domain analysis on spatial grids identified as under-vibration, the actual under-vibration region is spatially continuous due to insufficient vibration energy, usually covering multiple adjacent grids. The misjudgment of a single grid is spatially represented as an isolated point. Therefore, spatially adjacent under-vibration grids are merged into connected regions, and isolated regions with an area lower than a preset area threshold (e.g., below 3×3 grids) are filtered to eliminate noise misjudgments. The retained connected regions are the output under-vibration regions.

[0051] As can be seen from the above, this embodiment decomposes the physical process of concrete compaction into five evolutionary states that can be distinguished by visual observation. During the compaction process, state evolution information is accumulated in real time for each spatial grid. Multidimensional process features reflecting the integrity of the compaction process, the activity of bubble escape, the rate of slurry rise, and the dynamics of state switching are extracted from the state label sequence and state transition sequence. Combined with spatial adaptive standardization, the systematic bias caused by the distance attenuation of vibration energy is eliminated. Finally, the under-vibration judgment of each grid is made by a classifier, realizing the real-time perception of the under-vibration position during the compaction process. This allows construction personnel to promptly re-vibrate the under-vibration area before the initial setting of the concrete.

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

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

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

Claims

1. A method for quality inspection of cast-in-place composite slab strips based on visual recognition, characterized in that, include: A sequence of images of the cast-in-place slab strip during the vibration process was acquired, and the sequence images were decomposed into grayscale sequence images and saturation sequence images. Based on the grayscale sequence image and the saturation sequence image, calculate the frame-by-frame state observation of each spatial grid. The calculation of the frame-by-frame state observations of each spatial grid based on the grayscale sequence image and the saturation sequence image includes: For each spatial grid in each frame of a grayscale image sequence, calculate the mean brightness within the pixel range covered by the grid to obtain the local mean brightness; calculate the variance of the pixel brightness values ​​to obtain the local texture energy; calculate the absolute value of the difference between the mean brightness of the grid and the corresponding grid in the previous frame to obtain the inter-frame brightness variation amplitude. For each spatial grid in each frame of the saturation sequence image, calculate the average saturation value within the pixel range covered by that grid to obtain the local average saturation value; The mean local brightness, local texture energy, inter-frame brightness variation amplitude, and mean local saturation are used as the frame-by-frame state observations of the spatial grid in that frame. Based on the frame-by-frame state observation, the state attribution of each spatial grid is determined frame by frame, and the state label sequence and state transition sequence of each spatial grid during the vibration process are obtained. Based on the state label sequence and state transition sequence, extract the process feature components of each spatial grid, and combine the process feature components into a feature vector; Obtain the current vibration position, and based on the distance from each spatial grid to the current vibration position, perform spatial adaptive standardization on the feature vector to obtain a standardized feature vector; Based on the standardized feature vectors, under-vibration judgment is performed on each spatial grid, and the under-vibration region is obtained by connected component analysis of the under-vibration grid.

2. The method for quality inspection of cast-in-place composite slab strips based on visual recognition according to claim 1, characterized in that, The frame-by-frame determination of state attribution for each spatial grid based on the frame-by-frame state observations includes: Five states are predefined: initial stockpiling state, vibration activation state, bubble escape state, slurry upwelling state, and surface stable state. According to the preset five state priority order, the state assignment of each spatial grid in each frame is determined by the threshold combination conditions of local brightness average, local texture energy, inter-frame brightness change amplitude and local saturation average. If the conditions are met, the grid is assigned to the corresponding state and the subsequent determination stops. If none of the priorities are met, the current frame state is set to the previous frame state. The state numbers are assigned in the order of initial stockpiling state, vibration activation state, bubble escape state, slurry upflow state and surface stable state. The state numbers of all frames are arranged in chronological order to obtain the state label sequence. The state label pairs of adjacent frames in the state label sequence are recorded in chronological order to obtain the state transition sequence.

3. The method for quality inspection of cast-in-place composite slab strips based on visual recognition according to claim 2, characterized in that, The process characteristic components include: state proportion component, bubble escape event density component, slurry upwelling rate component, state transition structure component, and neighborhood cooperation component.

4. The method for quality inspection of cast-in-place composite slab strips based on visual recognition according to claim 3, characterized in that, The extraction of process feature components for each spatial grid based on the state label sequence and state transition sequence includes: For each spatial grid: The proportion of each state in the statistical state label sequence to the total number of vibration frames is used to obtain the state proportion component. In the state label sequence, the ratio of the number of occurrences of inter-frame brightness local maxima events to the total number of frames in the frame segment belonging to the bubble escape state is counted. If the bubble escape state never appears, the value is assigned to zero, and the bubble escape event density component is obtained. In the state label sequence, a linear regression is performed on the local texture energy time series within the frame segment belonging to the slurry upflow state. The absolute value of the regression slope is taken. If the slurry upflow state has never appeared, it is assigned a value of zero to obtain the slurry upflow rate component. Traverse the adjacent frame state number pairs in the state transition sequence, and record them as forward transition, reverse transition and stay in place according to the order of the number pairs. Calculate the proportion of the number of the three types of records to the total number of records in the state transition sequence to obtain the forward advancement ratio, reverse retreat ratio and stay in place ratio, and combine them to obtain the state transition structure components. The neighborhood cooperative component is obtained by statistically analyzing the mean and variance of the surface stable state proportion in the state proportion component of each spatial grid within the preset neighborhood centered on the grid, and the mean and variance of the reverse retreat ratio in the state transition structure component.

5. The method for quality inspection of cast-in-place composite slab strips based on visual recognition according to claim 4, characterized in that, The spatial adaptive standardization of the feature vector based on the distance from each spatial grid to the current vibration position includes: Feature vector samples from multiple spatial grids under fully vibrated conditions were pre-collected and grouped according to the distance from each spatial grid to the vibration position. The expected median curves of the bubble escape event density component and the slurry upwelling rate component as a function of distance were fitted respectively. The expected mean curves of the state proportion component, the three types of ratios in the state transition structure component, and the mean of the neighborhood cooperation component as a function of distance were fitted respectively to obtain the feature reference curve. For each spatial grid, the corresponding expected value is retrieved from the characteristic reference curve based on its distance from the current vibration position. The density component of the bubble escape event and the slurry upwelling rate component are divided by the corresponding expected median, and the expected mean is subtracted from the mean of each state proportion component, each transition ratio in the state transition structure component, and each mean in the neighborhood cooperation component to obtain the standardized feature vector.

6. The method for quality inspection of cast-in-place composite slab strips based on visual recognition according to claim 5, characterized in that, The undervibration judgment of each spatial grid based on the standardized feature vector includes: The standardized feature vectors of each spatial grid are input into a pre-trained undervibration discriminant classifier to obtain a binary classification result of whether each spatial grid is undervibrated. Spatial grids that are judged to be undervibrated are marked as undervibration grids.

7. The method for quality inspection of cast-in-place composite slab strips based on visual recognition according to claim 6, characterized in that, The training method for the undervibration discriminant classifier includes: Collect sequential images of multiple cast-in-place slab strip vibration processes and corresponding internal density detection records from historical construction sites. Based on the internal density detection records, label each spatial grid with a binary classification label indicating whether it is under-vibration or not, and construct a training dataset. Normalized feature vectors are extracted from the sequence images of each spatial grid in the training dataset and used as training samples; Using the training samples and the corresponding binary labels as training targets, an undervibration discriminant classifier is trained.

Images