Intelligent inspection method for appearance quality of concrete

Abstract

The invention discloses an intelligent inspection method for concrete appearance quality, and relates to the technical field of concrete appearance quality detection.The method comprises the following steps that the optimal inspection height is configured for intelligent inspection of an unmanned aerial vehicle based on the detection precision requirement and the detection efficiency requirement of voids and pitted surfaces; the unmanned aerial vehicle inspects the surface of the concrete structure according to the preset route and the set optimal inspection height, and the airborne high-resolution imaging system is used for continuously shooting the detection area. Through a dynamic inspection mechanism based on confidence score driving, the unmanned aerial vehicle can actively and adaptively adjust the inspection height and the inspection time when finding a suspected slight pitted surface, the image resolution and the defect feature detail expression capability are improved, the recognition capability of the convolutional neural network on the slight pitted surface and a dark honeycomb is enhanced, and the detection accuracy is improved. The defect omission ratio and the false detection rate in the inspection process are effectively reduced, and accurate, stable and efficient recognition of hidden defects such as slight pitted surfaces and deep honeycombs on the surface of the concrete structure is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete appearance quality detection, and specifically relates to an intelligent inspection method for concrete appearance quality. Background Art

[0002] Intelligent inspection of concrete appearance quality refers to the process of using artificial intelligence, computer vision, image recognition, and automation technologies to automatically and intelligently detect and evaluate the surface quality of concrete structures (such as walls, beams, columns, slabs, etc.). By deploying high-definition cameras, drones, robots, or handheld devices to collect concrete surface images, combined with AI algorithms, automatic identification, classification, positioning, and quantitative analysis of common defects such as cracks, honeycombing, exposed reinforcement, hollowing, and holes are carried out, replacing the traditional inspection method that relies on manual visual inspection and manual recording. This technology can improve the accuracy, consistency, and efficiency of detection, reduce human subjective errors, timely detect and warn of quality hazards, and provide scientific and reliable support for project quality acceptance, maintenance decision-making, and construction process control.

[0003] The existing technologies have the following deficiencies: The existing intelligent inspection technology for concrete appearance quality based on drones usually uses a constant inspection operation height to detect defects on the surface of concrete structures to achieve a compromise between inspection efficiency and recognition accuracy. However, due to the diverse forms of honeycombing on the concrete surface, especially some deep honeycombing (also known as hidden honeycombing), its appearance often only shows slight pitting, which is difficult to be accurately recognized at the conventional inspection height. The drones in the existing technologies lack an adaptive height adjustment mechanism based on the defect recognition results. When the AI detection model initially recognizes slight pitting, it still uses the original constant inspection height to complete the operation, which will result in insufficient image resolution and observation angle, unable to effectively distinguish slight pitting from deep honeycombing, and is extremely prone to missed detection or misdetection, seriously affecting the reliability of the inspection results and the ability to detect early structural safety risks.

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

[0005] The object of the present invention is to provide an intelligent inspection method for the appearance quality of concrete. Through a dynamic inspection mechanism driven by confidence score, the unmanned aerial vehicle (UAV) can actively and adaptively adjust the inspection height and inspection time when detecting suspected slight pitting, thereby significantly improving the image resolution and the ability to express defect feature details. At the same time, by combining feature depth analysis and high-recognizability feature vector input, the recognition ability of the convolutional neural network for slight pitting and hidden honeycombing is enhanced. Finally, it can effectively reduce the defect missed detection rate and false detection rate during the inspection process, realize the accurate, stable and efficient recognition of hidden defects such as slight pitting on the surface of the concrete structure and deep honeycombing, and significantly improve the reliability and practicability of the intelligent inspection of the concrete appearance quality, so as to solve the problems in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: An intelligent inspection method for the appearance quality of concrete, comprising the following steps: Configure the optimal inspection height for the intelligent inspection of the UAV based on the detection accuracy requirement and detection efficiency requirement of honeycombing and pitting; Fly the UAV along a preset route and at the set optimal inspection height to inspect the surface of the concrete structure, and continuously capture the inspection area by using the on-board high-resolution imaging system to form an inspection image data stream; During the inspection process, for each inspection area, collect the image data collected during its inspection process in real time, construct an analysis set for the inspection area, extract the key features that can characterize slight pitting from the analysis set, and perform in-depth analysis on the extracted key features to provide high-recognizability input features for the convolutional neural network and improve the recognition confidence of slight pitting; Input the analyzed features into a pre-trained convolutional neural network, output the confidence score of the inspection area, and based on the score result, identify whether there is slight pitting in the inspection area and perform a suspected slight pitting determination; When it is recognized that there is suspected slight pitting in the inspection area, determine the effective area of slight pitting in the inspection area; Based on the confidence score output by the convolutional neural network, adaptively lower the actual inspection height to improve the spatial resolution of the image, enhance the recognition ability of the subsequent AI detection model, and dynamically adjust the inspection time of the UAV according to the area size of slight pitting in the inspection area and the actual inspection height to ensure the accurate recognition of deep honeycombing.

