Outer wall heat insulation coating appearance quality detection method based on machine learning

Abstract

The invention provides an exterior wall heat insulation coating appearance quality detection method based on machine learning, an illumination intensity-shooting parameter mapping relation database is constructed through experiments, classification is carried out according to weather types to form a plurality of sub-databases, and shooting parameters in each sub-database are configured with different weight coefficients. During use, the system acquires external illumination intensity and weather type data in real time, adaptively adjusts shooting parameters in combination with the sub-database, analyzes a historical illumination trend to pre-judge a future illumination change direction and amplitude, and adjusts the parameters in advance to ensure image acquisition quality. And inputting an outer wall image acquired after parameter optimization into the trained defect identification model for defect identification, and outputting defect detailed information and an appearance quality evaluation grade. According to the method, through an intelligent environment adaptation and pre-judgment mechanism, the influence of illumination fluctuation on detection is effectively reduced, the accuracy and robustness of defect identification are improved, and automatic and efficient appearance quality detection is realized.

Description

Technical Field

[0001] This invention relates to the field of wall appearance inspection technology, and in particular to a method for inspecting the appearance quality of exterior wall thermal insulation coatings based on machine learning. Background Technology

[0002] As a core material for building energy conservation and structural protection, the appearance quality of exterior wall insulation coatings directly determines the building's thermal insulation performance, durability, and aesthetics. Defects such as cracking, peeling, discoloration, or bulging can not only increase building energy consumption by more than 30%, but also potentially lead to safety hazards such as water seepage and structural corrosion. Currently, the appearance quality inspection of exterior wall insulation coatings mainly relies on two types of technical methods, both of which have significant technical bottlenecks, making it difficult to meet the demands for large-scale, high-precision, and high-efficiency testing.

[0003] Traditional manual inspection methods require inspectors to work at heights using scaffolding, suspended platforms, and other equipment. This not only results in low work efficiency (less than 500 square meters of inspection area per person per day) but is also heavily influenced by subjective experience. The identification rate for defects such as tiny cracks with a width of less than 0.5 mm and slight discoloration with a color difference ΔE < 5 is less than 60%. In addition, working at heights poses safety risks such as falls and electric shocks. It is also extremely unsuitable for high-rise buildings and complex terrain scenarios. With the development of intelligent construction in the industry, manual inspection has gradually become unable to meet the timeliness and accuracy requirements of modern construction projects.

[0004] In recent years, automated inspection technology based on drones and machine learning has been gradually applied to the field of exterior wall inspection. However, in actual inspection, it is severely affected by ambient light, resulting in key problems such as unstable image quality and large fluctuations in model detection performance. When drones take pictures, the ambient light intensity changes dynamically (e.g., the light intensity can reach 100,000 lux at noon on a sunny day, only 5,000 lux on a cloudy day, and the light intensity fluctuation frequency can reach 1 time / minute in cloudy weather). Traditional fixed shooting parameters (e.g., fixed shutter speed 1 / 1000s, aperture f / 8) cannot be adapted to complex lighting environments, resulting in overexposure (loss of details under strong light), underexposure (increased noise under weak light), or color distortion (color deviation of paint under low color temperature). When such images are input into machine learning models, the defect recognition accuracy will decrease by 25%-40%, especially for small-sized, low-contrast defects (such as bulges and fine cracks), the false negative rate exceeds 50%, reducing the efficiency of inspection. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a machine learning-based method for inspecting the appearance quality of exterior wall thermal insulation coatings, which overcomes or at least partially solves the above problems.

[0006] This invention discloses a machine learning-based method for inspecting the appearance quality of exterior wall thermal insulation coatings, the method comprising: Optimal shooting parameters under different light intensities were collected through experiments, and a light intensity-shooting parameter mapping relationship database was established. The light intensity-shooting parameter mapping relationship database was then classified according to weather type, and multiple sub-databases of light intensity-shooting parameter mapping relationships for different weather types were established. Different weight coefficients were configured for the shooting parameters in each sub-database of light intensity-shooting parameter mapping relationships for each weather type. Real-time acquisition of external light intensity data and weather type; adaptive adjustment of shooting parameters based on real-time external light intensity data, weather type, and a sub-database of light intensity-shooting parameter mapping relationships for multiple weather types; analysis of light intensity change trends within a preset time period; prediction of future light change directions; advance adjustment of shooting parameters based on prediction results; and acquisition of exterior wall images based on the adjusted shooting parameters. A trained exterior wall thermal insulation coating appearance defect recognition model is used to identify appearance defects in exterior wall images and generate appearance quality inspection results for the exterior wall thermal insulation coating; the appearance quality inspection results include detailed defect information and an evaluation level of the appearance quality of the exterior wall thermal insulation coating.

[0007] This invention has the following advantages: This invention presents a machine learning-based method for inspecting the appearance quality of exterior wall insulation coatings. It experimentally constructs a database mapping the relationship between light intensity and shooting parameters, and categorizes this database into multiple sub-databases based on weather type. Each sub-database has shooting parameters configured with different weight coefficients for optimization. During operation, the system collects real-time data on external light intensity and weather type, adaptively adjusts shooting parameters based on the sub-databases, and analyzes historical light trends to predict future light changes in direction and magnitude, adjusting parameters in advance to ensure image acquisition quality. The optimized exterior wall images are then input into a trained defect recognition model for defect identification, outputting detailed defect information and an appearance quality assessment level. This method, through intelligent environmental adaptation and prediction mechanisms, effectively reduces the impact of light fluctuations on detection, improves the accuracy and robustness of defect identification, achieves automated and efficient appearance quality inspection, reduces manual costs and subjective errors, and enhances the efficiency and reliability of building construction quality monitoring. Attached Figure Description

