An intelligent inspection method for concrete appearance quality

Abstract

The present invention discloses an intelligent inspection method for the appearance quality of concrete, which relates to the technical field of concrete appearance quality detection. The method comprises the following steps: configuring an optimal inspection height for an unmanned aerial vehicle (UAV) intelligent inspection based on the detection accuracy and efficiency requirements of honeycomb pitting; inspecting the surface of a concrete structure using a pre-set route and the set optimal inspection height, and continuously photographing the inspection area using an onboard high-resolution imaging system. The present invention utilizes a dynamic inspection mechanism driven by confidence scores, enabling the UAV to proactively and adaptively adjust its inspection height and time when suspected minor pitting is detected. This improves image resolution and the ability to express defect feature details, enhances the convolutional neural network's ability to identify minor pitting and dark honeycombs, effectively reduces the defect missed detection rate and false detection rate during the inspection process, and achieves accurate, stable, and efficient identification of hidden defects such as minor pitting and deep honeycombs on the surface of concrete structures.

Description

Technical Field

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

[0002] Intelligent inspection of concrete appearance quality utilizes artificial intelligence (AI), computer vision, image recognition, and automation technologies to automatically and intelligently inspect and evaluate the surface quality of concrete structures (such as walls, beams, columns, and slabs). By deploying high-definition cameras, drones, robots, or handheld devices to capture concrete surface images, AI algorithms are used to automatically identify, classify, locate, and quantify common defects such as cracks, honeycombing, exposed rebar, hollows, and holes. This replaces traditional inspection methods that rely on visual inspection and manual record-keeping. This technology improves inspection accuracy, consistency, and efficiency, reduces human error, and enables timely detection and early warning of quality risks, providing scientific and reliable support for project quality acceptance, maintenance decisions, and construction process control.

[0003] The existing technology has the following deficiencies:

[0004] Existing drone-based intelligent inspection technology for concrete appearance quality typically uses a constant inspection height to detect defects on the surface of concrete structures, achieving a compromise between detection efficiency and recognition accuracy. However, due to the diverse manifestations of honeycomb pockmarks on the concrete surface, especially some deep honeycombs (also known as dark honeycombs), their appearance often only appears as slight pockmarks, making them difficult to accurately identify at conventional inspection heights. Drones in existing technologies lack an adaptive height adjustment mechanism based on defect recognition results. When the AI ​​detection model initially identifies slight pockmarks, it still uses the original constant inspection height to complete the operation, resulting in insufficient image resolution and observation angles, making it impossible to effectively distinguish between slight pockmarks and deep honeycombs. This makes it very easy to miss or misdetect, seriously affecting the reliability of inspection results and the ability to detect structural safety risks early.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose 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 scoring, the drone can actively and adaptively adjust the inspection height and inspection time when suspected slight pitting is found, 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-recognition feature vector input, the convolutional neural network's recognition ability for slight pitting and dark honeycombs is enhanced. Ultimately, it can effectively reduce the defect missed detection rate and false detection rate during the inspection process, and achieve accurate, stable and efficient identification of hidden defects such as slight pitting and deep honeycombs on the surface of concrete structures, significantly improving the reliability and practicality of intelligent inspection of concrete appearance quality, so as to solve the problems in the above-mentioned background technology.

[0007] In order 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:

[0008] Based on the requirements for detection accuracy and efficiency of honeycomb surface, the optimal inspection height is configured for drone intelligent inspection;

[0009] The drone inspects the surface of the concrete structure according to the preset route and the set optimal inspection height. The onboard high-resolution imaging system continuously captures the inspection area to form an inspection image data stream.

[0010] During the inspection process, for each inspection area, the image data collected during the inspection process is aggregated in real time to construct an analysis set for the inspection area. Key features that can characterize slight pockmarks 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, thereby improving the recognition confidence of slight pockmarks.

[0011] The analyzed features are input into a pre-trained convolutional neural network, which outputs a confidence score for the detection area. Based on the score, the system identifies whether there is slight pockmarking in the detection area and performs a suspected slight pockmarking judgment.

[0012] When suspected slight pitting is identified in the detection area, an effective area of ​​the slight pitting in the detection area is determined;

[0013] Based on the confidence score output by the convolutional neural network, the actual inspection height is adaptively lowered to improve the spatial resolution of the image and enhance the recognition ability of the subsequent AI detection model. The inspection time of the drone is dynamically adjusted according to the area size of the slight pockmarks in the detection area and the actual inspection height to ensure accurate identification of deep honeycombs.

[0014] Preferably, in order to meet the dual requirements of honeycomb surface detection accuracy and detection efficiency, the optimal inspection height is configured for the drone intelligent inspection. The specific steps include:

[0015] Based on the actual project needs, the minimum detection size and required image resolution of honeycomb surface defects should be clarified;

[0016] Secondly, the flight altitude range that can meet the resolution requirement is calculated based on the parameters of the imaging equipment carried by the UAV;

[0017] Then, the optimal inspection altitude is preliminarily determined by comprehensively considering inspection efficiency, spatial characteristics of the inspection area, and flight safety.