[0007] Preferably, for the dual requirements of honeycombing and pitting detection accuracy and detection efficiency, configuring the optimal inspection height for the intelligent inspection of the UAV specifically includes the following steps: Combined with the actual engineering requirements, clarify the minimum detection size of honeycombing and pitting defects and the required image resolution; Secondly, calculate the flight altitude range that can meet the resolution requirement based on the parameters of the imaging device carried by the UAV; Then, comprehensively consider the inspection operation efficiency, the spatial characteristics of the detection area, and flight safety to preliminarily determine the optimal inspection altitude; Next, verify the clarity and feature representation ability of the images at the selected altitude through simulation or on-site test flights, ensuring that at this altitude, it can not only guarantee the effective recognition of ordinary honeycomb pitted surfaces but also complete efficient inspections at a reasonable flight speed; Finally, combine the feature distribution of the training data of the AI recognition model to make necessary fine-tuning and optimization of the selected inspection altitude, and finally determine the optimal inspection operation altitude suitable for honeycomb pitted surface detection, providing reasonable initial parameters for subsequent automatic inspection tasks.

[0008] Preferably, for each detection area, extract the key features representing slight pitting from the analysis set formed by the image data. Among them, the extracted features include the discreteness of local texture directional changes and the density of micro-depressions. After in-depth analysis of the extracted key features, generate a texture direction discreteness index and a micro-depression density index respectively. Use the texture direction discreteness index and the micro-depression density index as quantitative indicators reflecting the presence of slight pitting in the detection area, providing highly recognizable inputs for the convolutional neural network and improving the recognition confidence of slight pitting.

[0009] Preferably, input the feature vector composed of the texture direction discreteness index and the micro-depression density index after analysis into the pre-trained convolutional neural network. Output the confidence coefficient through the convolutional neural network, and based on the confidence coefficient, identify whether there is slight pitting in the detection area and perform a suspected slight pitting determination.

[0010] Preferably, compare and analyze the confidence coefficient generated when predicting the presence of slight pitting in the detection area by the pre-trained convolutional neural network with the pre-set confidence coefficient reference threshold, and perform a suspected slight pitting determination. The specific steps are as follows: If the confidence coefficient is greater than the confidence coefficient reference threshold, divide the detection area into suspected slight pitting; If the confidence coefficient is less than or equal to the confidence coefficient reference threshold, divide the detection area into a normal plane.

[0011] Preferably, when it is identified that there is suspected slight pitting in the detection area, determine the effective area of slight pitting in the detection area. The specific steps are as follows: After it is identified that there is suspected slight pitting in the detection area, to further determine its specific effective area of slight pitting, it can usually be completed according to the following steps: First, based on the output result of the convolutional neural network, locate the initial defect position points of suspected slight pitting in the detection area; Secondly, around the initial position point, local area expansion is carried out using information such as the high-frequency features, texture changes, gray-scale differences, and depression density of the image to form a slightly pitted candidate area; Next, an adaptive region growing method or an edge detection + region segmentation algorithm is used to accurately segment the actual contour of the slightly pitted surface and eliminate the background or irrelevant noise regions; Subsequently, connected component analysis is performed on the segmented region to eliminate pseudo-defects with too small an area or abnormal morphology, and only the connected regions that conform to the characteristics of the slightly pitted surface are retained; Finally, parameters such as the actual area, boundary, and position center coordinates of the defect region are calculated to form a complete description of the slightly pitted surface range, providing reliable range information support for subsequent defect level evaluation and inspection strategy adjustment.

[0012] Preferably, after the convolutional neural network outputs a confidence coefficient for the detection region, first compare it with a pre-set confidence reference threshold. If the condition that the confidence coefficient is greater than the confidence coefficient reference threshold is met, it is determined that there is a suspected slightly pitted surface in this region. At this time, to further improve the imaging resolution of the suspected defect region, the following adaptive height reduction strategy is used to adaptively reduce the height of the UAV inspection: , where: is the actual inspection height after adjustment, is the pre-configured optimal inspection height (such as 15m), is the inspection height reduction coefficient (range 0 - 1), used to control the reduction amplitude, is the currently recognized confidence coefficient, is the pre-set confidence reference threshold.

[0013] Preferably, after the height reduction is completed and a high-resolution image is obtained, according to the effective area of the slightly pitted surface within the determined detection region, subsequently, in combination with the current actual inspection height and the effective area of the slightly pitted surface, the inspection stay time of the UAV in this detection region is dynamically determined. The following area-height joint control model is used, and the specific expression is: , where: is the actually adjusted inspection stay time, is the minimum stay time of the UAV under normal circumstances, is the detected effective area of the slightly pitted surface, is the standardized defect area (which can be set according to engineering experience, for example, 50 cm²), is the area ratio adjustment coefficient, controlling the influence of the area on the time, is the height reduction ratio adjustment coefficient, controlling the influence of the reduction degree on the time.

[0014] Preferably, for each detection area, the specific steps for deeply analyzing the local texture directionality change dispersion of the detection area to generate a texture direction dispersion index are as follows: First, divide the detection area image into multiple image blocks of a fixed size (such as 32×32 pixels), and calculate the multi-directional gradient responses for each image block. The common directions include 0°, 45°, 90°, and 135°. Using a directional gradient filter (such as Sobel or Gabor transform), calculate the directional energy response values of the image block in each direction respectively, and form the directional energy vector of the image block E , , where represents the texture energy in the direction . Then, aggregate the directional energy vectors of all image blocks in the entire detection area to construct a global directional energy distribution model; Normalize the energy in each direction to obtain the directional probability distribution weights, and calculate the texture direction dispersion index based on the directional probability distribution weights as a quantitative index reflecting the presence of slight pitting in the detection area. The calculation expression of the texture direction dispersion index is: , where: is the directional probability distribution weight, and the calculation expression is: , is a non-linear adjustment factor that adjusts the sensitivity to small probability directions, is the texture direction dispersion index, which reflects the degree of texture direction disorder in the detection area. The exponential form strengthens the response to direction perturbations "when approaching an equilibrium distribution" and is suitable for capturing slight anomalies.