[0008] Figure 1 This is a flowchart of the steps of a machine learning-based method for inspecting the appearance quality of exterior wall thermal insulation coatings provided in an embodiment of the present invention; Figure 2 This is a flowchart for detecting the appearance quality of exterior wall thermal insulation coatings provided in an embodiment of the present invention. Detailed Implementation

[0009] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0010] Reference Figure 1 The diagram illustrates a flowchart of a machine learning-based method for inspecting the appearance quality of exterior wall thermal insulation coatings, as provided in an embodiment of the present invention. The method may specifically include the following steps: Step 101: Collect the optimal shooting parameters under different light intensities through experiments, establish a light intensity-shooting parameter mapping relationship database, and classify the light intensity-shooting parameter mapping relationship database according to weather type to establish multiple weather type light intensity-shooting parameter mapping relationship sub-databases; configure different weight coefficients for the shooting parameters in the light intensity-shooting parameter mapping relationship sub-database of each weather type. Step 102: Collect real-time data on ambient light intensity and weather type. Based on the real-time ambient light intensity data and weather type, as well as the sub-database of light intensity-shooting parameter mapping relationships for multiple weather types, adaptively adjust the shooting parameters. At the same time, analyze the trend of light intensity change within a preset time period, predict the direction of light change in the future, adjust the shooting parameters in advance based on the prediction results, and collect images of the exterior wall based on the adjusted shooting parameters. Step 103: Use the trained exterior wall thermal insulation coating appearance defect recognition model to identify appearance defects in the exterior wall image and generate appearance quality inspection results for the exterior wall thermal insulation coating; the appearance quality inspection results include detailed defect information and an evaluation level of the appearance quality of the exterior wall thermal insulation coating.

[0011] In an optional embodiment of the present invention, optimal shooting parameters under different light intensities are collected experimentally to establish a database of light intensity-shooting parameter mapping relationships, including: Under controlled lighting conditions, different light intensity gradients are set, and for each light intensity, multiple sets of shooting parameters are used to acquire images. The image quality is evaluated based on an image quality assessment system, and the comprehensive score of each parameter combination is calculated by weighted scoring to select the best shooting parameter combination; the image quality assessment system includes sharpness, color reproduction, and noise level. A database of the mapping relationship between light intensity and shooting parameters is established based on each light intensity and the optimal shooting parameter combination corresponding to each light intensity.

[0012] In an optional embodiment of the present invention, the shooting parameters are adaptively adjusted based on real-time ambient light intensity data, weather type, and a sub-database of light intensity-shooting parameter mapping relationships for multiple weather types, including: Based on the real-time weather type, select the light intensity-shooting parameter mapping relationship sub-database corresponding to the weather type, and select the shooting parameters corresponding to the real-time external light intensity data in the light intensity-shooting parameter mapping relationship sub-database. Apply the weight coefficients corresponding to the light intensity-shooting parameter mapping relationship sub-database to the shooting parameters corresponding to the real-time external light intensity data. When the weather type changes abruptly, a tiered switching strategy is used to switch the sub-database of the light intensity-shooting parameter mapping relationship; When the weather type changes gradually, a linear interpolation algorithm is used to achieve a smooth transition of the weight coefficients between the sub-databases of different light intensities and shooting parameters.

[0013] In an optional embodiment of the present invention, the shooting parameters are adaptively adjusted based on real-time ambient light intensity data, weather type, and a sub-database of light intensity-shooting parameter mapping relationships for multiple weather types. The method further includes: The shooting parameters are adjusted in a graded manner according to the rate of change of light. When the rate of change of light is less than or equal to the preset rate of change of light threshold, the shooting parameters are gradually adjusted according to the preset step size. When the rate of change of light is greater than the preset rate of change of light threshold, the shooting parameters are adjusted in a fast adjustment mode and the shooting parameters are adjusted within the specified time.

[0014] In an optional embodiment of the present invention, the method further includes: Based on the type of building scene, a preset scene correction value is applied to the weighting coefficients to optimize the image acquisition quality of special building scenes.

[0015] In an optional embodiment of the present invention, analyzing the trend of light intensity change within a preset time period, predicting the direction of light change at future moments, and adjusting shooting parameters in advance based on the prediction results includes: The continuously collected ambient light intensity data is filtered, and the LSTM time series prediction model is used to model the light intensity change trend of the filtered ambient light intensity data, generating the light intensity change direction and range of future time. Based on the generated prediction results, a differentiated pre-adjustment strategy is executed: when the predicted light intensity changes positively, the shutter speed is pre-adjusted using the first adjustment strategy; when the predicted light intensity changes negatively and the change is less than a preset threshold, the aperture is pre-adjusted using the second adjustment strategy; when the predicted light intensity change is stable, the parameter maintenance mode is activated for fine-tuning and calibration.

[0016] In an optional embodiment of the present invention, after the step of acquiring the exterior wall image, the method further includes: Denoising of exterior wall images based on median filtering; Using the white area of ​​the wall as a reference, adjust the channel gain of R, G, and B in the exterior wall image; An enhancement algorithm is used to enhance the external wall image and highlight the features of the target area.

[0017] In an optional embodiment of the present invention, before the step of using a trained exterior wall thermal insulation coating appearance defect recognition model to identify appearance defects in the exterior wall image, the method further includes: The sharpness of the acquired exterior wall images is evaluated. If multiple consecutive frames fail to meet the standard, a reshoot mechanism is automatically initiated and the current illumination parameter deviation value is recorded. The illumination parameter deviation value is used for weight coefficient optimization.

