Bubble defect detection method and system for radiation refrigeration coating based on machine vision

Abstract

The invention relates to the technical field of image recognition, in particular to a bubble defect detection method and system for a radiation refrigeration coating based on machine vision, and the method comprises the steps: obtaining a visible light image, a near-infrared image and infrared thermal imaging data; performing environment adaptive preprocessing on the visible light image, the near-infrared image and the infrared thermal imaging data to correct illumination fluctuation influence and construct multi-scale temperature gradient features; extracting spatial texture features and material component features in the preprocessed image data, enhancing feature expressions of the spatial texture features and the material component features in combination with an attention module, and performing feature fusion to obtain fused features and bubble density; performing bubble defect pixel-level segmentation on the fusion feature to obtain a bubble segmentation mask; and extracting multi-dimensional bubble characteristic parameters from the bubble segmentation mask, and predicting the performance loss of the radiation refrigeration coating in combination with the multi-dimensional bubble characteristic parameters, the multi-scale temperature gradient characteristics and the bubble density to obtain a coating defect detection result.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically a method and system for detecting bubble defects in radiation-cooled coatings based on machine vision. Background Technology

[0002] Radiation-cooling coatings are key passive cooling materials for addressing the climate crisis and the energy consumption crisis of refrigeration. They require no external energy input, achieving sustainable cooling through natural processes. They precisely utilize the 8-13 micrometer infrared transparency window of the Earth's atmosphere, allowing heat from the object's surface to dissipate into deep space through thermal radiation. Simultaneously, they maintain extremely high reflectivity to solar radiation in the 0.3-2.5 micrometer wavelength range, ensuring that even during direct sunlight in broad daylight, the object's surface temperature remains below ambient temperature, achieving zero energy consumption, zero emissions, zero noise, and refrigerant-free cooling. As an important industrialization vehicle for radiation-cooling technology, the material systems of radiation-cooling coatings are diversifying, with polymer-based composite materials being the mainstream. By combining highly efficient solar diffusers such as titanium dioxide and barium sulfate with a high-infrared-emissivity polymer matrix, parameters such as particle size and dispersion stability are continuously optimized to pursue higher solar reflectivity and infrared emissivity.

[0003] In the preparation of radiation-cooled coatings, air bubbles are a key defect affecting their performance and must be strictly controlled throughout the entire process to ensure that the coating density reaches over 99.5%. These bubbles mainly form during the dispersion, application, and curing stages. During dispersion, air entrained by high-speed stirring or sand milling, and air released from the breakup of rare earth particle agglomerates, can form bubbles. During curing, rapid solvent evaporation or small molecule gases released from partial resin chemical reactions, if unable to escape in time, can also form bubbles. Excessive dispersant dosage or excessively high coating viscosity can further exacerbate the stable retention of bubbles. The presence of bubbles can cause random light scattering due to the difference in refractive index between air and the coating, leading to a 5% to 10% decrease in solar reflectivity. It can also reduce the effective load-bearing area of ​​the coating, resulting in a decrease in tensile and impact strength. Furthermore, it can act as a reservoir for moisture and contaminants, accelerating substrate corrosion in salt spray environments and significantly shortening salt spray resistance time.

[0004] Currently, the detection of bubble defects in coatings mainly relies on manual visual inspection or traditional optical inspection methods. These methods have drawbacks such as low detection accuracy, poor efficiency, and inability to predict the impact of bubble defects on coating performance. In particular, for materials with special optical properties, such as radiation-cooled coatings, traditional single-band imaging methods are difficult to accurately reflect the impact of bubble defects on coating performance. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method and system for detecting bubble defects in radiation-cooled coatings based on machine vision, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The present invention provides a machine vision-based method for detecting bubble defects in radiation-cooled coatings, comprising the following steps:

[0008] Acquire image data of the radiation-cooled coating, the image data including visible light images, near-infrared images, and infrared thermal imaging data;

[0009] The visible light image, the near-infrared image, and the infrared thermal imaging data are subjected to environment-adaptive preprocessing to correct the effects of illumination fluctuations and construct multi-scale temperature gradient features, thereby obtaining preprocessed image data.

[0010] A dual-branch deep learning network is used to extract spatial texture features and material composition features from the preprocessed image data. An attention module is then used to enhance the feature representation of the spatial texture features and the material composition features and to perform feature fusion to obtain fused features and bubble density.

[0011] The fused features are segmented at the pixel level to obtain a bubble segmentation mask;

[0012] Multi-dimensional bubble feature parameters are extracted from the bubble segmentation mask. By combining the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density, the performance loss of the radiation-cooled coating is predicted to obtain the coating defect detection result.

[0013] This application also provides a machine vision-based system for detecting bubble defects in radiation-cooled coatings, including:

[0014] An acquisition unit is used to acquire image data of the radiation-cooled coating, the image data including visible light images, near-infrared images and infrared thermal imaging data;

[0015] The preprocessing unit is used to perform environmentally adaptive preprocessing on the visible light image, the near-infrared image, and the infrared thermal imaging data to correct the effects of illumination fluctuations and construct multi-scale temperature gradient features to obtain preprocessed image data.

[0016] The feature extraction unit is used to extract spatial texture features and material composition features from the preprocessed image data through a dual-branch deep learning network, and to enhance the feature representation of the spatial texture features and material composition features by combining an attention module and performing feature fusion to obtain fused features and bubble density.

[0017] A segmentation unit is used to perform pixel-level segmentation of the fused features to obtain a bubble segmentation mask;

[0018] The prediction unit is used to extract multi-dimensional bubble feature parameters from the bubble segmentation mask, and combine the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density to predict the performance loss of the radiation-cooled coating, so as to obtain the coating defect detection result.