[0015] Preferably, for each detection area, the specific steps for deeply analyzing the micro-pit distribution density of the detection area to generate a micro-pit density index are as follows: Extract the local micro-pit features on the surface by performing a morphological bottom-hat transform on the high-resolution image in the detection area to obtain a set of micro-pit candidate regions . Subsequently, based on the spatial geometric constraints of the micro-pits, screen out an effective set of micro-pit points that meet the depression depth threshold and the minimum depression area threshold , , that is, satisfying: , where: represents the maximum depth of the micro-pit point , represents the area of the micro-pit point , and are the depression depth threshold and the minimum depression area threshold respectively, which are used to screen out the effective micro-depression point set; For the effective micro-depression point set Perform spatial distribution analysis. By calculating the density and distribution compactness of the micro-depression point set, a micro-depression density index is comprehensively constructed. The expression for constructing the micro-depression density index is: , where: is the micro-depression density index, represents the number of effective micro-depression point sets, A is the actual detection area of the detection region, is the average nearest neighbor distance of the depression point set, that is: , where: represents the micro-depression point to the distance of its nearest neighbor micro-depression point, is the compactness enhancement coefficient, which is used to balance the importance of density and distribution aggregation.

[0016] In the above technical solution, the technical effects and advantages provided by the present invention are: Through the dynamic inspection mechanism driven by confidence score, the present invention enables the drone to actively and adaptively adjust the inspection height and inspection time when detecting suspected slight pitting, thereby significantly improving the image resolution and the ability to express defect feature details. At the same time, combined with feature depth analysis and high-identification feature vector input, the recognition ability of the convolutional neural network for slight pitting and dark honeycombing is enhanced. Finally, it can effectively reduce the defect missed detection rate and false detection rate during the inspection process, realize the accurate, stable and efficient recognition of hidden defects such as slight pitting and deep honeycombing on the surface of the concrete structure, and significantly improve the reliability and practicability of the intelligent inspection of the concrete appearance quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0018] Figure 1 is the method flow chart of an intelligent inspection method for the concrete appearance quality of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art.

[0020] The present invention provides an intelligent inspection method for the appearance quality of concrete as Figure 1 shown, comprising the following steps: Based on the detection accuracy requirements and detection efficiency requirements for honeycombing and pitting, configure the optimal inspection altitude for the intelligent inspection of the unmanned aerial vehicle (UAV). The specific steps for configuring the optimal inspection altitude for the intelligent inspection of the UAV in view of the dual requirements of detection accuracy and detection efficiency for honeycombing and pitting are as follows: First, in combination with the actual engineering requirements, clarify the minimum detection size of honeycombing and pitting defects and the required image resolution (for example, it is required to identify pitting features of ≥5 mm); Second, calculate the flight altitude range that can meet this resolution requirement based on the parameters of the imaging equipment carried by the UAV (such as camera resolution, focal length, and field of view angle); Then, comprehensively consider the inspection operation efficiency, the spatial characteristics of the inspection area (such as building height, shape complexity), and flight safety, and preliminarily determine the optimal inspection altitude; Next, verify the clarity and feature representation ability of the images at the selected altitude through simulation or on-site test flights to ensure that at this altitude, the effective identification of ordinary honeycombing and pitting can be guaranteed, and efficient inspection can be completed at a reasonable flight speed (for example, it can be determined through simulation analysis that the image resolution of a certain concrete surface at a height of 15 m can meet the effective resolution ability for features >5 mm, that is, set 15 m as the initial optimal inspection altitude for this scenario); Finally, in combination with the feature distribution of the AI recognition model training data, make necessary fine-tuning and optimization of the selected inspection altitude, and finally determine the optimal inspection operation altitude applicable to the detection of honeycombing and pitting, providing reasonable initial parameters for subsequent automatic inspection tasks.

[0021] Fly the UAV to inspect the surface of the concrete structure according to the preset flight path and the set optimal inspection altitude, and use the on-board high-resolution imaging system (such as RGB camera, multispectral camera) to continuously capture the inspection area to form an inspection image data stream. During the inspection of the concrete structure surface, to ensure the comprehensiveness and high quality of the inspection image data, the real-time flight path of the drone is controlled to dynamically adjust the flight attitude and shooting rhythm, ensuring that the image overlap rate and flight strip coverage during the inspection meet the predetermined requirements. Specifically, the overlap rate refers to the overlapping part of adjacent captured images in space, usually between 70% and 85%, to ensure the stability of image stitching, feature matching, and 3D modeling; the flight strip coverage refers to the flight line distribution that can comprehensively cover the detection area of the concrete surface, avoiding image blind spots or inspection dead ends. By real-time monitoring the position, attitude, and environmental conditions of the drone, combined with the path optimization ability of the flight control system, the flight path and shooting points can be dynamically fine-tuned according to the geometric characteristics of the concrete surface, local complexity, or the undulation of the building, so as to achieve the acquisition of non-missing, high-definition, and continuous images of the concrete structure surface, providing high-quality and complete raw data guarantee for subsequent defect detection, feature extraction, and AI detection models.