[0018] In an optional embodiment of the present invention, a trained exterior wall thermal insulation coating appearance defect recognition model is used to identify appearance defects in an exterior wall image, generating an appearance quality inspection result for the exterior wall thermal insulation coating, including: The exterior wall image is input into the trained exterior wall thermal insulation coating appearance defect recognition model. The trained exterior wall thermal insulation coating appearance defect recognition model performs multi-scale feature inference and outputs detailed defect information, including the number of defects, defect type, defect location, defect distribution range, and defect severity. A quality score is calculated based on the number of defects, the severity of defects, and the distribution range of defects. The appearance quality assessment level is determined based on the quality score. The system integrates information to generate an appearance quality inspection report for exterior wall insulation coatings, including basic image information, detailed defect information, and appearance quality assessment level.

[0019] In an optional embodiment of the present invention, the exterior wall thermal insulation coating appearance defect recognition model is based on the YOLOv8-nano architecture, with dynamic mosaic enhancement added to the input end, a lightweight attention mechanism added to the backbone network, multi-scale feature fusion optimized in the neck network, and a four-classification detection head set in the output layer. The training process of the appearance defect recognition model for exterior wall thermal insulation coatings includes: Collect and label the dataset of exterior wall images, and after quality verification, divide it into training set, validation set and test set according to a predetermined ratio. Combine the sampled small defect samples and generate easily confused defect differential samples to enhance the data. The model is trained using a phased strategy, including: a first warm-up phase, freezing the backbone network and using a low learning rate to allow the model to quickly adapt to the data; a second fine-tuning phase, unfreezing all network layers, using an adaptive sample selection mechanism to focus on difficult examples, and using cosine annealing to balance convergence and stability; and a third fine-tuning phase, reducing the learning rate and increasing the weight of the defect edge loss term to enhance the ability to locate small targets. The model optimization employs pruning operations, quantization precision processing, and knowledge distillation techniques. Based on the evaluation results of the validation and test sets, supplementary training data is used to iterate the model for false positives and false negatives, thus completing the training and optimization of the model.

[0020] The following will be combined with the appendix Figure 2 The present invention will be described in further detail below: S1. Equip a light intensity sensor, a camera, and a control module on a drone. The control module is electrically connected to the light intensity sensor and the high-definition camera, respectively. Collect the optimal shooting parameters under different light intensities through experiments and establish a database of light intensity-shooting parameter mapping relationship. Dynamic weighting coefficients are introduced to create sub-databases for different weather types, with each sub-database corresponding to a different weighting coefficient. S2. During the flight of the drone, the light intensity sensor collects the external light intensity data in real time and transmits it to the control module. The control module adaptively adjusts the shooting parameters of the high-definition camera based on the established mapping relationship database to complete the acquisition of the external wall image. The control module analyzes the trend of light intensity change over 5 consecutive seconds and predicts the direction of light change 500ms in advance, thus achieving predictive and adaptive forward control. S3. Perform preprocessing operations on the acquired images to obtain standardized image data; S4. Input the preprocessed image to be detected into the trained exterior wall insulation coating appearance defect recognition model, output the appearance quality detection result of the exterior wall insulation coating, and output the detection report. The report includes: the defect category, location, range, specific information and the evaluation level of the coating appearance quality in the image; wherein, the exterior wall insulation coating appearance defect recognition model is obtained by constructing an exterior wall insulation coating appearance defect recognition model and training and optimizing the model using a pre-collected and labeled standardized image dataset.

[0021] Traditional exterior wall inspection is easily affected by lighting and weather conditions (such as overexposure in strong light or underexposure in cloudy weather, which can render images invalid). However, the method of this invention achieves stable inspection in complex environments through a dynamic database and look-ahead control.

[0022] This invention establishes a light intensity-shooting parameter mapping database through experiments, and divides it into sub-databases according to weather type and assigns dynamic weights to avoid the failure of a single parameter to adapt to different weather conditions. For example, it automatically reduces the exposure and increases the shutter speed in bright sunlight, and increases the ISO (sensitivity) and extends the exposure in cloudy weather, ensuring consistent image clarity under different lighting conditions.

[0023] The control module of this invention analyzes the trend of light changes within 5 seconds and predicts the direction of light 500ms in advance, such as a sudden drop in light caused by cloud cover or an increase in light caused by the movement of the sun, and adjusts the shooting parameters in advance. Compared with the lagging control that adjusts after the light changes, it can avoid blurry, overexposed / underexposed images caused by untimely parameter adjustment, reduce invalid data collection, and is especially suitable for outdoor scenes with frequent light fluctuations.

[0024] Traditional exterior wall inspection (especially for high-rise buildings) relies on manual suspended platforms / climbing operations, which are inefficient and pose high safety risks. This invention significantly optimizes the process by using drones and automation. The drone carries equipment to complete image acquisition, eliminating the need for manual high-altitude work and completely avoiding the risks of falls and electric shock. It also saves time on setting up suspended platforms and transporting equipment, increasing the efficiency of a single inspection by 3-5 times. For example, a 10-story building requires 2-3 hours of traditional manual inspection, while automated drone data acquisition takes only 30-40 minutes.

[0025] This invention, from illumination acquisition → parameter adjustment → image acquisition → preprocessing → model recognition → report output, requires no manual intervention throughout the entire process. The control module automatically adjusts the shooting parameters, the machine learning model automatically identifies defects, and finally directly outputs a standardized report. This avoids the tedious operation of manually screening images and manually annotating defects, reducing labor costs. Traditional inspection requires 2-3 people to work together, while this invention only requires 1 person to operate the drone.

[0026] Based on real-time data from a light sensor and a preset mapping database, this invention allows for adaptive adjustment of camera parameters (exposure time, ISO, white balance, focal length), avoiding the problem of large differences in brightness / contrast between different areas of the same wall caused by traditional fixed-parameter shooting, and ensuring that the captured exterior wall images are highly uniform in color, clarity, and brightness.