[0019] The beneficial effects of this invention are as follows: The machine vision-based method and system for detecting bubble defects in radiation-cooled coatings firstly captures the optical reflection characteristics, material composition differences, and thermal distribution features of bubble defects through multimodal data acquisition using visible light, near-infrared, and infrared thermal imaging. This overcomes the limitations of traditional single-band imaging, which struggles to cover the multidimensional influences of bubbles, laying a data foundation for accurately linking defects with performance. Secondly, environmentally adaptive preprocessing corrects for illumination fluctuation interference and constructs multi-scale temperature gradient features, avoiding feature distortion caused by environmental factors. Simultaneously, a dual-branch deep learning network specifically extracts spatial texture and material composition features, combined with an attention module to enhance the expression of key bubble features, and a dynamic weight fusion strategy... The system achieves efficient and accurate integration of features, and the synchronously calculated bubble density provides a basis for subsequent efficient inference. Furthermore, pixel-level segmentation accurately obtains bubble contour masks through improved algorithms, ensuring the accuracy of multi-dimensional bubble feature parameter extraction. Finally, these accurately extracted feature parameters, multi-scale temperature gradient features, and bubble density are combined, and a physics-based quantitative evaluation system is used to directly predict core performance losses such as solar reflectivity, mechanical properties, and salt spray corrosion resistance, realizing a direct correlation between defects and performance impact. At the same time, the efficient design of the dual-branch network architecture and feature fusion, along with the bubble density adaptive inference logic, ensures the speed of detection and performance evaluation, ultimately achieving rapid and accurate feedback on the impact of bubble defects on coating performance. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0021] Figure 1 This is a flowchart of a machine vision-based method for detecting bubble defects in a radiation-cooled coating according to an embodiment of this application.

[0022] Figure 2 This is a structural diagram of a machine vision-based radiation-cooled coating bubble defect detection system according to an embodiment of this application. Detailed Implementation

[0023] Figure 1 This is a flowchart of a machine vision-based method for detecting bubble defects in radiation-cooled coatings according to an embodiment of this application, as shown below. Figure 1 As shown, the machine vision-based method for detecting bubble defects in radiation-cooled coatings in this application includes:

[0024] S110, acquire image data of the radiation-cooled coating, the image data including visible light images, near-infrared images and infrared thermal imaging data;

[0025] S120, perform environmental adaptive preprocessing on the visible light image, the near-infrared image and the infrared thermal imaging data to correct the influence of light fluctuations and construct multi-scale temperature gradient features to obtain preprocessed image data;

[0026] S130, through a dual-branch deep learning network, spatial texture features and material composition features are extracted from the preprocessed image data. The attention module is combined to enhance the feature representation of the spatial texture features and the material composition features and perform feature fusion to obtain fused features and bubble density.

[0027] S140, Perform pixel-level segmentation of the fused features to obtain a bubble segmentation mask;

[0028] S150, extract multi-dimensional bubble feature parameters from the bubble segmentation mask, and combine the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density to predict the performance loss of the radiation cooling coating, so as to obtain the coating defect detection result.

[0029] Among them, multispectral imaging refers to acquiring images of multiple specific wavelengths, which can capture the characteristics of targets in different spectral ranges; multimodal imaging refers to integrating different types of imaging data (such as optical and thermal) to make up for the information limitations of single-modal data; the visible light band (400-700nm) can reflect the spatial texture and appearance characteristics of targets, the near-infrared band (700-1300nm) is sensitive to differences in material composition, and infrared thermal imaging can capture the temperature distribution and gradient changes of targets.

[0030] Image preprocessing is a key step in eliminating environmental interference and improving image quality. Environmental factors such as light fluctuations and temperature noise can cause image feature distortion and affect detection accuracy.

[0031] A dual-branch deep learning network is an architecture in deep learning that includes two parallel feature extraction branches that can process different types of input data separately. The features are then integrated through a fusion module to take into account the characteristics of different data and improve the model's feature representation ability.

[0032] Spatial texture features are the geometric structure and gray-level distribution patterns of the target in the image (such as contours and roughness), reflecting the appearance of the target; material composition features are the reflection and absorption characteristics of the target material in a specific spectral band, reflecting the composition and distribution of the material.

[0033] Attention modules are used in deep learning to enhance key features and suppress irrelevant information. They enable the network to focus on target regions or important feature channels, improving the model's recognition accuracy and efficiency. Types include channel attention, spatial attention, and group attention.

[0034] Feature fusion is the process of integrating features from different sources and of different types into a unified feature vector. The purpose is to combine the advantages of each feature to improve the model's decision-making ability. This includes weighted fusion, concatenation fusion, attention fusion, etc.

[0035] Density is the number of targets per unit area, reflecting the degree of target distribution density, and is a key parameter for describing defect distribution.

[0036] Pixel-level segmentation is the core task of deep learning semantic segmentation, which refers to classifying each pixel in an image into its corresponding target category (such as defect / background), and can accurately obtain the contour and location of the target.

[0037] A segmentation mask is the output of pixel-level segmentation. It is a binary image (or multi-valued image) with the same size as the original image, in which the target region (such as a defect) and the background region are distinguished by different pixel values ​​(such as 1 for the defect region and 0 for the background).

[0038] Multidimensional feature parameters are a set of parameters that describe the characteristics of a target from different perspectives. They can comprehensively characterize the target's physical, geometric, optical and other properties, and improve the accuracy of subsequent evaluation or prediction.

[0039] Performance loss prediction is based on a correlation model between defect characteristics and material properties. It calculates the degree of decline in key material properties through defect parameters, providing a scientific basis for quality control.

[0040] In some embodiments, acquiring image data of the radiation-cooled coating includes:

[0041] An industrial camera is used to image the radiation-cooled coating in the visible light band (400-700nm) to obtain the visible light image.

[0042] The near-infrared image is obtained by imaging the radiation-cooling coating in the 700-1300nm band using a near-infrared camera.

[0043] The surface temperature distribution of the radiation-cooling coating is collected using an infrared thermal imager to obtain the infrared thermal imaging data.

[0044] Among them, industrial cameras are the core imaging devices of machine vision inspection systems. They have the characteristics of high resolution, high frame rate, and strong anti-interference, and can stably acquire optical images of target objects in industrial environments. The visible light band of 400-700nm is the band that human vision can perceive. Imaging in this band can clearly reflect the appearance features of objects such as surface morphology and spatial texture.

[0045] Near-infrared cameras operate in the 700-1300nm near-infrared band. Light in this band can penetrate the shallow layers of an object's surface and is highly sensitive to differences in the chemical composition and internal structure of materials. It can capture material distribution anomalies that cannot be identified in the visible light band. Its imaging principle is based on the differences in the reflection and absorption characteristics of different materials to near-infrared light, and it is suitable for detecting hidden defects inside or under the surface of an object.