[0018] Next, verify the image clarity and feature representation capabilities at the selected altitude through simulation or on-site test flights to ensure that both effective recognition of common honeycomb surfaces and efficient inspections can be completed at this altitude at a reasonable flight speed.

[0019] Finally, combined with the characteristic distribution of the AI ​​recognition model training data, necessary fine-tuning and optimization are performed on the selected inspection height, and the optimal inspection operation height suitable for honeycomb surface detection is finally determined, providing reasonable initial parameters for subsequent automatic inspection tasks.

[0020] Preferably, for each detection area, key features characterizing slight pockmarks are extracted from the analysis set formed by the image data, wherein the extracted features include the discreteness of local texture directional changes and the distribution density of micro-depressions. After in-depth analysis of the extracted key features, texture direction discrete index and micro-depression density index are generated respectively. The texture direction discrete index and micro-depression density index are used as quantitative indicators reflecting the presence of slight pockmarks in the detection area, providing highly recognizable input for the convolutional neural network and improving the recognition confidence of slight pockmarks.

[0021] Preferably, the analyzed feature vector composed of texture direction discrete indicators and micro-depression density indicators is input into a pre-trained convolutional neural network, and the confidence coefficient is output by the convolutional neural network. Based on the confidence coefficient, it is identified whether there are slight pockmarks in the detection area, and a suspected slight pockmark judgment is performed.

[0022] Preferably, the confidence coefficient generated when the pre-trained convolutional neural network predicts the presence of slight pockmarks in the detection area is compared with a pre-set confidence coefficient reference threshold to perform suspected slight pockmark determination. The specific steps are as follows:

[0023] If the confidence coefficient is greater than the confidence coefficient reference threshold, the detection area is classified as suspected slight pockmarks;

[0024] If the confidence coefficient is less than or equal to the confidence coefficient reference threshold, the detection area is divided into a normal plane.

[0025] Preferably, when it is identified that there is suspected slight pockmarking in the detection area, the effective area of ​​the slight pockmarking in the detection area is determined, and the specific steps are:

[0026] After identifying suspected minor pitting in the inspection area, further determining the specific effective area of ​​the minor pitting can usually be accomplished by following the following steps:

[0027] First, based on the output of the convolutional neural network, the initial defect location of suspected slight pitting in the inspection area is located;

[0028] Secondly, around the initial position point, the local area is expanded using the image's high-frequency features, texture changes, grayscale differences, concave density and other information to form a candidate area for slightly pockmarked surfaces;

[0029] Then, the adaptive region growing method or edge detection + region segmentation algorithm is used to accurately segment the actual contour of the slightly pockmarked surface and eliminate the background or irrelevant noise areas;

[0030] Then, the segmented areas are analyzed for connected domains to eliminate pseudo-defects that are too small or have abnormal shapes, and only connected areas that meet the characteristics of slight pockmarks are retained.

[0031] Finally, the actual area, boundary, center coordinates and other parameters of the defect area are calculated to form a complete description of the range of the slight pitting surface, providing reliable range information support for subsequent defect level assessment and inspection strategy adjustment.

[0032] Preferably, after the convolutional neural network outputs the confidence coefficient for the detection area, it is first compared with a preset confidence reference threshold. If the confidence coefficient is greater than the confidence coefficient reference threshold, it is determined that there is suspected slight pitting in the area. At this time, in order to further improve the imaging resolution of the suspected defect area, the following adaptive height reduction strategy is used to adaptively reduce the height of the drone inspection:

[0033] ,in: is the actual inspection height after adjustment, For the pre-configured optimal inspection height (e.g. 15m), It is the inspection height reduction coefficient (range 0~1), used to control the reduction range. is the confidence coefficient of the current recognition, is the preset confidence reference threshold.

[0034] Preferably, after completing the altitude reduction and obtaining a high-resolution image, the effective area of ​​the slightly pitted surface within the determined detection area is then combined with the current actual inspection altitude and the effective area of ​​the slightly pitted surface to dynamically determine the inspection dwell time of the drone in the detection area. The following area-altitude joint control model is used, and the specific expression is:

[0035] ,in: 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 ​​the detected slight pockmarks, is the standardized defect area (which can be set according to engineering experience, for example, 50cm²), is the area ratio adjustment coefficient, which controls the impact of area on time. It is the adjustment coefficient of height reduction ratio, which controls the impact of reduction degree on time.