[0027] This invention further eliminates interference information in images through preprocessing operations such as cropping, noise reduction (e.g., removing image blur caused by high-altitude wind and drone shaking noise), and grayscale / color correction, transforming the original image into standardized data that meets the model input requirements, thus avoiding the problems of noise being misjudged as defects and defects not being identified due to image interference.

[0028] Traditional appearance inspection relies on manual visual judgment, which is easily affected by the experience and fatigue of the inspectors, leading to missed or incorrect defects (such as small cracks and local color differences are easily overlooked). The core advantage of the method in this invention lies in the accurate recognition capability of the model.

[0029] This invention trains a model using a large number of pre-collected and labeled standardized defect datasets (covering common defects in exterior wall insulation coatings such as cracks, bulges, peeling, color differences, and impurities). The model can learn the characteristics of different defects (such as the line features of cracks, the raised contour features of bulges, and the color gradient features of color differences). Compared with the human eye, it can identify smaller (such as cracks with a width of <0.1mm) and more hidden (such as slight local color differences) defects, reducing the false negative rate.

[0030] The model of this invention is based on a fixed algorithm and feature matching logic to output results, and is not affected by differences in the experience of the inspectors, fatigue from long hours of work, or subjective judgment bias (such as different people may have different judgments on whether "slight color difference is a defect"). This ensures that the test results of the same batch and different batches are highly consistent, and the detection accuracy rate can reach more than 95% (the accuracy rate of traditional manual inspection is usually 80%-85%).

[0031] The practicality of the test results directly impacts subsequent quality control. The report output by this invention is characterized by complete, accurate, and actionable information. The report not only clearly defines the defect category (e.g., cracks, bulges) but also includes specific locations (e.g., the area on the east facade of the building, floors 3-5) and extents (e.g., crack length 2.5m, width 0.15mm, bulge area 0.8㎡). Compared to traditional, vague reports that only describe the existence of defects, this provides a quantitative basis for quality assessment. The report includes an assessment level for the coating's appearance quality and clearly defines the specific information about the defects, facilitating construction and supervision teams to quickly locate problem areas, develop targeted rectification plans, and promote the transformation of test results into actual quality control.

[0032] The steps in step S1 for collecting the optimal shooting parameters under different light intensities are as follows: The experimental environment was set up, including a tunable light source system, a standard test target, and a constant temperature and humidity experimental chamber; The equipment is calibrated, the light intensity gradient is set, and the combination of imaging parameters is designed. Construct a multi-dimensional image quality evaluation system, including evaluation of sharpness, color reproduction, and noise level; Data was collected, and based on the constructed image quality evaluation system, the optimal combination of shooting parameters under the current lighting intensity was selected, specifically: Set the target light intensity and start the test after stabilizing for 30 seconds; select 200 typical parameter combinations for shooting using the orthogonal experimental method; Five images were taken consecutively for each set of parameters, and the average value was used as the evaluation criterion. Record ambient temperature, humidity, and equipment operating status parameters; By employing the analytic hierarchy process (AHP) to assign weights to each evaluation index, calculating the comprehensive score for each group of parameters, and selecting parameter combinations with a comprehensive score ≥ 90, a database of the mapping relationship between light intensity and shooting parameters is obtained.

[0033] The experiment uses an LED light source array that simulates natural white light, and the output intensity of the light source can be precisely adjusted by the control system to ensure stable simulation of different outdoor lighting scenarios. The installation position of the light source must be consistent with the height and angle of the drone camera. A combination target of ISO 12233 resolution test chart + 24-color standard color chart is used. The resolution test chart is used to calibrate image sharpness, and the 24-color standard color chart is used to calibrate color reproduction. It contains 24 standard color values ​​of red, green, blue, and yellow, and can compare the color difference between the captured image and the standard color chart. The test target must be fixed on a substrate that is consistent with the material of the exterior wall (such as the cement mortar substrate or paint sample substrate commonly used for exterior walls) to avoid the test results being affected by the difference in reflectivity of the substrate material.

[0034] The cabin temperature is controlled within a range of 15-35℃ (covering most outdoor testing environments), and the humidity is controlled within a range of 30%-70% (to avoid fogging of the equipment lens due to high humidity and electrostatic interference due to low humidity). A non-reflective background board (such as a matte black board) must be installed inside the cabin to eliminate interference from environmental reflections on the shooting. At the same time, the cabin is reserved with equipment installation interfaces to ensure that the fixed positions of the drone camera and light intensity sensor are consistent with those during outdoor flight (e.g., the distance between the camera and the test target is 5-8m, simulating the drone's high-altitude shooting distance).

[0035] Using a standard lux meter (accuracy ±5 lux) as a benchmark, the sensor and the standard lux meter were placed in the experimental chamber at the same time. The readings of both were recorded under different light intensities (such as 100 lux, 500 lux, 1000 lux...10000 lux). The sensor error was corrected by linear regression (ensuring that the deviation between the sensor reading and the standard lux meter is ≤3%) to avoid subsequent parameter adjustment errors due to sensor inaccuracy.

[0036] First, lens distortion calibration is performed (using a checkerboard calibration board, taking checkerboard images from different angles, and eliminating radial / tangential distortion of the lens through algorithms); then, color calibration is performed (taking pictures of a 24-color standard color chart, comparing the RGB values ​​of each color block in the image with the Lab values ​​of the standard color chart, and correcting through a color matrix to ensure that the color difference ΔE≤2, which meets the color detection standard for exterior wall coatings); finally, sharpness calibration is performed (taking pictures of an ISO 12233 resolution test chart to ensure that the camera can clearly distinguish the 1000 line pairs / inch line group on the test chart, avoiding the impact of lens blur on defect identification).