[0022] During the inspection process, for each detection area, the image data collected during the inspection is real-time aggregated to construct an analysis set for the detection area. Key features that can characterize slight pitting are extracted from the analysis set, and the extracted key features are deeply analyzed to provide highly recognizable input features for the convolutional neural network, improving the recognition confidence of slight pitting. For each detection area, key features that characterize slight pitting are extracted from the analysis set formed by the image data. Among them, the extracted features include the dispersion degree of local texture directional change and the distribution density of micro depressions. After deeply analyzing the extracted key features, a texture direction dispersion index and a micro depression density index are respectively generated. The texture direction dispersion index and the micro depression density index are used as quantitative indicators to reflect the presence of slight pitting in the detection area, providing highly recognizable input for the convolutional neural network and improving the recognition confidence of slight pitting.

[0023] For each detection area, if the dispersion degree of local texture directional change in this detection area is relatively large, it usually means that the risk of slight pitting in this detection area is greater. The reason is that on the surface of normally cast and formed concrete, the mortar distribution and surface forming process will make the surface texture direction relatively uniform, and the texture direction has continuity and smoothness; while slight pitting is usually due to insufficient mortar flow, local exposure of stones, or inadequate vibration, resulting in irregular unevenness, particle exposure, or tiny holes on the surface. These defects will destroy the original texture continuity, making the local texture direction show diverse, disordered, or divergent distribution, and thus significantly increasing the texture direction dispersion degree (i.e., the statistical distribution difference of textures in different directions). Therefore, when a relatively large texture direction dispersion degree appears in the detection area, it usually reflects the risk of slight pitting caused by construction defects on the surface of this area.

[0024] For each detection area, the specific steps for deeply analyzing the discreteness of the local texture directionality change in the detection area to generate a texture direction discreteness index are as follows: First, divide the detection area image into multiple image blocks of a fixed size (such as 32×32 pixels). Calculate the multi-directional gradient responses for each image block. The common directions include 0°, 45°, 90°, and 135°. Using a directional gradient filter (such as Sobel or Gabor transform), calculate the directional energy response value of the image block in each direction respectively to form the directional energy vector of the image block E , , where represents the texture energy in the direction . Then aggregate the directional energy vectors of all image blocks in the entire detection area to construct a global directional energy distribution model; The higher the flatness of this distribution, the smaller the difference in texture energy in each direction, the more chaotic the directionality, and the greater the possibility of slight pitting.

[0025] Normalize the energy in each direction to obtain the direction probability distribution weight. Calculate the texture direction discreteness index based on the direction probability distribution weight as a quantitative index reflecting the presence of slight pitting in the detection area. The calculation expression of the texture direction discreteness index is: , where: is the direction probability distribution weight, and the calculation expression is: , is a non-linear adjustment factor that adjusts the sensitivity to small probability directions, is the texture direction discreteness index, which reflects the degree of chaos in the texture directionality in the detection area. The exponential form strengthens the response to direction perturbations "when approaching an even distribution" and is suitable for capturing slight anomalies.

[0026] From the texture direction dispersion index, it can be seen that for each detection area, the larger the performance value of the texture direction dispersion index generated by deeply analyzing the local texture direction change dispersion degree of the detection area, the greater the risk of slight pitting in the detection area. The reason is as follows: Due to good construction process control, the surface texture of normal concrete usually shows relatively orderly, single-direction or regular distribution, that is, the surface texture energy dominates in one or a few main directions, with concentrated directionality, and the texture direction dispersion index is naturally low; while in the slightly pitted area, due to insufficient mortar flow, improper vibration compaction or local exposure of stones during concrete pouring, the distribution of the surface texture shows the characteristics of no obvious dominant direction or dispersed direction energy, that is, the energy in multiple directions is relatively balanced, the directionality is chaotic, and the dispersion degree increases significantly. Therefore, after feature mapping, the texture direction dispersion index will increase significantly and become an important index for judging the risk of slight pitting. The higher the texture direction dispersion index, the more lacking the main directionality of the local surface of the area, and the closer it is to the pitting feature. On the contrary, the lower the texture direction dispersion index, the relatively regular the surface can be judged, and the lower the risk of slight pitting.

[0027] For each detection area, when the distribution density of micro-depressions in the detection area is large, it usually indicates that the risk of slight pitting in this area is greater. The fundamental reason is that slight pitting is usually caused by insufficient mortar flow, improper vibration compaction or local formwork leakage during concrete pouring. Although these defects will not form obvious deep holes or exposed steel bars, they will form a large number of small and dense depressions, honeycomb-like shallow holes or micro-depression textures on the concrete surface. Compared with the normal flat concrete surface layer, the surface details of the slightly pitted area show local unevenness and granularity, making the distribution density of the depression features significantly increase in high-resolution images or point clouds. Especially when these depressions show the characteristics of extensive distribution, relatively uniform size but dense quantity, they are often important signals of the early or hidden manifestation of slight pitting. Therefore, the larger the distribution density of micro-depressions, the higher the probability that the area has slight pitting can be effectively reflected statistically.