[0036] Preferably, for each detection area, the specific steps of performing an in-depth analysis on the local texture directional variation discreteness of the detection area to generate a texture direction discrete index are as follows:

[0037] First, the detection area image is divided into multiple fixed-size image blocks (such as 32×32 pixels), and multi-directional gradient response calculations are performed on each image block. Common directions include 0°, 45°, 90°, and 135°. Using a directional gradient filter (such as Sobel or Gabor transform), the directional energy response value of the image block is calculated in each direction to form the directional energy vector of the image block. E , ,in Indicates the direction Then, the directional energy vectors of all image blocks are aggregated in the entire detection area to construct a global directional energy distribution model;

[0038] The energy in each direction Normalization is performed to obtain the directional probability distribution weight. The texture directional discrete index is calculated based on the directional probability distribution weight as a quantitative indicator reflecting the presence of slight pockmarks in the detection area. The calculation expression of the texture directional discrete index is:

[0039] ,in: is the direction probability distribution weight, and the calculation expression is: , is a nonlinear adjustment factor that adjusts the sensitivity to the direction of small probability. It is a texture direction discrete index that reflects the degree of texture directional disorder in the detection area. The exponential form strengthens the response to directional disturbance when the distribution is close to equilibrium, which is suitable for capturing slight anomalies.

[0040] Preferably, for each detection area, the specific steps of performing an in-depth analysis on the micro-depression distribution density of the detection area to generate a micro-depression density index are as follows:

[0041] The micro-depression features of the local surface are extracted by performing morphological bottom-hat transform on the high-resolution image in the detection area, and a set of micro-depression candidate areas is obtained. Then, based on the spatial geometric constraints of micro-depression, the effective micro-depression point set that meets the depression depth threshold and the minimum depression area threshold is screened out. , , that is, satisfying:

[0042] ,in: Indicates micro-depression points The maximum depth of Indicates micro-depression points The area, and They are respectively the depression depth threshold and the minimum depression area threshold, which are used to screen out the effective micro-depression point set;

[0043] For the effective micro-depression point set Spatial distribution analysis is performed, and the density and distribution compactness of the micro-depression point set are calculated to comprehensively construct the micro-depression density index. The expression of the constructed micro-depression density index is:

[0044] ,in: is the micro-depression density index, is the number of valid micro-depression point sets, A is the actual detection area of ​​the detection area, is the average nearest neighbor distance of the concave point set, that is:

[0045] ,in: Indicates micro-depression points The distance to its nearest neighbor micro-depression point, is the compactness enhancement coefficient, which is used to balance the importance of density and distribution aggregation.

[0046] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0047] This invention uses a dynamic inspection mechanism driven by confidence scores, enabling drones to proactively and adaptively adjust their inspection height and time when suspected minor pitting is detected, significantly improving image resolution and the ability to express detail in defect features. Simultaneously, the combination of feature depth analysis and high-recognition feature vector input enhances the convolutional neural network's ability to identify minor pitting and dark honeycombs. Ultimately, it can effectively reduce the defect omission and false detection rates during inspections, enabling accurate, stable, and efficient identification of hidden defects such as minor pitting and deep honeycombs on the surface of concrete structures, significantly improving the reliability and practicality of intelligent inspections of concrete appearance quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0049] Figure 1 The present invention is a flowchart of an intelligent inspection method for concrete appearance quality. DETAILED DESCRIPTION

[0050] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0051] The present invention provides Figure 1 The intelligent inspection method for the appearance quality of concrete shown includes the following steps:

[0052] Based on the requirements for detection accuracy and efficiency of honeycomb surface, the optimal inspection height is configured for drone intelligent inspection;

[0053] In response to the dual requirements of honeycomb surface detection accuracy and detection efficiency, the specific steps for configuring the optimal inspection height for drone intelligent inspection include: first, based on the actual project needs, clarify the minimum detection size of honeycomb surface defects and the required image resolution (for example, requiring the recognition of surface features ≥5mm); second, based on the imaging equipment parameters carried by the drone (such as camera resolution, focal length, and field of view), calculate the flight altitude range that can meet the resolution requirements; then, comprehensively consider the inspection efficiency, spatial characteristics of the inspection area (such as building height and shape complexity) and flight safety, preliminarily determine the optimal inspection height; then, through simulation or on-site The test flight verifies the image clarity and feature expression capabilities at the selected altitude, ensuring that at this altitude, both common honeycomb surfaces can be effectively identified and efficient inspections can be completed at a reasonable flight speed. (For example, simulation analysis can be used to determine that the image resolution of a concrete surface at an altitude of 15m can effectively resolve features >5mm, i.e., 15m is set as the initial optimal inspection altitude for this scenario.) Finally, based on the feature distribution of the AI ​​recognition model training data, the selected inspection altitude is fine-tuned and optimized as necessary, ultimately determining the optimal inspection operating altitude suitable for honeycomb surface detection, providing reasonable initial parameters for subsequent automatic inspection tasks.

[0054] The drone inspects the surface of the concrete structure according to a preset route and the set optimal inspection height. The onboard high-resolution imaging system (such as an RGB camera or a multispectral camera) continuously captures the inspection area to form an inspection image data stream.