[0037] Focusing on the needs of exterior wall inspection, four key parameters were selected: exposure time (1 / 100s-1 / 1000s), ISO sensitivity (100-800), white balance (auto / cloudy / sunny / fluorescent mode), and aperture (f / 2.8-f / 8, adjusting the amount of light entering the lens at a fixed focal length).

[0038] 200 combinations were designed using the orthogonal experimental method: based on the orthogonal array, the levels were evenly selected within the effective range of each parameter to ensure that the 200 combinations could cover the interaction and influence of parameters, while avoiding repeated testing (compared to full factorial experiments, orthogonal experiments can reduce the number of test groups by more than 80%, improving experimental efficiency while ensuring data validity).

[0039] Edge Sharpness quantization was used. Five regions, including the center and four corners, were selected from the ISO 12233 resolution test card image. The grayscale gradient of the black and white line edges in each region was calculated using the Sobel operator. The average gradient value of the five regions was taken as the sharpness score.

[0040] After setting the target light intensity, wait 30 seconds to ensure that the output of the adjustable light source system is stable (to avoid brightness fluctuations immediately after the light source is adjusted) and the sensor reading is stable (to eliminate reading deviations caused by sensor response delay). Starting the shooting at this time can ensure that the parameter combination test is accurately matched with the target light intensity. To avoid accidental errors caused by minor equipment vibrations (such as airflow in the experimental chamber or fluctuations in equipment current) during a single shot, the average of the evaluation metrics (sharpness, color reproduction, and noise) of the five images can more objectively reflect the actual effect of the parameter combination. The real-time temperature (accurate to 0.1℃), humidity (accurate to 1%), light source operating current (to avoid intensity deviation caused by light source aging), and camera operating voltage (to avoid voltage fluctuations affecting parameter stability) inside the experimental chamber need to be recorded simultaneously. This data can be used for subsequent database optimization.

[0041] The Analytic Hierarchy Process (AHP) weighting process involves inviting 5-8 exterior wall inspection experts and image processing engineers to form an evaluation panel. This panel compares the importance of sharpness, color reproduction, and noise level in a pairwise manner, constructs a judgment matrix, and calculates the weights. Typically, in exterior wall defect identification, sharpness has the highest weight (approximately 0.5), followed by color reproduction (approximately 0.3), and noise level has the lowest weight (approximately 0.2) (because defects such as cracks and bulges primarily rely on sharpness for identification, color difference defects rely on color reproduction, and noise can be partially eliminated through preprocessing). The overall score is calculated as follows: (Clarity Score × 0.5) + (Color Reproduction Score × 0.3) + (Noise Level Score × 0.2). Parameter combinations with an overall score ≥ 90 are selected because they demonstrate excellent performance across all three indicators (e.g., clarity ≥ 90, color reproduction ≥ 85, noise ≤ 10), ensuring that the acquired images fully meet the requirements of subsequent model recognition. If no combination with a score ≥ 90 is found under a certain illumination gradient, the top 3 highest-scoring combinations are selected as candidates (further screening will be conducted through actual outdoor testing), forming a mapping database of 1-3 optimal parameters for each illumination intensity.

[0042] The steps for adaptively adjusting shooting parameters by introducing dynamic weighting coefficients are as follows: Based on meteorological classification standards and special effects of external wall inspection scenarios, the database is divided into 6 basic categories, each corresponding to an independent sub-database; The system automatically identifies weather types by combining spectral data from a light sensor, temperature and humidity data from a drone-borne temperature and humidity sensor, and the API interface of a ground weather station. When the weather type changes abruptly, a tiered switching mechanism is adopted; When the weather type changes gradually, a linear interpolation algorithm is used to achieve a smooth transition of the weighting coefficients; For specific architectural scenarios, a scenario correction value of ±10% is added to the basic weighting coefficient; After each shot, the system fine-tunes the current weight coefficients based on the image quality assessment results and records the correction values ​​for subsequent database iterations.

[0043] The basic categories include: Sunny day (direct sunlight) Its characteristics include high light intensity (50,000-100,000 lux), concentrated spectrum, obvious shadows, and high light stability; The sub-database contains 1200 sets of lighting-parameter mapping relationships; Shutter speed (0.4), aperture (0.3), ISO (0.1), exposure compensation (0.2); The adjustment logic prioritizes controlling the amount of light entering the camera through shutter speed, adjusting the depth of field in conjunction with the aperture, and reducing ISO usage to lower noise.

[0044] Cloudy (diffuse light) Its characteristics include moderate light intensity (5000-30000 lux), uniform spectrum, no obvious shadows, and high light stability; The sub-database contains 1000 sets of lighting-parameter mapping relationships; Aperture (0.4), ISO (0.2), Shutter speed (0.2), Exposure compensation (0.2); Adjust the logic to prioritize opening the aperture to increase the amount of light entering the camera, and moderately increase the ISO to ensure the shutter speed (to avoid blurring caused by drone shaking).

[0045] Partly cloudy (intermittent light) Its characteristics include large fluctuations in light intensity (10,000-80,000 lux), high frequency of change (period 1-5 minutes), and complex spectrum; The sub-database contains 1500 sets of illumination-parameter mapping relationships (including dynamic change sequences); Shutter speed (0.5), exposure compensation (0.2), aperture (0.2), ISO (0.1); The adjustment logic responds quickly to changes in lighting conditions with shutter speed and corrects them in real time through exposure compensation, keeping the aperture relatively stable.

[0046] Sunrise / Sunset (Low color temperature light) Its characteristics include low to medium light intensity (1000-20000 lux), low color temperature (2000-3500K), and large color deviation; The sub-database contains 800 sets of lighting-parameter mapping relationships; White balance (0.3), exposure compensation (0.3), aperture (0.2), shutter speed (0.2); The adjustment logic prioritizes white balance correction to ensure color reproduction, while exposure compensation enhances shadow details.