[0028] For each detection area, the specific steps for deeply analyzing the distribution density of micro-depressions in the detection area to generate the micro-depression density index are as follows: Extract the local micro-depression features on the surface by performing morphological bottom-hat transform on the high-resolution image in the detection area to obtain the set of micro-depression candidate areas . Subsequently, based on the spatial geometric constraints of the micro-depressions, filter out the effective micro-depression point set that meets the depression depth threshold and the minimum depression area threshold , , that is, satisfy: , where: Represents the maximum depth of the micro-dent points , Represents the area of the micro-dent points , and are the depression depth threshold and the minimum depression area threshold respectively, used to screen out effective micro-dent points; This step accurately extracts the set of effective depression feature points, recognizes the subtle undulations on the surface through morphological operators, avoids interference from illumination, texture artifacts, etc., and effectively extracts the real depression signals of slightly pitted surfaces.

[0029] Perform a spatial distribution analysis on the set of effective micro-dent points . By calculating the density and distribution compactness of the micro-dent point set, a micro-dent density index is comprehensively constructed. The expression for constructing the micro-dent density index is: , where: is the micro-dent density index, represents the number of the set of effective micro-dent points, A is the actual detection area of the detection region, is the average nearest neighbor distance of the depression point set, that is: , where: represents the distance from the micro-dent point to its nearest neighbor micro-dent point, is the compactness enhancement coefficient, used to balance the importance of density and distribution aggregation; The micro-dent density index generated in this step simultaneously considers the number of micro-dents (the denser the distribution, the more inclined to a pitted surface) and the distribution compactness of micro-dents (the more aggregated the micro-dent group, the more in line with the performance of a slightly pitted surface), directly quantifies the micro-dent distribution with spatial density-aggregation characteristics, and effectively improves the sensitivity and discrimination ability for slightly pitted surfaces.

[0030] As can be seen from the micro-dent density index, for each detection area, the larger the performance value of the micro-dent density index generated by in-depth analysis of the micro-dent distribution density in the detection area, the greater the risk of slight pitting on the surface of the detection area. The reason is that the formation mechanism of slight pitting mainly stems from problems such as insufficient slurry, inadequate vibration compaction, or air bubble accumulation during the concrete pouring process, resulting in a large number of shallow depressions with small sizes but dense distribution on the surface. When, through feature extraction and spatial distribution analysis, the number of effective micro-dents in the detection area is large (high density) and shows a certain degree of aggregation, the micro-dent density index increases significantly, indicating insufficient continuity and integrity of the surface structure in this area and the manifestation of slight pitting characteristics. If the micro-dent density index is small, it means that the number of micro-dents in this area is small, the distribution is sparse, the surface finish is good, and it conforms to the surface characteristics of normally poured and formed concrete. Therefore, the risk is relatively low. The micro-dent density index can effectively capture the typical distribution law of high density, weak depression, and spatial compactness unique to the pitted area and is an important quantitative index reflecting the risk of slight pitting.

[0031] Input the analyzed features into a pre-trained convolutional neural network, output the confidence score of the detection area, and based on the score result, identify whether there is slight pitting in the detection area and perform a suspected slight pitting determination. Input the feature vector composed of the texture direction discrete index and the micro-dent density index after analysis into a pre-trained convolutional neural network, output the confidence coefficient through the convolutional neural network, and based on the confidence coefficient, identify whether there is slight pitting in the detection area and perform a suspected slight pitting determination.

[0032] A pre-trained convolutional neural network refers to a deep neural network model that, before the UAV concrete honeycomb and pitted surface inspection system is officially put into inspection applications, has completed the structural design, parameter optimization, and model training of the neural network in advance using a large amount of concrete surface data highly relevant to the target defect detection task through deep learning methods, and has been verified to have good recognition capabilities. Here, "pre-training" does not mean that the model is completely separated from the actual inspection environment, but rather that the learning process of the model has been completed on a large number of labeled concrete surface image datasets, covering a rich distribution of surface defect features, including honeycomb and pitted surfaces, slightly pitted surfaces, ordinary pitted surfaces, defect-free concrete surfaces, and various variations under different lighting, viewing angles, and imaging conditions. During the training process, through multiple stages such as convolution, pooling, feature extraction, feature fusion, and feature classification of images, the model has learned the multi-level feature expression rules of different types of pitted surfaces and their potential distributions, and formed a mapping relationship between features and target categories, that is, network weights. After repeated training, verification, and optimization, the convolutional neural network model can make high-precision classifications and confidence predictions for the input feature vectors. Therefore, when the system extracts the feature vectors of the detection area during the actual inspection task (such as the feature vector composed of the discrete index of texture direction and the density index of micro-depressions), there is no need to train from scratch, but the pre-trained network can be directly called for inference, and quickly output the class confidence corresponding to the feature vector, that is, the possibility of the existence of slightly pitted surfaces in the detection area.