[0046] Near-infrared images are generated by near-infrared cameras by receiving near-infrared light reflected or radiated by a target object. Their grayscale values ​​are related to the material composition, water content, and internal structure of the object. Compared with visible light images, they can better reflect the differences in the object's internal characteristics and can be used to identify targets that have no obvious external abnormalities but have internal defects.

[0047] Infrared thermal imagers work based on the infrared radiation characteristics of objects. They detect the infrared radiation energy on the surface of an object and convert it into a temperature value, generating thermal imaging data that reflects the temperature distribution on the object's surface. Their core advantage is that they can quickly capture temperature differences on the surface of an object without contacting it. They are especially suitable for detecting abnormal heat conduction caused by internal defects (such as uneven local temperature distribution caused by low thermal conductivity media such as air and voids).

[0048] In this embodiment, by specifying the specific acquisition devices and corresponding bands for visible light, near-infrared, and infrared thermal imaging data, high-quality and highly synchronous multimodal raw data support is provided for the entire detection method: the combination of the 400-700nm visible light band and an industrial camera ensures clear capture of the geometric morphology of bubble defects; the combination of the 700-1300nm near-infrared band and a near-infrared camera enables accurate identification of abnormal material composition distribution caused by bubbles; and the acquisition of surface temperature distribution by an infrared thermal imager captures the difference in thermal conduction between bubbles and normal coatings. The targeted acquisition of these three types of data makes up for the shortcomings of traditional single-band, single-device acquisition of incomplete data information.

[0049] In some embodiments, the environmentally adaptive preprocessing of the visible light image, the near-infrared image, and the infrared thermal imaging data includes:

[0050] The light intensity L is obtained by measuring the light intensity of the detection environment in which the radiation-cooling coating is located using a photometric sensor.

[0051] Based on the light intensity L, the light compensation coefficient is obtained, wherein the light compensation coefficient is expressed as follows:

[0052]

[0053] In the formula, This is the illumination compensation coefficient. This is the adjustment coefficient; Light intensity; This represents the average of historical illumination data.

[0054] Based on the illumination compensation coefficient, the illumination intensity to be compensated in the visible light image is compensated to obtain the compensated illumination intensity, wherein the compensated illumination intensity is expressed as follows:

[0055]

[0056] In the formula, To compensate for the light intensity, The light intensity to be compensated;

[0057] Temperature gradient calculations are performed on the infrared thermal imaging data to obtain the temperature gradient of the bubble region, wherein the temperature gradient of the bubble region is represented as follows:

[0058]

[0059] In the formula, The temperature gradient in the bubble region. The partial derivative of temperature in the x-direction. This is the temperature partial derivative in the y-direction;

[0060] The temperature gradient in the bubble region is subjected to multi-scale fusion processing to obtain multi-scale temperature gradient features, wherein the multi-scale temperature gradient features are represented as follows:

[0061]

[0062] In the formula, It features multi-scale temperature gradient characteristics. For scale indexing, The total number of scales. Let be the weight coefficient for the s-th scale. Let be the temperature gradient after Gaussian filtering at the s-th scale.

[0063] Among them, a photometric sensor is an electronic device used to measure the intensity of ambient light. Based on the photoelectric effect, it converts light signals into electrical signals and has the characteristics of fast response speed, high measurement accuracy, and strong anti-interference ability. It is widely used in industrial inspection, automatic control and other scenarios, and can provide real-time feedback on the dynamic changes of ambient light.

[0064] The illumination compensation coefficient is a parameter used to correct the impact of ambient light fluctuations on image quality. It uses a mathematical model to link real-time illumination with standard illumination conditions, ensuring that the image maintains consistent brightness and contrast under different lighting environments. It is a key parameter for eliminating environmental interference in image preprocessing.

[0065] The adjustment coefficient is a constant used in a mathematical model to control the rate of change or degree of influence of parameters. It is determined through experimental calibration and can be used to adjust the response characteristics of the model according to the needs of actual application scenarios, so that the model output is more in line with actual needs.

[0066] Historical data mean is a value obtained by statistically averaging monitoring data under similar scenarios in the past. It can be used as a reference benchmark for current data to eliminate random fluctuations in data and improve the robustness and stability of the model.

[0067] The light intensity to be compensated is the brightness value of the image under the original lighting environment. Affected by fluctuations in ambient light, it may have problems such as uneven brightness and insufficient contrast. The compensated light intensity is the brightness value corrected by the light compensation coefficient, which can eliminate ambient light interference and restore the true appearance features of the target object.

[0068] Temperature gradient is a physical quantity that describes the rate of temperature change in space. It reflects the difference in temperature distribution on the surface of an object. The larger the value, the more drastic the temperature change. For objects with internal defects (such as voids and bubbles), a significant temperature gradient will be formed due to the difference in thermal conductivity between the defects and the base material.

[0069] Partial derivatives are the rates of change of a certain independent variable in a multivariable function. The partial derivative of temperature in the x-direction reflects the rate of temperature change along the horizontal direction, while the partial derivative of temperature in the y-direction reflects the rate of temperature change along the vertical direction. The combination of the two can comprehensively describe the temperature distribution gradient in a two-dimensional plane.

[0070] Multi-scale fusion processing is a technique that integrates feature data at different scales with weights. By assigning corresponding weight coefficients to features at different scales, it takes into account both the global information of large-scale features and the local details of small-scale features, avoids missing key information in single-scale features, and improves the representation ability of features.

[0071] The scale index is an identifier that distinguishes features at different scales, the total number of scales is the number of scales involved in the fusion, and the weight coefficient is a parameter used to assign the importance of features at different scales. Together, these three constitute the core elements of multi-scale fusion, ensuring that the fused features can fully cover targets of different sizes.

[0072] Gaussian filtering is an image smoothing technique that uses a Gaussian function to apply a weighted average to the data, effectively suppressing noise interference while preserving the overall contour of the target features. The temperature gradient after Gaussian filtering is smoothed temperature change rate data, which has higher stability and reliability.