[0047] Foggy weather (diffuse light) The characteristics include low light intensity (1000-10000 lux), low contrast, and severe scattering due to high levels of particulate matter in the air. The sub-database contains 600 sets of lighting-parameter mapping relationships; ISO (0.3), exposure compensation (0.3), aperture (0.2), shutter speed (0.2); Adjust the ISO to ensure brightness, increase exposure compensation to enhance contrast, and use a medium aperture to balance depth of field and light intake.

[0048] Rainy day (low light + reflection) Its characteristics include low light intensity (500-8000 lux), strong surface reflectivity, and high ambient humidity. The sub-database contains 500 sets of lighting-parameter mapping relationships; Aperture (0.3), shutter speed (0.3), exposure compensation (-0.2), ISO (0.2); Adjust the logic to reduce reflections by narrowing the aperture, increase the shutter speed to avoid raindrop blur, and reduce exposure compensation to prevent overexposure.

[0049] Complex weather types When the weather changes from sunny to cloudy, the dynamic weighting coefficient smoothly transitions from sunny mode to cloudy mode (transition time 3 minutes). When the weather changes from cloudy to rainy, the weighting coefficient changes from cloudy to rainy at a rate of 10% per minute. When the weather clears up after rain, the fog weighting factor is initially activated, and the mode is gradually switched to sunny mode as the humidity decreases.

[0050] The specific steps for real-time acquisition of external light intensity data in step S2 are as follows: After the drone enters the inspection route, the light intensity sensor collects data at a high frequency, simultaneously acquiring visible light and near-infrared dual-spectrum light data at a frequency of 30 times per second. After analog-to-digital conversion, a data packet containing the absolute value of light intensity, spectral distribution characteristic value and the rate of change within 1 second is generated and transmitted in real time to the edge computing unit of the control module via an industrial-grade CAN bus. After receiving the data, the control module calls the weather type recognition submodule. Combining the built-in temperature and humidity sensor data with the meteorological feature database of the preset geographical area, the weather type is determined within 50ms. The corresponding sub-database is automatically matched, and the interpolation algorithm is used to query the three sets of parameter records in the weather sub-database that are closest to the current light intensity. The initial parameter combination is generated by weighting the data according to the dynamic weight coefficient.

[0051] The absolute value of light intensity in this invention is not a single numerical value, but a comprehensive intensity value after dual-spectrum fusion, calculated as follows: Total light intensity = (visible light intensity × 0.6) + (near-infrared intensity × 0.4); The weighting is based on the fact that the camera's imaging mainly relies on visible light (accounting for 60%), but near-infrared intensity can correct the deviation of visible light overexposure under strong light (e.g., near-infrared intensity is high at noon on a sunny day, so the camera's exposure parameters need to be reduced, hence accounting for 40%). Spectral distribution characteristic values ​​are used to describe the spectral composition of illumination, such as: The visible light component is high (approximately 70%), the near-infrared component is low (approximately 30%), and the eigenvalues ​​are biased towards a high visible light weight. On cloudy days, there is more diffuse reflected light, and the proportion of near-infrared light increases (about 45%), so the eigenvalues ​​are biased towards high near-infrared weights. This parameter is used to assist in determining the weather type (e.g., if the feature value matches the high near-infrared ratio, it can be preliminarily determined to be a cloudy day), and at the same time, it corrects the camera's white balance parameters (e.g., on cloudy days, the white balance color temperature needs to be increased to avoid the image being too blue). The calculation formula is: Rate of change = [(maximum illuminance in the current second - minimum illuminance in the current second) / average illuminance in the current second] × 100%; The core function of the rate of change is to predict the stability of illumination. When the rate of change is >5%, it indicates that the illumination fluctuates frequently (such as the rapid movement of clouds), and the control module needs to strengthen the forward-looking adjustment. When the rate of change is <2%, the illumination is stable, and the frequency of parameter adjustment can be reduced to reduce the power consumption of the equipment.

[0052] The specific steps of adaptive adjustment in step S2 are as follows: For the parameter adjustment execution stage, the control module adopts a hierarchical response mechanism. When the rate of change in illumination is ≤10% / s, the parameters are adjusted gradually in 10% steps to avoid sudden image changes. When the rate of change is >10% / s, a fast adjustment mode is triggered, prioritizing the locking of the shutter speed and simultaneously calculating the aperture and ISO co-values. All parameters are switched within 100ms. The adjustment command is precisely controlled by the PWM signal to control the camera actuator. The shutter speed adjustment accuracy reaches ±1 / 1000s, and the aperture control error is ≤0.3f. The control module captures a 10% area from the center of the image in real time for sharpness assessment. If three consecutive frames fail to meet the standard, a reshoot mechanism is automatically initiated, and the current illumination parameter deviation value is recorded and uploaded to the ground database for weight coefficient optimization. Combined with GPS positioning information, latitude and longitude, acquisition time, and illumination parameter labels are added to each image to form an image dataset with environmental metadata, thus obtaining the acquired image data.

[0053] This invention adjusts parameters gradually in 10% increments when the lighting changes slowly (e.g., from sunny / cloudy to overcast, the lighting decreases from 8000 lux to 7200 lux within 10 seconds, a change rate of 8% / s). This avoids sudden changes in image brightness / contrast (such as sudden overexposure or underexposure) caused by one-step adjustments. This gradual adjustment allows continuously acquired images to maintain a smooth visual transition, preventing misjudgment of color difference defects due to sudden changes in brightness in the same area during subsequent model recognition. It also reduces mechanical wear on the camera's actuators (frequent large-scale adjustments can shorten the lifespan of shutter and aperture components).