[0033] Technically speaking, a pre-trained convolutional neural network is equivalent to the core discriminator of the system and is the key tool for solving the problem of automatic recognition of honeycomb and slightly pitted surfaces. Compared with traditional image analysis methods that only rely on a small number of features designed manually, convolutional neural networks can effectively utilize the high-dimensional information hidden in a large amount of concrete surface data through end-to-end deep feature learning, and have stronger recognition and generalization capabilities. Moreover, for defects such as slightly pitted surfaces with weak features and hidden manifestations, convolutional neural networks can focus on multi-dimensional feature combinations such as depressions, textures, and uneven lighting through multi-level convolution extraction of feature maps, and establish a more complex and effective defect description ability than manually designed ones, thus significantly improving the recognizability of defects. Therefore, a pre-trained convolutional neural network not only has the advantages of fast response, stability, and reliability, but also can adapt to the diverse concrete surface manifestations under different construction environments, and is a key component for the inspection system to achieve intelligence.

[0034] The mentioned "pre-trained convolutional neural network" is not just an ordinary image classification model. It undertakes the crucial task of accurately distinguishing slightly pitted surfaces from ordinary pitted surfaces or normal concrete surfaces. Specifically, during the inspection process, based on the images collected by the drone, after the feature extraction stage, the system extracts two key features from the set of analyzed images in the detection area: the local texture directional change and the micro-depression distribution density. These two features are further quantified through in-depth analysis to form a texture direction discrete index and a micro-depression density index, which together constitute a set of discriminant feature vectors. Taking this feature vector as the input and directly feeding it into the pre-trained convolutional neural network, the model can, based on the distribution law of the massive sample features learned during training, perform feature map mapping, feature space projection, and class discrimination on the input features, and finally output a confidence coefficient, which reflects the possibility that the input area is determined to be a slightly pitted surface.

[0035] Compared with the traditional direct global discrimination method based on images, this solution, through the collaborative work of "feature extraction + in-depth feature analysis + pre-trained convolutional neural network classification", enables the convolutional neural network to focus on the feature subspace closely related to slightly pitted surfaces, avoiding the interference of a large amount of interfering information in the original images (such as light changes, surface color differences, casting marks, etc.) on the recognition results, significantly improving the recognition confidence, and reducing the probability of misjudgment and missed judgment. More importantly, the pre-trained convolutional neural network has good generalization ability and can stably output reasonable confidence coefficients under different inspection environments, different concrete materials, and different imaging conditions, providing a reliable basis for the inspection system to implement adaptive adjustment strategies (such as flight altitude, stagnation time, image re-acquisition). In this way, without significantly increasing the inspection time cost, the system can effectively improve the recognition accuracy of difficult-to-detect defects such as slightly pitted surfaces and potential deep honeycombing, and truly achieve the intelligent inspection ability of concrete appearance quality for engineering applications.

[0036] There is no specific limitation on the convolutional neural network model here. Any convolutional neural network model that can comprehensively analyze the texture direction discrete index and the micro-depression density index to generate a confidence coefficient is acceptable. To implement the technical solution of the present invention, the present invention provides a specific implementation method; the expression for generating the confidence coefficient is: , where , are the preset proportionality coefficients of the texture direction discrete index and the micro-depression density index respectively, and are both greater than 0.

[0037] The preset proportionality coefficients, namely those in the text and , refer to two constant coefficients used to weight the contributions of the texture direction discrete index and the micro-dent density index when calculating the confidence coefficient. The role of these two coefficients is to balance and regulate the weight ratio of the two features in calculating the total confidence, preventing the numerical characteristics (such as dimension, variation range, distribution) of a certain feature from dominating absolutely in the calculation, resulting in an imbalance of the confidence coefficient. Specifically and are proportionality factors artificially set or obtained through optimization based on a large number of experiments or data fitting during the model design or training stage. They can be flexibly adjusted according to different engineering objects, image features, or model characteristics to ensure that the confidence coefficient can reasonably and accurately reflect the possibility of the existence of slight pockmarked defects. For example, if it is found in a large number of training samples that the micro-dent density contributes more to defect recognition, the value of can be appropriately increased to increase the influence of the micro-dent density on the confidence, thereby improving the final recognition accuracy. Therefore, the essence of the preset proportionality coefficients is a weighted parameter for feature fusion, and its rationality is directly related to the reliability and determination accuracy of the system's recognition of honeycomb pockmarks.

[0038] It can be seen from the confidence coefficient that for each detection area, the larger the performance value of the texture direction discrete index generated by deeply analyzing the local texture directionality change dispersion of the detection area, and the larger the performance value of the micro-dent density index generated by deeply analyzing the micro-dent distribution density of the detection area, that is, the larger the performance value of the confidence coefficient generated when predicting the existence of slight pockmarks in the detection area through a pre-trained convolutional neural network, the greater the risk of slight pockmarks in the detection area. Conversely, it indicates that the risk of slight pockmarks in the detection area is smaller.

[0039] Compare and analyze the confidence coefficient generated when predicting the existence of slight pockmarks in the detection area through a pre-trained convolutional neural network with a preset confidence coefficient reference threshold to perform a suspected slight pockmark determination. The specific steps are as follows: If the confidence coefficient is greater than the confidence coefficient reference threshold, then divide the detection area into suspected slight pockmarks; If the confidence coefficient is less than or equal to the confidence coefficient reference threshold, then divide the detection area into a normal plane.