[0073] In this embodiment, an environment-adaptive preprocessing scheme constructed using illumination compensation and multi-scale temperature gradient features effectively addresses the shortcomings of traditional preprocessing, such as ignoring environmental interference and missing key information due to single-scale features. A photometric sensor collects illumination data in real time, and the illumination compensation coefficient is calculated by combining historical illumination averages and adjustment coefficients. This accurately corrects the impact of ambient illumination fluctuations on visible light images, ensuring that the spatial texture features of bubble defects remain stable and discernible under different illumination environments. The temperature gradient in the bubble region is calculated using two-dimensional temperature partial derivatives, capturing the thermal differences between bubbles and normal coatings. Multi-scale temperature gradient features are then obtained through multi-scale Gaussian filtering and fusion, taking into account the thermal characteristics of bubbles of different sizes and enhancing the distinguishability between micro and large bubbles. The entire preprocessing process adapts to dynamic environmental changes, eliminating illumination noise interference and enriching the dimensions of temperature features. This provides high-quality, highly stable input data for subsequent feature extraction by a dual-branch deep learning network, directly improving the overall detection method's anti-interference capability and defect recognition accuracy.

[0074] In some implementations, the step of extracting spatial texture features and material composition features from the preprocessed image data, and enhancing the feature representation of the spatial texture features and material composition features using an attention module, includes:

[0075] The compensated light intensity Input the residual network to perform feature extraction, and obtain the spatial texture features. ;

[0076] The near-infrared image Input the residual network to perform feature extraction, and obtain the material composition features. ;

[0077] The channel attention weight matrix is ​​generated through the channel attention submodule. The spatial attention weight matrix is ​​generated through the spatial attention submodule. ;

[0078] According to the channel attention weight matrix The spatial attention submodule generates a spatial attention weight matrix. For the spatial texture features and the material composition characteristics Feature enhancement processing is performed separately to obtain enhanced spatial texture features. and the compositional characteristics of the reinforced material .

[0079] Among them, residual networks are a classic architecture in the field of deep learning. The core of the architecture is to pass the input directly to the output of the subsequent layers through residual connections (skip connections), which effectively solves the gradient vanishing problem in the training of deep networks and allows the network to maintain training stability and feature extraction ability while increasing the number of layers. Its essence is to learn the residual (difference) between input and output, which can capture the key features of the target more efficiently and is widely used in computer vision tasks such as image classification and object detection.

[0080] The channel attention submodule is an important branch of the attention mechanism, focusing on optimizing the channel dimension of the feature map. By learning the importance weights of different feature channels, it strengthens the channel responses that are useful for target recognition and suppresses noise interference from irrelevant channels. Its typical implementation process is as follows: global average pooling and max pooling are performed on the feature map, the dimension is compressed and the weights are learned through a multilayer perceptron (MLP), and finally the channel attention weight matrix is ​​generated.

[0081] The spatial attention submodule focuses on optimizing the spatial dimension of the feature map. By learning the importance weights of different spatial locations, it strengthens the feature response of the target area and suppresses invalid information in the background area. Its typical implementation process is as follows: the feature map is subjected to average pooling and max pooling along the channel dimension, and after concatenation, a spatial attention weight map is generated through convolution operation to achieve accurate focusing on the spatial location of the target.

[0082] The channel attention weight matrix is ​​the output of the channel attention submodule. Its dimension is the same as the number of channels in the input feature map. Each element corresponds to the importance coefficient of a feature channel (within the range of [0,1]). The larger the coefficient, the higher the contribution of the channel to target recognition.

[0083] The spatial attention weight matrix is ​​the output of the spatial attention submodule. Its size is consistent with the spatial size (height × width) of the input feature map. Each pixel position corresponds to an importance coefficient (value range [0,1]). The larger the coefficient, the higher the probability that there is a target at that spatial position.

[0084] The enhanced features are the product of the original features weighted by channel attention and spatial attention. They integrate key information from the channel dimension and target localization information from the spatial dimension. Compared with the original features, they have the advantages of lower noise, more prominent target features, and stronger representation ability.

[0085] In this embodiment, the residual network ensures the deep extraction and integrity of spatial texture features and material composition features through residual connections, avoiding the loss of key bubble information caused by gradient vanishing in deep networks. The channel attention submodule strengthens key channel features related to bubble defects by learning channel weights and suppresses noise interference from irrelevant channels. The spatial attention submodule accurately locates the spatial position of bubble defects by learning spatial weights and weakens background interference in normal coating areas. The two types of enhanced features generated in the end not only retain the core features of the bubble (spatial texture and abnormal material composition) but also significantly improve the distinguishability between features and interference, providing high-quality and highly targeted input data for subsequent feature fusion and bubble segmentation, directly supporting the accuracy of the overall detection method.

[0086] In some implementations, the feature fusion includes:

[0087] Spatial texture features The visible light branch feature activation degree is calculated, and the visible light branch feature activation degree is represented as follows:

[0088]

[0089] In the formula, Avis represents the activation degree of the visible light branch feature. The length of the feature vector. It is the Sigmoid activation function. For the transpose of the learnable weight vector, The i-th element of the spatial texture feature;

[0090] Material composition characteristics Near-infrared branch feature activations are calculated to obtain the near-infrared branch feature activations, which are represented as follows:

[0091]

[0092] In the formula, Near-infrared branch feature activation, The i-th element is a characteristic of the material composition.

[0093] Based on the activation values ​​of the visible light branch and the near-infrared branch, dynamic weights are assigned to obtain the dynamic weights α and β of the visible light branch and the near-infrared branch, respectively. The dynamic weight assignment is represented as follows:

[0094]

[0095] Spatial texture features Material composition characteristics Channel attention fusion is performed to obtain channel attention fusion features, which are represented as follows:

[0096]

[0097] In the formula, For channel attention fusion features, Output weights for the convolutional block attention module in the visible light branch. Output weights for the grouped sparse attention module of the near-infrared branch;

[0098] Spatial texture features Material composition characteristics Spatial attention fusion is performed to obtain spatial attention fusion features, which are represented as follows:

[0099]

[0100] In the formula, For spatial attention fusion features, Output weights for the convolutional block attention module of the near-infrared branch. Output weights for the grouped sparse attention module of the visible light branch;

[0101] The channel attention fusion features and spatial attention fusion features are fused to obtain the final fused features, which are represented as follows:

[0102]

[0103] In the formula, For the final fusion feature, These are the weighting coefficients for thermal imaging features. Features of infrared thermal imaging;

[0104] Based on the final fusion features, the total number of bubbles n and the actual physical area S within the detection region are calculated. Then, the bubble density is determined based on these two parameters, where the bubble density is expressed as follows:

[0105]

[0106] In the formula, Bubble density, The total number of bubbles in the detection area. This represents the actual physical area of ​​the detection zone.