[0055] During inspections of concrete surfaces, to ensure the comprehensiveness and high quality of inspection image data, the drone's real-time trajectory control dynamically adjusts its flight attitude and shooting rhythm to ensure that image overlap and flight strip coverage meet predetermined requirements. Specifically, overlap refers to the spatial overlap between adjacent images, typically requiring a range of 70% to 85% to ensure the stability of image stitching, feature matching, and 3D modeling. Flight strip coverage refers to the ability of flight routes to fully cover the inspection area of ​​the concrete surface, avoiding blind spots or inspection blind spots. By monitoring the drone's position, attitude, and environmental conditions in real time, combined with the flight control system's path optimization capabilities, the flight track and shooting points can be dynamically fine-tuned based on the concrete surface's geometric features, local complexity, or the building's height. This ensures comprehensive, high-definition, and continuous image acquisition of the concrete structure's surface, providing high-quality, complete raw data for subsequent defect detection, feature extraction, and AI detection models.

[0056] During the inspection process, for each inspection area, the image data collected during the inspection process is aggregated in real time to construct an analysis set for the inspection area. Key features that can characterize slight pockmarks 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, thereby improving the recognition confidence of slight pockmarks.

[0057] For each detection area, key features characterizing slight pockmarks are extracted from the analysis set formed by the image data. The extracted features include the discrete degree of local texture directional changes and the distribution density of micro-depressions. After in-depth analysis of the extracted key features, texture directional discrete index and micro-depression density index are generated respectively. The texture directional discrete index and micro-depression density index are used as quantitative indicators to reflect the presence of slight pockmarks in the detection area, providing highly recognizable input for the convolutional neural network and improving the recognition confidence of slight pockmarks.

[0058] For each inspection area, if the dispersion of the local texture directional changes in the inspection area is large, it generally means that the risk of slight pitting in the inspection area is greater. The reason is that the mortar distribution and surface forming process of a normally cast concrete surface will make the surface texture directionality relatively uniform, and the texture direction will be continuous and smooth. However, slight pitting is usually caused by insufficient mortar flow, partial exposure of gravel, or inadequate vibration, resulting in irregular bumps, exposed particles, or tiny holes on the surface. These defects will destroy the original texture continuity and make the local texture direction appear diversified, disordered, or divergent, thereby significantly increasing the texture directional dispersion (i.e., the statistical distribution difference of texture in different directions). Therefore, when a large texture directional dispersion appears in the inspection area, it usually reflects the risk of slight pitting on the surface of this area caused by construction defects.

[0059] For each detection area, the specific steps for deeply analyzing the local texture directional change discreteness of the detection area to generate the texture direction discrete index are as follows:

[0060] First, the detection area image is divided into multiple fixed-size image blocks (such as 32×32 pixels), and multi-directional gradient response calculations are performed on each image block. Common directions include 0°, 45°, 90°, and 135°. Using a directional gradient filter (such as Sobel or Gabor transform), the directional energy response value of the image block is calculated in each direction to form the directional energy vector of the image block. E , ,in Indicates the direction Then, the directional energy vectors of all image blocks are aggregated in the entire detection area to construct a global directional energy distribution model;

[0061] The flatter the distribution is, the smaller the difference in texture energy in different directions is, the more chaotic the directionality is, and the greater the possibility of slight pockmarks.

[0062] The energy in each direction Normalization is performed to obtain the directional probability distribution weight. The texture directional discrete index is calculated based on the directional probability distribution weight as a quantitative indicator reflecting the presence of slight pockmarks in the detection area. The calculation expression of the texture directional discrete index is:

[0063] ,in: is the direction probability distribution weight, and the calculation expression is: , is a nonlinear adjustment factor that adjusts the sensitivity to the direction of small probability. It is a texture direction discrete index that reflects the degree of texture directional disorder in the detection area. The exponential form strengthens the response to directional disturbance when the distribution is close to equilibrium, which is suitable for capturing slight anomalies.

[0064] The texture direction dispersion index (TDDI) shows that for each inspection area, the greater the value of the TDDI, generated by in-depth analysis of the local texture directional variation dispersion, the greater the risk of minor pitting in that area. This is because, due to well-controlled construction processes, normal concrete surfaces typically exhibit a relatively orderly, single-directional, or regular distribution of surface texture. This means that surface texture energy is dominant in one or a few principal directions, resulting in a concentrated directionality, and the TDDI is naturally low. However, in slightly pitted areas, due to insufficient mortar flow, incomplete vibration, or exposed gravel, the surface texture distribution exhibits a lack of a clear dominant direction or a dispersion of directional energy. This means that energy is relatively balanced across multiple directions, directionality is chaotic, and the dispersion is significantly increased. Therefore, after feature mapping, the TDDI significantly increases, becoming an important indicator for determining the risk of minor pitting. A higher TDDI indicates a less dominant directionality in the local surface and a closer approximation to pitting characteristics. Conversely, a lower TDDI indicates a relatively regular surface and a lower risk of minor pitting.

[0065] For each inspection area, a higher density of micro-depressions generally indicates a greater risk of minor pitting in that area. The fundamental reason is that minor pitting is often caused by insufficient mortar flow, loose vibration, or localized mortar leakage in the formwork during concrete pouring. While these defects don't create obvious deep holes or exposed rebar, they do form a large number of tiny, dense depressions, honeycomb-like shallow holes, or micro-dimpled textures on the concrete surface. Compared to a normal, smooth concrete surface, the surface details of the slightly pitted area exhibit localized unevenness and a grainy feel, significantly increasing the distribution density of these depression features in high-resolution images or point clouds. In particular, when these depressions are widely distributed, relatively uniform in size, but densely populated, they are often important signals of early or hidden manifestations of minor pitting. Therefore, a higher density of micro-depressions statistically indicates a higher probability of the presence of minor pitting in that area.