[0054] The specific steps of look-ahead control are as follows: The control module has a built-in light trend prediction submodule, which synchronously starts the time series analysis function during the flight of the UAV. With a 5-second analysis window, it preprocesses 150 sets of raw data collected by the light intensity sensor, removes instantaneous interference noise through the Kalman filter algorithm, and retains the trend curve. The LSTM time series prediction model is used to model the trend of the filtered data. The model input includes three dimensions of features: the mean light intensity in the last 5 seconds, the standard deviation of the rate of change per second, and the intensity range of the current light intensity. The training samples are updated in real time through a sliding window. The model completes a prediction calculation within 30ms and outputs the direction and range of light intensity change in the next 500ms. For different prediction results, the control module executes a differentiated forward adjustment strategy. When it is predicted that the light intensity will increase by more than 10%, the shutter speed is pre-adjusted to 80% of the parameters corresponding to the predicted value, leaving a 20% buffer space. If it is predicted that the light intensity will decrease and the rate of change is small, the aperture is pre-increased by 0.5-1 stops while keeping the ISO stable. For a stable trend, the parameter maintenance mode is activated, and only ±3% fine-tuning calibration is performed.

[0055] The preprocessing in step S3 includes image denoising, color correction, and image enhancement. Image denoising is based on median filtering. Color correction is performed by adjusting the R, G, and B channel gains based on the white area of ​​the wall. Image enhancement is based on enhancement algorithms to highlight the features of the target area.

[0056] The appearance defect recognition model in step S4 is based on the YOLOv8-nano architecture. The input end is enhanced with dynamic mosaic, the backbone network is equipped with a lightweight attention mechanism, the neck network is optimized for multi-scale feature fusion, and the output layer is equipped with a four-class detection head.

[0057] The training and optimization steps for the appearance defect recognition model in step S4 are as follows: Data is collected in advance, and the quality of annotation is verified by manual sampling and automatic tools. The training, validation and test sets are divided equally in a 7:2:1 ratio. Then, the data is enhanced by sampling small defect samples and generating easily confused defect differentiation samples. In the warm-up phase, the backbone network is frozen and the data is quickly adapted using a low learning rate; in the fine-tuning phase, the entire network is unfrozen, and difficult cases are selected using adaptive samples, while cosine annealing learning rate is used to balance convergence and stability; in the fine-tuning phase, the learning rate is reduced and the weight of the defect edge loss term is increased to strengthen the ability to locate small targets. The model optimization employs pruning to reduce the number of parameters by 40%, quantizes to INT8 accuracy with minimal accuracy loss, and then performs knowledge distillation. During evaluation, both core and sub-indicators are considered, and supplementary data is used to iterate the model for false positives and false negatives, thus completing the training and optimization of the model.

[0058] The specific steps in step S4 are as follows: The preprocessed image data is input into the trained exterior wall thermal insulation coating appearance defect recognition model. The model performs multi-scale feature inference and outputs JSON data containing defect category, location, range and specific information, retaining high confidence results. Next, a quality score is calculated based on the number, severity, and distribution of defects, and an assessment level is assigned. The calculated quality score is then substituted into the assigned assessment level. The system integrates information to generate a PDF inspection report, including a report header, basic image information, a defect details table, quality assessment results, and attachments.

[0059] The beneficial effects of this invention are as follows: 1. This invention combines dynamic weighting coefficients with a weather sub-database to match optimal shooting parameters according to the weather, and proactively adjusts parameters in advance to cope with sudden changes in lighting, avoid overexposure and underexposure of images, ensure high image quality, lay the foundation for subsequent inspection, the system is highly adaptable and can cope with complex lighting scenarios, the model can be flexibly deployed and is suitable for inspection of buildings of different sizes, the report is stored in the cloud for easy tracking of paint quality, and it provides support for the maintenance of buildings throughout their entire life cycle, with a wide range of application scenarios.

[0060] 2. This invention replaces manual inspection with adaptive data collection by drones, which greatly shortens the inspection time and avoids risks; it achieves full-process automation, fast reasoning, strong batch processing capability, significantly improves efficiency, and reduces human intervention and errors.

[0061] 3. This invention has high defect detection accuracy, improving the detection precision of coating defects and minor cracks, solving the problem of missed detection and misjudgment by manual methods, quantifying scoring and classifying levels, and providing detailed information in the report to ensure compliance of the assessment and provide accurate reference for repair.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The various embodiments in this specification are described in a related manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for inspecting the appearance quality of exterior wall thermal insulation coatings based on machine learning, characterized in that, The method includes: Optimal shooting parameters under different light intensities were collected through experiments, and a light intensity-shooting parameter mapping relationship database was established. The light intensity-shooting parameter mapping relationship database was then classified according to weather type, and multiple sub-databases of light intensity-shooting parameter mapping relationships for different weather types were established. Different weight coefficients were configured for the shooting parameters in each sub-database of light intensity-shooting parameter mapping relationships for each weather type. Real-time acquisition of external light intensity data and weather type; adaptive adjustment of shooting parameters based on real-time external light intensity data, weather type, and a sub-database of light intensity-shooting parameter mapping relationships for multiple weather types; analysis of light intensity change trends within a preset time period; prediction of future light change directions; advance adjustment of shooting parameters based on prediction results; and acquisition of exterior wall images based on the adjusted shooting parameters. A trained exterior wall thermal insulation coating appearance defect recognition model is used to identify appearance defects in exterior wall images and generate appearance quality inspection results for the exterior wall thermal insulation coating; the appearance quality inspection results include detailed defect information and an evaluation level of the appearance quality of the exterior wall thermal insulation coating.