[0107] Feature activation is an indicator that measures the contribution of a feature map to target recognition. It quantifies the feature response intensity to reflect the current feature’s ability to identify the target. The value range is usually [0,1]. The higher the value, the stronger the correlation between the feature and the target.

[0108] The Sigmoid activation function is a non-linear activation function in deep learning, expressed as σ(x)=1 / (1+exp(-x)). It can map any real number input to the interval [0,1] and is often used in scenarios such as probability prediction and feature response normalization. It can effectively compress the range of feature values ​​and improve computational stability.

[0109] The learnable weight vector is a parameter vector that is continuously optimized during the training of a deep learning model. Its transpose (w^T) is used to perform an inner product operation with the feature vector to achieve weighted filtering of features, highlighting feature elements that are relevant to the target and suppressing irrelevant elements.

[0110] The feature vector length is the number of elements in the one-dimensional vector obtained after flattening the feature map. It reflects the dimensionality of the feature and its value is determined by the number of channels, height, and width of the feature map (N = number of channels × height × width).

[0111] Dynamic weight allocation is a strategy that adaptively adjusts the contribution ratio of different branches based on the real-time feature quality. Compared with fixed weights, it can more flexibly adapt to feature differences in different scenarios and improve the representational ability of fused features.

[0112] Channel attention fusion is a feature fusion method that combines channel attention weights. By assigning channel-level weights to features of different branches, it strengthens the synergistic effect of key channels and generates fused features dominated by a certain branch.

[0113] Spatial attention fusion is a feature fusion method that combines spatial attention weights. By assigning spatial weights to features of different branches, it strengthens the collaborative response of the target spatial location and generates a fusion feature dominated by another branch.

[0114] The weighting coefficient is a constant used to balance the contribution of multimodal features. It is determined through experimental calibration to ensure that information imbalance will not occur due to differences in numerical values ​​or importance of different modal features during fusion.

[0115] Infrared thermal imaging features are temperature distribution-related features extracted from infrared thermal imaging data. They reflect the difference in heat conduction between the target and the background and can serve as a supplement to optical features, improving the target recognition capability in complex scenes.

[0116] The final fusion feature is the core feature vector after integrating multi-branch and multi-modal features. It combines the advantages of each original feature and has a more comprehensive target representation capability. It is a key input for subsequent model inference (such as segmentation and classification).

[0117] The total number of detection areas is a statistical count of the number of defects within the target area, while the actual physical area of ​​the detection area is the true spatial size of the target detection range (which needs to be converted in conjunction with the spatial resolution of the equipment). Both are used together to calculate the defect density, reflecting the degree of density of defect distribution.

[0118] In this embodiment, the collaborative design of feature activation calculation, dynamic weight allocation, multimodal fusion, and bubble density statistics effectively solves the shortcomings of fixed weights, imbalance of multimodal information, and insufficient quantification of defect distribution in traditional feature fusion: feature activation accurately quantifies the recognition degree of bubble defects by dual-branch features through the Sigmoid function and learnable weight vector; dynamic weight allocation adaptively adjusts the branch contribution ratio according to activation, avoiding the problem of insufficient adaptability of fixed weights in different scenarios; channel attention fusion and spatial attention fusion are dominated by two types of core features, which not only retain their respective advantages but also achieve complementarity. After superimposing thermal imaging features, the feature dimensions are further enriched, so that the final fused features have a more comprehensive bubble representation capability; bubble density statistics provide key quantitative indicators for subsequent adaptive inference strategies and performance loss prediction. At the same time, the entire fusion process ensures computational stability through operations such as Sigmoid activation and weight normalization, which greatly improves the accuracy and efficiency of subsequent bubble segmentation.

[0119] In some embodiments, performing pixel-level segmentation of the fused features to obtain a bubble segmentation mask includes:

[0120] Fusion features Input the feature pyramid network, replace the third to fifth convolutional layers of the feature pyramid network with trident modules, and perform feature processing using dilated convolutions with dilation rates d=2, 4, and 8 to obtain multi-scale features. The dilated convolution is represented as follows:

[0121]

[0122] In the formula, For dilated convolution at pixels The output at that location, For fusion features In pixels Pixel value at that location, The kernel size is [size]. For the convolution kernel at position The weight of the position, For expansion rate, The coordinates of the sampled pixels from the dilated convolution;

[0123] Candidate bubble regions are obtained by aligning regions of interest with multi-scale features to extract candidate region features.

[0124] Calculate the centroid coordinates of the candidate bubble region, where the centroid coordinates are represented as follows:

[0125]

[0126] In the formula, Using the centroid coordinates, Let's say the pixel area of ​​the bubble. These are the pixel values ​​of the bubble mask;

[0127] Calculate the radial distance between the centroid coordinates and the pixels within the candidate bubble region to obtain the radial distance from the pixel to the centroid. The radial distance from the pixel to the centroid is represented as follows:

[0128]

[0129] In the formula, For pixels to centroid coordinates The radial distance;

[0130] A constraint loss is constructed for the radial distance, resulting in a radial distance correction loss, which is expressed as follows:

[0131]

[0132] In the formula, For radial distance correction loss, Let be the expected function. Let the bubble's equivalent radius be 1. ;

[0133] Based on the radial distance correction loss, the masking loss and radial distance correction loss of the standard masked region convolutional neural network are weighted and summed to obtain the total loss for bubble segmentation. The total loss for bubble segmentation is expressed as follows:

[0134]

[0135] In the formula, The total loss for bubble segmentation, For mask loss, The weighting coefficients for radial distance correction loss;

[0136] The candidate bubble region is segmented and optimized based on the total loss of bubble segmentation to obtain the bubble segmentation mask.