[0040] When it is recognized that there are suspected slight pockmarks in the detection area, determine the effective area of the slight pockmarks in the detection area; The specific steps are: After identifying a suspected slight pitting in the detection area, to further determine its specific effective area of slight pitting, the following steps can generally be completed: First, based on the output result of the convolutional neural network, locate the initial defect position points of the suspected slight pitting in the detection area; Secondly, around the initial position points, use information such as the high-frequency features, texture changes, gray-scale differences, and depression density of the image to perform local area expansion to form a candidate area for slight pitting; Then, adopt the adaptive region growing method or the edge detection + region segmentation algorithm to accurately segment the actual contour of the slight pitting and eliminate the background or irrelevant noise areas; Subsequently, perform connected component analysis on the segmented area to eliminate pseudo-defects with too small area or abnormal morphology, and only retain the connected areas that conform to the characteristics of slight pitting; Finally, calculate parameters such as the actual area, boundary, and position center coordinates of the defect area to form a complete description of the slight pitting range, providing reliable range information support for subsequent defect level evaluation and inspection strategy adjustment.

[0041] Based on the confidence score output by the convolutional neural network, adaptively lower the actual inspection height, improve the spatial resolution of the image, enhance the recognition ability of the subsequent AI detection model, and dynamically adjust the inspection time of the unmanned aerial vehicle according to the area size of the slight pitting in the detection area and the actual inspection height to ensure accurate recognition of deep honeycombs; When the convolutional neural network outputs a confidence coefficient for the detection area, first compare it with the pre-set confidence reference threshold. If the confidence coefficient meets the condition that it is greater than the confidence coefficient reference threshold, it is determined that there is a suspected slight pitting in this area. At this time, to further improve the imaging resolution of the suspected defect area, the following adaptive height lowering strategy is used to adaptively lower the height of the unmanned aerial vehicle inspection: , where: is the adjusted actual inspection height, is the pre-configured optimal inspection height (such as 15m), is the inspection height lowering coefficient (range 0-1) used to control the lowering amplitude, is the currently recognized confidence coefficient, is the pre-set confidence reference threshold; This step uses confidence-driven relative ratio lowering. Instead of directly reducing the height rigidly, the lowering amplitude changes non-linearly with the proximity of the confidence. When the confidence is closer to 1, the height lowering is more significant. This can adaptively improve the image resolution, obtain higher-quality images, provide rich and clear surface information for subsequent convolutional neural network rejudgment and deep honeycomb detection, and prevent ordinary pitting or false features from being misjudged as honeycombs.

[0042] After the height is adjusted downward and a high-resolution image is obtained, based on the effective area of the slight pitting in the determined detection area, subsequently, in combination with the current actual inspection height and the effective area of the slight pitting, the inspection stay time of the UAV in this detection area is dynamically determined. The following area-height joint regulation model is adopted, and the specific expression is: , where: is the actual inspection stay time after dynamic adjustment, is the minimum stay time of the UAV under normal circumstances, is the detected effective area of the slight pitting, is the standardized defect area (which can be set according to engineering experience, for example, 50 cm²), is the area ratio adjustment coefficient, which controls the influence of the area on the time, is the height reduction ratio adjustment coefficient, which controls the influence of the reduction degree on the time; This step is jointly driven by the area and the height to dynamically adjust the stay time of the UAV, ensuring that: the larger the area, the longer the stay time, and ensuring that enough defect detail images are obtained; the more the height is reduced, it means that the risk of the current area is higher (the confidence level is higher), and the same requires a longer time to supplement the image acquisition.

[0043] Through the above scheme, the problems of "difficult to accurately identify slight pitting" and "easy to miss deep honeycombing" existing in the existing UAV inspection of concrete appearance can be effectively solved, and the height adaptability and defect recognition accuracy of the inspection process are improved. Through the dynamic inspection mechanism driven by the confidence score, the UAV can actively and adaptively adjust the inspection height and inspection time when detecting suspected slight pitting, thereby significantly improving the image resolution and the expression ability of defect feature details. At the same time, combined with the feature depth analysis and the input of high-identifiability feature vectors, the recognition ability of the convolutional neural network for slight pitting and dark honeycombing is enhanced. Finally, the defect miss rate and false alarm rate in the inspection process can be effectively reduced, and the accurate, stable and efficient recognition of hidden defects such as slight pitting and deep honeycombing on the surface of the concrete structure can be realized, and the reliability and practicability of the intelligent inspection of the concrete appearance quality are significantly improved.

[0044] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. The preset parameters in the formulas are set by technicians in this field according to the actual situation.

[0045] As described above, this is only the specific implementation manner of the present application. However, the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described above.