2. The method according to claim 1, characterized in that, Optimal shooting parameters under different light intensities were collected through experiments, and a database of light intensity-shooting parameter mapping relationships was established, including: Under controlled lighting conditions, different light intensity gradients are set, and for each light intensity, multiple sets of shooting parameters are used to acquire images. The image quality is evaluated based on an image quality assessment system, and the comprehensive score of each parameter combination is calculated by weighted scoring to select the best shooting parameter combination; the image quality assessment system includes sharpness, color reproduction, and noise level. A database of the mapping relationship between light intensity and shooting parameters is established based on each light intensity and the optimal shooting parameter combination corresponding to each light intensity.

3. The method according to claim 1, characterized in that, Based on real-time ambient light intensity data, weather type, and a sub-database mapping light intensity to shooting parameters for multiple weather types, the shooting parameters are adaptively adjusted, including: Based on the real-time weather type, select the light intensity-shooting parameter mapping relationship sub-database corresponding to the weather type, and select the shooting parameters corresponding to the real-time external light intensity data in the light intensity-shooting parameter mapping relationship sub-database. Apply the weight coefficients corresponding to the light intensity-shooting parameter mapping relationship sub-database to the shooting parameters corresponding to the real-time external light intensity data. When the weather type changes abruptly, a tiered switching strategy is used to switch the sub-database of the light intensity-shooting parameter mapping relationship; When the weather type changes gradually, a linear interpolation algorithm is used to achieve a smooth transition of the weight coefficients between the sub-databases of different light intensities and shooting parameters.

4. The method according to claim 1, characterized in that, Based on real-time ambient light intensity data, weather type, and a sub-database mapping light intensity to shooting parameters for multiple weather types, the system adaptively adjusts shooting parameters, including: The shooting parameters are adjusted in a graded manner according to the rate of change of light. When the rate of change of light is less than or equal to the preset rate of change of light threshold, the shooting parameters are gradually adjusted according to the preset step size. When the rate of change of light is greater than the preset rate of change of light threshold, the shooting parameters are adjusted in a fast adjustment mode and the shooting parameters are adjusted within the specified time.

5. The method according to claim 1, characterized in that, The method further includes: Based on the type of building scene, a preset scene correction value is applied to the weighting coefficients to optimize the image acquisition quality of special building scenes.

6. The method according to claim 1, characterized in that, Analyze the trend of light intensity changes within a preset time period, predict the direction of light changes in the future, and adjust shooting parameters in advance based on the prediction results, including: The continuously collected ambient light intensity data is filtered, and the LSTM time series prediction model is used to model the light intensity change trend of the filtered ambient light intensity data, generating the light intensity change direction and range of future time. Based on the generated prediction results, a differentiated pre-adjustment strategy is executed: when the predicted light intensity changes positively, the shutter speed is pre-adjusted using the first adjustment strategy; when the predicted light intensity changes negatively and the change is less than a preset threshold, the aperture is pre-adjusted using the second adjustment strategy; when the predicted light intensity change is stable, the parameter maintenance mode is activated for fine-tuning and calibration.

7. The method according to claim 1, characterized in that, After the step of acquiring images of the exterior walls, the following steps are also included: Denoising of exterior wall images based on median filtering; Using the white area of ​​the wall as a reference, adjust the channel gain of R, G, and B in the exterior wall image; An enhancement algorithm is used to enhance the external wall image and highlight the features of the target area.

8. The method according to claim 1, characterized in that, Before using the trained exterior wall thermal insulation coating appearance defect recognition model to identify appearance defects in exterior wall images, the following steps are also included: The sharpness of the acquired exterior wall images is evaluated. If multiple consecutive frames fail to meet the standard, a reshoot mechanism is automatically initiated and the current illumination parameter deviation value is recorded. The illumination parameter deviation value is used for weight coefficient optimization.

9. The method according to claim 1, characterized in that, A trained exterior wall insulation coating appearance defect recognition model is used to identify appearance defects in exterior wall images, generating appearance quality inspection results for the exterior wall insulation coating, including: The exterior wall image is input into the trained exterior wall thermal insulation coating appearance defect recognition model. The trained exterior wall thermal insulation coating appearance defect recognition model performs multi-scale feature inference and outputs detailed defect information, including the number of defects, defect type, defect location, defect distribution range, and defect severity. A quality score is calculated based on the number of defects, the severity of defects, and the distribution range of defects. The appearance quality assessment level is determined based on the quality score. The system integrates information to generate an appearance quality inspection report for exterior wall insulation coatings, including basic image information, detailed defect information, and appearance quality assessment level.

10. The method according to claim 1, characterized in that, The exterior wall thermal insulation coating appearance defect recognition model is based on YOLOv8-nano architecture. Dynamic mosaic enhancement is added to the input layer, a lightweight attention mechanism is added to the backbone network, multi-scale feature fusion is optimized in the neck network, and a four-class classification detection head is set in the output layer. The training process of the appearance defect recognition model for exterior wall thermal insulation coatings includes: Collect and label the dataset of exterior wall images, and after quality verification, divide it into training set, validation set and test set according to a predetermined ratio. Combine the sampled small defect samples and generate easily confused defect differential samples to enhance the data. The model is trained using a phased strategy, including: a first warm-up phase, freezing the backbone network and using a low learning rate to allow the model to quickly adapt to the data; a second fine-tuning phase, unfreezing all network layers, using an adaptive sample selection mechanism to focus on difficult examples, and using cosine annealing to balance convergence and stability; and a third fine-tuning phase, reducing the learning rate and increasing the weight of the defect edge loss term to enhance the ability to locate small targets. The model optimization employs pruning operations, quantization precision processing, and knowledge distillation techniques. Based on the evaluation results of the validation and test sets, supplementary training data is used to iterate the model for false positives and false negatives, thus completing the training and optimization of the model.

Images