[0137] Among them, the feature pyramid network is a classic architecture in the field of object detection and semantic segmentation. By fusing features from different levels in the convolutional neural network (shallow high-resolution features and deep high-semantic features), it constructs a multi-scale feature pyramid to achieve accurate identification and segmentation of targets of different sizes. Its core advantage is that it takes into account both the local details and global semantic information of the target.

[0138] The Trident module is a multi-branch feature extraction structure that expands the receptive field of the network without increasing the computational load by using dilated convolutions with different dilation rates in different branches, while maintaining the feature map resolution, thus adapting to the needs of target detection and segmentation of different sizes.

[0139] Dilated convolution expands the receptive field without changing the feature map size by introducing holes (zero padding) into the convolution kernel.

[0140] Centroid coordinates are the geometric center of an object. For a target in an image, the core spatial location of the target is reflected by a weighted average of the pixel coordinates (with the weight being the pixel value). This is the basis for distance calculation and shape analysis.

[0141] Radial distance is the straight-line distance from any pixel in an image to the centroid of the target. It is calculated by taking the square root of the sum of the squares of the coordinate differences and is used to describe the spatial positional relationship of a pixel relative to the centroid.

[0142] Correction loss is a loss function that constrains the consistency between the model's prediction results and the actual physical parameters. By quantifying the difference between the predicted and actual values, it guides the model to optimize its parameters and improve prediction accuracy.

[0143] The weighting coefficients are hyperparameters that balance the contributions of multiple loss terms. They are determined through experimental calibration to ensure that different loss terms play a reasonable role in model training and to avoid a single loss term dominating the training process.

[0144] In this embodiment, through the collaborative design of feature pyramid network optimization, multi-diffraction dilated convolution, accurate candidate region extraction, and radial distance correction loss, the technical challenges of inaccurate recognition of overlapping and microbubble contours and insufficient segmentation integrity in traditional pixel-level segmentation are effectively solved: the feature pyramid network significantly expands the receptive field of the network without reducing feature resolution, achieving comprehensive capture of features of bubbles of different sizes; region of interest alignment eliminates quantization errors in candidate region extraction, ensuring accurate mapping of bubble features; the calculation of centroid coordinates and radial distance provides a geometric basis for distinguishing the contours of overlapping bubbles; the radial distance correction loss forces the model to learn the true geometric shape of bubbles by constraining the difference between the predicted value and the true equivalent radius, avoiding misjudgment of pixels in overlapping areas; the total loss function balances the mask loss and correction loss through the weight coefficient λ, taking into account both the basic accuracy of segmentation and the adaptability to special scenarios (overlapping, microbubbles).

[0145] In some implementations, extracting multi-dimensional bubble feature parameters from the bubble segmentation mask includes:

[0146] The reflectance of the defective and non-defective regions corresponding to the bubble segmentation mask is extracted to obtain the reflectance difference in the visible light band.

[0147] Near-infrared reflectance is extracted from the defective and non-defective regions corresponding to the bubble segmentation mask to obtain the near-infrared band reflectance difference.

[0148] Pixel statistics are performed on the bubble segmentation mask to obtain the bubble pixel area;

[0149] The bubble volume is calculated by combining the bubble pixel area, coating thickness, and spatial resolution, and the equivalent diameter of the bubble is obtained from the bubble volume.

[0150] Based on the bubble pixel area, the equivalent circumference is calculated. Edge detection is performed on the bubble segmentation mask to extract edge pixels and calculate the circumference, thus obtaining the actual circumference of the bubble. The roundness of the bubble is then obtained based on the actual circumference.

[0151] Among them, the reflectivity difference refers to the relative change in reflectivity between the target defect area and the defect-free area in a specific wavelength band. It is the core indicator for quantifying the degree of damage to the optical properties of materials. Its positive or negative value and magnitude directly reflect the impact of defects on light reflection ability (negative numbers indicate a decrease in reflectivity).

[0152] Near-infrared reflectivity is closely related to material composition and internal structure. Its differences can characterize the uneven material distribution or internal structural damage caused by defects. Compared with the visible light band, it can better reflect the impact of latent defects on the material's essence.

[0153] The bubble pixel area is the total number of pixels in the bubble defect region of the segmentation mask. It is the basis for converting image features into physical features. Its value, combined with the spatial resolution of the imaging device, can be converted into the actual physical area of ​​the bubble.

[0154] Bubble volume is a core parameter characterizing the size of the space occupied by the bubble. For bubbles inside the coating, it can be estimated based on pixel area, coating thickness and spatial resolution (assuming that the bubbles are uniformly distributed in the coating thickness direction, approximately columnar or spherical).

[0155] The equivalent diameter is the diameter of an irregularly shaped bubble after it has been converted into an ideal sphere. It is used to standardize and quantify the size of bubbles, avoiding confusion in size characterization caused by differences in bubble shape. It is a standardized geometric parameter in industrial testing.

[0156] The equivalent circumference is the ideal circumference of a circle calculated based on the actual physical area of ​​the bubble. It serves as a benchmark for measuring the degree of deviation of the bubble shape, and its core is the theoretical circumference when the bubble is assumed to be a perfect circle.

[0157] The actual perimeter of the bubble is obtained by extracting the edge pixels of the bubble using an edge detection algorithm and then converting the actual contour length according to the spatial resolution, reflecting the actual shape characteristics of the bubble.

[0158] Roundness is a parameter that characterizes the degree to which the shape of an object deviates from an ideal circle. Its value range is usually [0,1). The larger the value, the more irregular the shape. Its core is to quantify the shape characteristics by the difference between the actual circumference and the equivalent circle circumference.

[0159] In this embodiment, by extracting multi-dimensional and comprehensive bubble feature parameters from a precise bubble segmentation mask, the shortcomings of traditional feature extraction—such as single parameters, insufficient quantification, and low correlation with coating performance—are effectively addressed. The difference in reflectance between visible and near-infrared bands quantifies the impact of bubbles from the perspectives of optical properties and material composition, providing a direct basis for predicting solar reflectance loss. Bubble pixel area, volume, and equivalent diameter construct a complete size quantification system, unifying the size representation of bubbles of different shapes. Equivalent circumference, actual circumference, and roundness supplement shape features, reflecting the potential impact of bubble irregularities on coating performance. All feature parameters are extracted based on a precise segmentation mask, ensuring the accuracy and reliability of the data.