[0066] For each detection area, the specific steps for performing an in-depth analysis of the micro-depression distribution density in the detection area to generate a micro-depression density index are as follows:

[0067] The micro-depression features of the local surface are extracted by performing morphological bottom-hat transform on the high-resolution image in the detection area, and a set of micro-depression candidate areas is obtained. Then, based on the spatial geometric constraints of micro-depression, the effective micro-depression point set that meets the depression depth threshold and the minimum depression area threshold is screened out. , , that is, satisfying:

[0068] ,in: Indicates micro-depression points The maximum depth of Indicates micro-depression points The area, and They are respectively the depression depth threshold and the minimum depression area threshold, which are used to screen out effective micro-depression points;

[0069] This step accurately extracts a set of effective concave feature points, identifies subtle surface undulations through morphological operators, avoids interference from lighting and texture artifacts, and effectively extracts true concave signals of slightly pockmarked surfaces.

[0070] For the effective micro-depression point set Spatial distribution analysis is performed, and the density and distribution compactness of the micro-depression point set are calculated to comprehensively construct the micro-depression density index. The expression of the constructed micro-depression density index is:

[0071] ,in: is the micro-depression density index, is the number of valid micro-depression point sets, A is the actual detection area of ​​the detection area, is the average nearest neighbor distance of the concave point set, that is:

[0072] ,in: Indicates micro-depression points The distance to its nearest neighbor micro-depression point, is the compactness enhancement coefficient, which is used to balance the importance of density and distribution aggregation;

[0073] The micro-depression density index generated in this step takes into account both the number of micro-depressions (the denser the distribution, the more inclined to pockmarks) and the compactness of the micro-depression distribution (the more clustered the micro-depression group, the more consistent with the appearance of mild pockmarks). It directly quantifies the micro-depression distribution using the spatial density-clustering characteristics, effectively improving the sensitivity and discrimination ability for mild pockmarks.

[0074] The micro-depression density index, generated through an in-depth analysis of the micro-depression distribution density within each test area, indicates that a higher value indicates a greater risk of minor pitting in that area. This is because minor pitting primarily arises from insufficient slurry, inadequate vibration, or bubble accumulation during concrete pouring, leading to the formation of a large number of small, densely distributed shallow depressions. When feature extraction and spatial distribution analysis reveal a high number (high density) of effective micro-depressions within the test area and a certain degree of clustering, the micro-depression density index increases significantly, reflecting a lack of continuity and integrity in the surface structure and the emergence of minor pitting. On the other hand, a lower micro-depression density index indicates a relatively low number of micro-depressions, a sparse distribution, and a good surface finish, consistent with the surface characteristics of properly poured concrete, thus indicating a relatively low risk. The micro-depression density index effectively captures the typical distribution pattern of high density, weak depressions, and compact spatial distribution characteristic of pitting areas, making it an important quantitative indicator of minor pitting risk.

[0075] The analyzed features are input into a pre-trained convolutional neural network, which outputs a confidence score for the detection area. Based on the score, the system identifies whether there is slight pockmarking in the detection area and performs a suspected slight pockmarking judgment.

[0076] The analyzed feature vector composed of texture direction discrete indicators and micro-depression density indicators is input into a pre-trained convolutional neural network. The convolutional neural network outputs a confidence coefficient, and based on the confidence coefficient, it is identified whether there is slight pockmarking in the detection area, and a suspected slight pockmarking judgment is performed.

[0077] A pre-trained convolutional neural network (CNN) refers to a deep learning-based deep neural network model that has been pre-trained using a large amount of concrete surface data highly relevant to the target defect detection task before the UAV-based concrete honeycomb inspection system is officially deployed. This deep neural network model has been validated to achieve excellent recognition capabilities through pre-training, utilizing a large amount of concrete surface data highly relevant to the target defect detection task. "Pre-trained" here does not mean that the model is completely independent of the actual inspection environment. Rather, it means that the model's learning process has been completed on a large dataset of labeled concrete surface images, covering a rich distribution of surface defect characteristics, including honeycomb, mild, normal, and defect-free concrete surfaces, as well as variations under various lighting, viewing angles, and imaging conditions. During training, through multiple stages of image convolution, pooling, feature extraction, feature fusion, and classification, the model learns the multi-level feature representation patterns of different types of pitting and their underlying distributions, and forms a mapping relationship between features and target categories, namely, network weights. Through repeated training, validation, and fine-tuning, the CNN model is able to make highly accurate classifications and confidence predictions for input feature vectors. Therefore, when the system extracts the feature vector of the detection area during an actual inspection task (for example, a feature vector composed of texture direction discrete indicators and micro-depression density indicators), there is no need to train from scratch. Instead, the pre-trained network can be directly called for inference to quickly output the category confidence corresponding to the feature vector, that is, the possibility of whether there is slight pockmarking in the detection area.