[0046] Only some exemplary embodiments of the present invention have been described above by way of illustration. Without doubt, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

Claims

1. An intelligent inspection method for the appearance quality of concrete, characterized in that, The steps include: Based on the detection accuracy requirement and detection efficiency requirement of honeycomb pitted surface, configure the optimal inspection height for UAV intelligent inspection; According to the preset flight path and the set optimal inspection height, the UAV conducts inspection on the surface of the concrete structure, and uses the onboard high-resolution imaging system to continuously capture the inspection area to form an inspection image data stream; During the inspection process, for each detection area, the image data collected during its inspection process is collected in real time, a analysis set for the detection area is constructed, the key features that can represent slight pitting are extracted from the analysis set, and the extracted key features are deeply analyzed to provide highly recognizable input features for the convolutional neural network and improve the recognition confidence of slight pitting; Input the analyzed features into the pre-trained convolutional neural network, output the confidence score of the detection area, and based on the score result, determine the suspected slight pitting within the detection area; When it is recognized that there is suspected slight pitting in the detection area, determine the effective area of slight pitting within the detection area; Based on the confidence score output by the convolutional neural network, adaptively lower the actual inspection height, and dynamically adjust the inspection time of the UAV according to the area size of slight pitting within the detection area and the actual inspection height to accurately identify deep honeycombing; 2. The intelligent inspection method for the appearance quality of concrete according to claim 1, characterized in that, For the dual requirements of honeycomb pitted surface detection accuracy and detection efficiency, configure the optimal inspection height for UAV intelligent inspection. The specific steps include: Combined with the actual engineering requirements, clarify the minimum detection size of honeycomb pitted surface defects and the required image resolution; Calculate the flight height range that meets the resolution requirement based on the parameters of the imaging equipment carried by the UAV; Comprehensively consider the inspection operation efficiency, the spatial characteristics of the detection area and flight safety, and initially determine the optimal inspection height; Verify the clarity and feature performance ability of the image at the selected height through simulation or on-site test flight to ensure that at this height, it can not only ensure the effective recognition of ordinary honeycomb pitted surface, but also complete efficient inspection at the flight speed; Combined with the feature distribution of the AI recognition model training data, fine-tune and optimize the selected inspection height, and finally determine the optimal inspection operation height applicable to honeycomb pitted surface detection; 3. The intelligent inspection method for the appearance quality of concrete according to claim 1, characterized in that For each detection area, extract the key features that represent slight pitting from the analysis set formed by the image data. Among them, the extracted features include the local texture directional change dispersion and the micro-depression distribution density. After deeply analyzing the extracted key features, generate the texture direction dispersion index and the micro-depression density index respectively. Use the texture direction dispersion index and the micro-depression density index as quantitative indicators to reflect the presence of slight pitting within the detection area, provide highly recognizable input for the convolutional neural network, and improve the recognition confidence of slight pitting; 4. The intelligent inspection method for the appearance quality of concrete according to claim 3, characterized in that, Input the feature vector composed of the texture direction dispersion index and the micro-depression density index after analysis into the pre-trained convolutional neural network. Through the convolutional neural network, output the confidence coefficient, and based on the confidence coefficient, identify whether there is slight pitting in the detection area and perform the determination of suspected slight pitting.

5. The intelligent inspection method for the appearance quality of concrete according to claim 4, characterized in that, When predicting the presence of slight pitting in the detection area using a pre-trained convolutional neural network, compare the confidence coefficient generated with a pre-set confidence coefficient reference threshold, and perform a determination of suspected slight pitting. The specific steps are as follows: If the confidence coefficient is greater than the confidence coefficient reference threshold, then classify this detection area as a suspected slight pitting area; If the confidence coefficient is less than or equal to the confidence coefficient reference threshold, then classify this detection area as a normal plane.

6. The intelligent inspection method for the appearance quality of concrete according to claim 5, wherein, When it is recognized that there is a suspected slight pitting in the detection area, determine the effective area of the slight pitting in the detection area. The specific steps are as follows: After recognizing that there is a suspected slight pitting in the detection area, to further determine its specific effective area of slight pitting, complete it according to the following steps: Based on the output result of the convolutional neural network, locate the initial defect position points of the suspected slight pitting in the detection area; Around the initial position points, perform local area expansion on the image to form a candidate area for slight pitting; Adopt an adaptive region growing method or an edge detection + region segmentation algorithm to accurately segment the actual contour of the slight pitting; Perform connected component analysis on the segmented area, eliminate pseudo-defects with too small area or abnormal morphology, and only retain the connected areas that conform to the characteristics of slight pitting; Calculate the defect area parameters to form a complete description of the slight pitting range.

7. An intelligent inspection method for the appearance quality of concrete according to claim 6, characterized in that After the convolutional neural network outputs a confidence coefficient for the detection area, first compare it with the pre-set confidence reference threshold. If the condition that the confidence coefficient is greater than the confidence coefficient reference threshold is met, then determine that there is a suspected slight pitting in this area. At this time, to further improve the imaging resolution of the suspected defect area, adopt the following adaptive height reduction strategy to adaptively reduce the height of the drone inspection: , where: is the adjusted actual inspection height, is the pre-configured optimal inspection height, is the inspection height reduction coefficient for controlling the reduction amplitude, is the currently recognized confidence coefficient, is the preset confidence reference threshold.

8. An intelligent inspection method for the appearance quality of concrete according to claim 7, characterized in that, After completing the height reduction and obtaining a high-resolution image, according to the determined effective area of the slight pitting in the detection area, then, combining the current actual inspection height and the effective area of the slight pitting, dynamically determine the inspection stay time of the drone in this detection area, and adopt the following area-height joint control model. The specific expression is: , where: is the actual inspection stay time after dynamic adjustment, is the minimum stay time of the UAV under normal circumstances, is the effective area of detected slight pitting, is the standardized defect area, is the area ratio adjustment coefficient, controlling the influence of area on time, is the height reduction ratio adjustment coefficient, controlling the influence of the reduction degree on time.

Images