[0160] In some embodiments, predicting the performance loss of the radiation-cooling coating by combining the multi-dimensional bubble characteristic parameters, the multi-scale temperature gradient characteristics, and the bubble density includes:

[0161] The solar reflectance loss is obtained based on the solar reflectance of the defect-free coating, the total area of ​​the inspection area, the actual total physical area of ​​all bubbles, and the percentage decrease in reflectance in the visible light band. The solar reflectance loss is expressed as follows:

[0162]

[0163] In the formula, For the loss of solar reflectivity, For a defect-free coating, solar reflectivity The total area of ​​the detection zone. This represents the actual total physical area of ​​all bubbles. This represents the percentage decrease in reflectivity in the visible light band.

[0164] Based on the mechanical strength of the defect-free coating, material sensitivity coefficient, defect distribution index, and bubble density, the mechanical performance loss is obtained, whereby the mechanical performance loss is expressed as follows:

[0165]

[0166] In the formula, For mechanical property loss, denoted as σ0, where σ0 is the mechanical strength of the defect-free coating, m is the material sensitivity coefficient, and p is the defect distribution index.

[0167] Based on the salt spray resistance time, corrosion rate coefficient, bubble density, ambient temperature, and salt spray concentration of the defect-free coating, the loss of salt spray corrosion resistance is obtained. The loss of salt spray corrosion resistance is expressed as follows:

[0168]

[0169] In the formula, To compensate for the loss of salt spray corrosion resistance, Salt spray resistance time for defect-free coatings The corrosion rate coefficient is... For ambient temperature, This refers to the salt spray concentration.

[0170] Among them, solar reflectivity is the core optical performance indicator of radiation-cooled coatings, specifically referring to the coating's ability to reflect solar radiation in the 0.3-2.5μm band. The value range is usually [0.85, 0.98]. The higher the value, the better the cooling effect. It is a key benchmark parameter for measuring the basic performance of the coating.

[0171] The total area of ​​the detection area refers to the actual physical size of the coating area used for defect detection and performance evaluation. It needs to be calculated by combining the spatial resolution of the imaging equipment and the area of ​​image pixels. It is a basic parameter for quantifying the proportion of defects.

[0172] The total defect area is the sum of the physical areas of all defects within the detection area. It reflects the overall coverage of defects and is a key intermediate parameter that correlates defect size with performance loss.

[0173] The percentage decrease in reflectivity is a quantitative indicator of the degree of degradation in the optical performance of a coating caused by defects. It reflects the extent to which defects impair the reflectivity of light in a specific wavelength band and is directly related to the refractive index and distribution density of the defects.

[0174] Solar reflectance loss is the amount of degradation in the core optical performance of the coating caused by defects. It directly determines the degree of attenuation of the cooling effect and is a key output indicator for evaluating coating quality.

[0175] Mechanical strength is the ability of a coating to resist external forces such as tension and impact, including tensile strength and impact strength. The unit is usually MPa. It is a core mechanical indicator for evaluating the structural stability and service life of a coating.

[0176] The material sensitivity coefficient is an empirical parameter characterizing the sensitivity of a material to defects. It is determined by the molecular structure and micromorphology of the material itself, and the value varies for different polymers, ceramics and other materials. It is usually calibrated through a large number of defect-strength correlation experiments.

[0177] The defect distribution index is a parameter that describes the spatial distribution characteristics of defects and reflects the uniformity of defects in the coating. Its value range is usually [0.5, 1.2]. When the distribution is uniform, p≈0.8, and when the distribution is clustered, p>1.0.

[0178] Mechanical property loss is the amount of mechanical property reduction in the coating caused by defects. It directly affects the coating's wear resistance and impact resistance and is a key indicator for evaluating the reliability of the coating in actual use. The unit is consistent with mechanical strength (MPa).

[0179] Salt spray resistance time is an indicator of a coating's ability to resist salt spray corrosion. It is determined by salt spray testing (standards such as GB / T10125) and is measured in hours (h). The longer the time, the better the corrosion resistance. It is a core parameter for evaluating the outdoor service life of a coating.

[0180] The corrosion rate coefficient is an empirical parameter characterizing the corrosion rate of a coating in a salt spray environment. It is related to the coating material, substrate type, and salt spray concentration, and is obtained by fitting salt spray test data. The unit is h⁻¹・cm².

[0181] Ambient temperature is a key environmental factor affecting the corrosion reaction rate. Increased temperature will accelerate the electrochemical corrosion process. The standard temperature for salt spray testing is usually 35℃, but it needs to be adapted to the ambient temperature of different regions in practical applications.

[0182] Salt spray concentration refers to the mass fraction of NaCl in salt spray, which is a core parameter affecting corrosion intensity. The standard salt spray test concentration is 5%, and the concentration may be higher (5%-10%) in marine environments, coastal industrial areas and other scenarios.

[0183] Salt spray corrosion resistance loss is the amount of decrease in the coating's corrosion resistance caused by defects. It directly determines the coating's service life in corrosive environments and is a key output indicator for evaluating the coating's outdoor reliability. The unit is hours (h).

[0184] In this embodiment, by constructing three quantitative prediction models—solar reflectivity loss, mechanical performance loss, and salt spray corrosion resistance performance loss—the technical challenge of traditional detection methods being unable to quantitatively assess the impact of bubble defects on the core performance of the coating is effectively solved: the solar reflectivity loss model integrates the bubble area ratio and reflectivity attenuation ratio to accurately quantify the impact of optical performance attenuation on the cooling effect; the mechanical performance loss model, based on the Weibull intensity distribution, establishes a quantitative correlation between bubble density and mechanical strength through the material sensitivity coefficient and defect distribution index; and the salt spray corrosion resistance performance loss model incorporates practical application factors such as ambient temperature and salt spray concentration to comprehensively consider the destructive mechanism of bubbles on corrosion resistance.