[0078] From a technical perspective, the pre-trained convolutional neural network is equivalent to the core discriminator of the system and is a key tool for solving the problem of automatic identification of honeycombed and slight pitting. Compared with traditional image analysis methods that rely only on a small number of manually designed features, convolutional neural networks can effectively utilize the high-dimensional information hidden in massive amounts of concrete surface data through end-to-end deep feature learning, and have stronger recognition and generalization capabilities. Moreover, for defects such as slight pitting, which have weak features and hidden manifestations, convolutional neural networks can use multi-level convolution extraction of feature maps to focus on multi-dimensional feature combinations such as depressions, textures, and uneven lighting, establishing a more complex and effective defect description capability than manual design, thereby significantly improving the identifiability of defects. Therefore, pre-trained convolutional neural networks not only have the advantages of fast response, stability and reliability, but can also adapt to the diverse concrete surface manifestations under different construction environments. They are a key component for the intelligentization of inspection systems.

[0079] The "pre-trained convolutional neural network" mentioned is not just an ordinary image classification model, it undertakes the key task of accurately distinguishing slight pitting from ordinary pitting 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 image analysis set of the detection area: local texture directional changes and micro-depression distribution density. These two types of features are further quantified through deep analysis to form texture direction discrete indicators and micro-depression density indicators, which together constitute a set of feature vectors for discrimination. The feature vector is directly sent as input to the pre-trained convolutional neural network. Based on the distribution rules of massive sample features learned during training, the model can perform feature map mapping, feature space projection and category discrimination on the input features, and finally output a confidence coefficient. The confidence coefficient reflects the possibility of the input area being judged as a slight pitting surface.

[0080] Compared to traditional direct image-based global discrimination methods, this solution, through the collaborative work of "feature extraction + feature depth analysis + pre-trained convolutional neural network classification," enables the convolutional neural network to focus on the feature subspace closely related to minor pitting. This avoids the interference of a large amount of noise in the original image (such as illumination changes, surface color differences, and pouring marks), significantly improving recognition confidence and reducing the probability of false positives and missed detections. Importantly, the pre-trained convolutional neural network exhibits excellent generalization capabilities and can stably output reasonable confidence coefficients across various inspection environments, concrete materials, and imaging conditions. This provides a reliable basis for the inspection system to implement adaptive adjustment strategies (such as flight altitude, dwell time, and image re-acquisition). In this way, the system can effectively improve the recognition accuracy of difficult-to-detect defects such as minor pitting and potential deep honeycombing without significantly increasing inspection time and costs, truly realizing intelligent inspection capabilities for concrete appearance quality for engineering applications.

[0081] The convolutional neural network model is not specifically limited here, and can realize the discrete index of texture direction and micro-depression density index Conduct comprehensive analysis to generate confidence coefficients In order to realize the technical solution of the present invention, the present invention provides a specific implementation method; the confidence coefficient The generated expression is: , where 、 Texture direction discrete index and micro-depression density index The preset scaling factor of 、 Both are greater than 0.

[0082] The preset proportional coefficient is 、 , refer to the two constant coefficients used to weight the contributions of the texture direction discrete index and the micro-depression 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 when calculating the total confidence, to prevent the numerical characteristics of a certain feature (such as dimension, range of variation, distribution) from dominating the calculation and causing an imbalance in the confidence coefficient. Specifically, 、 It is a scaling factor that is set manually or obtained through optimization based on a large number of experiments or data fitting during the model design or training phase. It 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 pitting defects. For example, if it is found in a large number of training samples that the micro-depression density contributes more to defect recognition, the confidence coefficient can be appropriately increased. The value of increases the influence of micro-depression density on the confidence level, thereby improving the final recognition accuracy. Therefore, the preset scale factor is essentially a weighted parameter for feature fusion, and its rationality is directly related to the reliability and judgment accuracy of the system's honeycomb surface recognition.

[0083] It can be seen from the confidence coefficient that for each detection area, the greater the performance value of the texture direction discrete index generated by the in-depth analysis of the discreteness of the local texture directional change of the detection area, the greater the performance value of the micro-depression density index generated by the in-depth analysis of the micro-depression distribution density of the detection area. That is, the greater the performance value of the confidence coefficient generated when the pre-trained convolutional neural network predicts the presence of slight pockmarks in the detection area, the greater the risk of slight pockmarks in the detection area, and vice versa.

[0084] The confidence coefficient generated by the pre-trained convolutional neural network when predicting the presence of slight pockmarks in the detection area is compared with the pre-set confidence coefficient reference threshold to perform suspected slight pockmarks determination. The specific steps are as follows:

[0085] If the confidence coefficient is greater than the confidence coefficient reference threshold, the detection area is classified as suspected slight pockmarks;

[0086] If the confidence coefficient is less than or equal to the confidence coefficient reference threshold, the detection area is divided into a normal plane.

[0087] When suspected slight pitting is identified in the detection area, an effective area of ​​the slight pitting in the detection area is determined;

[0088] The specific steps are:

[0089] After identifying suspected minor pitting in the inspection area, further determining the specific effective area of ​​the minor pitting can usually be accomplished by following the following steps:

[0090] First, based on the output of the convolutional neural network, the initial defect location of suspected slight pitting in the inspection area is located;

[0091] Secondly, around the initial position point, the local area is expanded using the image's high-frequency features, texture changes, grayscale differences, concave density and other information to form a candidate area for slightly pockmarked surfaces;

[0092] Then, the adaptive region growing method or edge detection + region segmentation algorithm is used to accurately segment the actual contour of the slightly pockmarked surface and eliminate the background or irrelevant noise areas;

[0093] Then, the segmented areas are analyzed for connected domains to eliminate pseudo-defects that are too small or have abnormal shapes, and only connected areas that meet the characteristics of slight pockmarks are retained.

[0094] Finally, the actual area, boundary, center coordinates and other parameters of the defect area are calculated to form a complete description of the range of the slight pitting surface, providing reliable range information support for subsequent defect level assessment and inspection strategy adjustment.

[0095] The actual inspection height is adaptively lowered based on the confidence score output by the convolutional neural network, improving the spatial resolution of the image and enhancing the recognition capabilities of the subsequent AI detection model. The inspection time of the drone is also dynamically adjusted based on the size of the minor pockmarks in the detection area and the actual inspection height, ensuring accurate identification of deep honeycombs.

[0096] After the convolutional neural network outputs the confidence coefficient for the inspection area, it is first compared with the pre-set confidence reference threshold. If the confidence coefficient is greater than the confidence coefficient reference threshold, it is determined that there is suspected slight pitting in the area. At this time, in order to further improve the imaging resolution of the suspected defect area, the following adaptive height reduction strategy is used to adaptively reduce the height of the drone inspection:

[0097] ,in: is the actual inspection height after adjustment, For the pre-configured optimal inspection height (e.g. 15m), It is the inspection height reduction coefficient (range 0~1), used to control the reduction range. is the confidence coefficient of the current recognition, is the preset confidence reference threshold;

[0098] This step uses a confidence-driven relative ratio reduction. Rather than a rigid reduction, the reduction varies nonlinearly with the confidence level. As the confidence level approaches 1, the height reduction becomes more significant. This adaptively improves image resolution, resulting in higher-quality images. This provides rich, clear surface information for subsequent convolutional neural network re-interpretation and deep honeycomb detection, preventing common pitting or false features from being misidentified as honeycombs.

[0099] After completing the altitude reduction and obtaining a high-resolution image, the effective area of ​​the slightly pitted surface within the determined detection area is then dynamically determined based on the current actual inspection altitude and the effective area of ​​the slightly pitted surface. The following area-altitude joint control model is used, and the specific expression is:

[0100] ,in: 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 ​​the detected slight pockmarks, is the standardized defect area (which can be set according to engineering experience, for example, 50cm²), is the area ratio adjustment coefficient, which controls the impact of area on time. It is the adjustment coefficient of height reduction ratio, which controls the impact of reduction degree on time;

[0101] This step dynamically adjusts the drone's dwell time by jointly driving the area and altitude to ensure that: the larger the area, the longer the dwell time, ensuring that enough defect detail images are obtained; the more the altitude is lowered, the higher the risk of the current area (higher confidence), and it also takes longer to collect additional images.

[0102] The above solution can effectively solve the problems of "minor pitting is difficult to accurately identify" and "deep honeycombs are easy to miss" in existing drone inspections of concrete appearance, and achieve high adaptability of the inspection process and improved defect recognition accuracy. This solution uses a dynamic inspection mechanism driven by confidence scores to enable drones to actively and adaptively adjust the inspection height and inspection time when suspected minor pitting is found, thereby significantly improving image resolution and the ability to express defect feature details. At the same time, combined with feature depth analysis and high-recognition feature vector input, the convolutional neural network's ability to recognize minor pitting and dark honeycombs is enhanced. Ultimately, it can effectively reduce the defect omission rate and false detection rate during the inspection process, achieve accurate, stable and efficient identification of hidden defects such as minor pitting and deep honeycombs on the surface of concrete structures, and significantly improve the reliability and practicality of intelligent inspections of concrete appearance quality.

[0103] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0104] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0105] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. An intelligent inspection method for concrete appearance quality, characterized in that: The following steps are involved: Based on the requirements for detection accuracy and efficiency of honeycomb surface, the optimal inspection height is configured for drone intelligent inspection; The drone inspects the surface of the concrete structure according to the preset route and the set optimal inspection height. The onboard high-resolution imaging system continuously captures the inspection area to form an inspection image data stream. During the inspection process, for each inspection area, the image data collected during the inspection process is aggregated in real time to construct an analysis set for the inspection area. Key features that can characterize slight pockmarks 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, thereby improving the recognition confidence of slight pockmarks. The analyzed features are input into a pre-trained convolutional neural network, which outputs a confidence score for the detection area. Based on the score, a determination is made regarding suspected minor pockmarks within the detection area. When suspected slight pitting is identified in the detection area, an effective area of ​​the slight pitting in the detection area is determined; The actual inspection height is adaptively lowered based on the confidence score output by the convolutional neural network, and the inspection time of the drone is dynamically adjusted according to the size of the slight pockmarks in the detection area and the actual inspection height to accurately identify deep honeycombs.

2. The intelligent inspection method for concrete appearance quality according to claim 1, characterized in that: To meet the dual requirements of honeycomb surface detection accuracy and detection efficiency, the optimal inspection height is configured for drone intelligent inspection. The specific steps include: Based on the actual project needs, the minimum detection size and required image resolution of honeycomb surface defects should be clarified; Calculate the flight altitude range that meets the resolution requirement based on the imaging device parameters carried by the UAV; Taking into account the inspection efficiency, the spatial characteristics of the inspection area, and flight safety, the optimal inspection altitude is preliminarily determined; Verify the image clarity and feature representation capabilities at the selected altitude through simulation or on-site test flights to ensure effective recognition of common honeycomb surfaces at this altitude while enabling efficient inspections at the flight speed. Combined with the characteristic distribution of AI recognition model training data, the selected inspection height is fine-tuned and optimized, and finally the optimal inspection operation height suitable for honeycomb surface detection is determined.

3. The intelligent inspection method for concrete appearance quality according to claim 1, characterized in that: For each detection area, key features characterizing slight pockmarks are extracted from the analysis set formed by the image data. The extracted features include the discrete degree of local texture directional changes and the distribution density of micro-depressions. After in-depth analysis of the extracted key features, texture directional discrete index and micro-depression density index are generated respectively. The texture directional discrete index and micro-depression density index are used as quantitative indicators to reflect the presence of slight pockmarks in the detection area, providing highly recognizable input for the convolutional neural network and improving the recognition confidence of slight pockmarks.

4. The intelligent inspection method for concrete appearance quality according to claim 3, characterized in that: The analyzed feature vector composed of texture direction discrete indicators and micro-depression density indicators is input into a pre-trained convolutional neural network. The convolutional neural network outputs a confidence coefficient, and based on the confidence coefficient, it is identified whether there is slight pockmarking in the detection area, and a suspected slight pockmarking judgment is performed.

5. The intelligent inspection method for concrete appearance quality according to claim 4, characterized in that: The confidence coefficient generated by the pre-trained convolutional neural network when predicting the presence of slight pockmarks in the detection area is compared with the pre-set confidence coefficient reference threshold to perform suspected slight pockmarks determination. The specific steps are as follows: If the confidence coefficient is greater than the confidence coefficient reference threshold, the detection area is classified as suspected slight pockmarks; If the confidence coefficient is less than or equal to the confidence coefficient reference threshold, the detection area is divided into a normal plane.

6. The intelligent inspection method for concrete appearance quality according to claim 5, characterized in that: When suspected slight pitting is identified in the detection area, the effective area of ​​the slight pitting in the detection area is determined. The specific steps are as follows: After identifying suspected minor pitting in the inspection area, to further determine the specific effective area of ​​the minor pitting, follow the steps below: Based on the output of the convolutional neural network, the initial defect point of the suspected slight pitting surface in the inspection area is located; Around the initial position point, the image is expanded locally to form a candidate area of ​​slightly pockmarked surface; Adopting adaptive region growing method or edge detection + region segmentation algorithm to accurately segment the actual contour of the slightly pockmarked surface; Perform connected domain analysis on the segmented areas to eliminate pseudo defects that are too small or have abnormal shapes, and only retain connected areas that meet the characteristics of slight pockmarks; The defect area parameters are calculated to form a complete description of the slight pitting range.

7. The intelligent inspection method for concrete appearance quality according to claim 6, characterized in that: After the convolutional neural network outputs the confidence coefficient for the inspection area, it is first compared with the pre-set confidence reference threshold. If the confidence coefficient is greater than the confidence coefficient reference threshold, it is determined that there is suspected slight pitting in the area. At this time, in order to further improve the imaging resolution of the suspected defect area, the following adaptive height reduction strategy is used to adaptively reduce the height of the drone inspection: ,in: is the actual inspection height after adjustment, The optimal inspection height is pre-configured. It is the inspection height reduction coefficient, used to control the reduction range. is the confidence coefficient of the current recognition, is the preset confidence reference threshold.

8. The intelligent inspection method for concrete appearance quality according to claim 7, characterized in that: After completing the altitude reduction and obtaining a high-resolution image, the effective area of ​​the slightly pitted surface within the determined detection area is then dynamically determined based on the current actual inspection altitude and the effective area of ​​the slightly pitted surface. The following area-altitude joint control model is used, and the specific expression is: ,in: 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 ​​the detected slight pockmarks, is the normalized defect area, is the area ratio adjustment coefficient, which controls the impact of area on time. It is the adjustment coefficient of height reduction ratio, which controls the impact of reduction degree on time.

Images