[0185] This invention discloses a machine vision-based method for detecting bubble defects in radiation-cooled coatings. First, it comprehensively captures the optical reflection characteristics, material composition differences, and thermal distribution features of bubble defects through multimodal data acquisition using visible light, near-infrared, and infrared thermal imaging. This overcomes the limitations of traditional single-band imaging, which struggles to cover the multidimensional influences of bubbles, laying a data foundation for accurately linking defects to performance. Second, environmentally adaptive preprocessing corrects for illumination fluctuation interference and constructs multi-scale temperature gradient features, avoiding feature distortion caused by environmental factors. Simultaneously, a dual-branch deep learning network specifically extracts spatial texture and material composition features, and an attention module enhances the expression of key bubble features. A dynamic weight fusion strategy achieves high feature density. The efficient and precise integration of bubble density calculations provides a basis for subsequent efficient inference. Furthermore, pixel-level segmentation accurately obtains bubble contour masks through improved algorithms, ensuring the accuracy of multi-dimensional bubble feature parameter extraction. Finally, these precisely extracted feature parameters, multi-scale temperature gradient features, and bubble density are combined to directly predict core performance losses such as solar reflectivity, mechanical properties, and salt spray corrosion resistance through a physics-based quantitative evaluation system. This achieves a direct correlation between defects and performance impact. Meanwhile, the efficient design of the dual-branch network architecture and feature fusion, along with the adaptive inference logic for bubble density, ensures the speed of detection and performance evaluation, ultimately achieving rapid and accurate feedback on the impact of bubble defects on coating performance.

[0186] like Figure 2 As shown, this application also provides a machine vision-based system for detecting bubble defects in radiation-cooled coatings, comprising:

[0187] An acquisition unit is used to acquire image data of the radiation-cooled coating, the image data including visible light images, near-infrared images and infrared thermal imaging data;

[0188] The preprocessing unit is used to perform environmentally adaptive preprocessing on the visible light image, the near-infrared image, and the infrared thermal imaging data to correct the effects of illumination fluctuations and construct multi-scale temperature gradient features to obtain preprocessed image data.

[0189] The feature extraction unit is used to extract spatial texture features and material composition features from the preprocessed image data through a dual-branch deep learning network, and to enhance the feature representation of the spatial texture features and material composition features by combining an attention module and performing feature fusion to obtain fused features and bubble density.

[0190] A segmentation unit is used to perform pixel-level segmentation of the fused features to obtain a bubble segmentation mask;

[0191] The prediction unit is used to extract multi-dimensional bubble feature parameters from the bubble segmentation mask, and combine the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density to predict the performance loss of the radiation-cooled coating, so as to obtain the coating defect detection result.

[0192] This invention discloses a machine vision-based bubble defect detection system for radiation-cooled coatings. Firstly, it comprehensively captures the optical reflection characteristics, material composition differences, and thermal distribution features of bubble defects through multimodal data acquisition using visible light, near-infrared, and infrared thermal imaging. This overcomes the limitations of traditional single-band imaging, which struggles to cover the multidimensional influences of bubbles, laying a data foundation for accurately linking defects to performance. Secondly, environmentally adaptive preprocessing corrects for illumination fluctuation interference and constructs multi-scale temperature gradient features, avoiding feature distortion caused by environmental factors. Simultaneously, a dual-branch deep learning network specifically extracts spatial texture and material composition features, and an attention module enhances the expression of key bubble features. A dynamic weight fusion strategy achieves high feature accuracy. The efficient and precise integration of bubble density calculations provides a basis for subsequent efficient inference. Furthermore, pixel-level segmentation accurately obtains bubble contour masks through improved algorithms, ensuring the accuracy of multi-dimensional bubble feature parameter extraction. Finally, these precisely extracted feature parameters, multi-scale temperature gradient features, and bubble density are combined to directly predict core performance losses such as solar reflectivity, mechanical properties, and salt spray corrosion resistance through a physics-based quantitative evaluation system. This achieves a direct correlation between defects and performance impact. Meanwhile, the efficient design of the dual-branch network architecture and feature fusion, along with the adaptive inference logic for bubble density, ensures the speed of detection and performance evaluation, ultimately achieving rapid and accurate feedback on the impact of bubble defects on coating performance.

Claims

1. A machine vision-based method for detecting bubble defects in radiation-cooled coatings, characterized in that, Including the following steps: Acquire image data of the radiation-cooled coating, the image data including visible light images, near-infrared images, and infrared thermal imaging data; The visible light image, the near-infrared image, and the infrared thermal imaging data are subjected to environment-adaptive preprocessing to correct the effects of illumination fluctuations and construct multi-scale temperature gradient features, thereby obtaining preprocessed image data. A dual-branch deep learning network is used to extract spatial texture features and material composition features from the preprocessed image data. An attention module is then used to enhance the feature representation of the spatial texture features and the material composition features and to perform feature fusion to obtain fused features and bubble density. The fused features are segmented at the pixel level to obtain a bubble segmentation mask; Multi-dimensional bubble feature parameters are extracted from the bubble segmentation mask. By combining the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density, the performance loss of the radiation-cooled coating is predicted to obtain the coating defect detection result.

2. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 1, characterized in that, The acquisition of image data of the radiation-cooled coating includes: An industrial camera is used to image the radiation-cooled coating in the visible light band (400-700nm) to obtain the visible light image. The near-infrared image is obtained by imaging the radiation-cooling coating in the 700-1300nm band using a near-infrared camera. The surface temperature distribution of the radiation-cooling coating is collected using an infrared thermal imager to obtain the infrared thermal imaging data.

3. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 1, characterized in that, The environmental adaptive preprocessing of the visible light image, the near-infrared image, and the infrared thermal imaging data includes: The light intensity L is obtained by measuring the light intensity of the detection environment in which the radiation-cooling coating is located using a photometric sensor. Based on the light intensity L, the light compensation coefficient is obtained, wherein the light compensation coefficient is expressed as follows: In the formula, This is the illumination compensation coefficient. This is the adjustment coefficient; Light intensity; This represents the average of historical illumination data. Based on the illumination compensation coefficient, the illumination intensity to be compensated in the visible light image is compensated to obtain the compensated illumination intensity, wherein the compensated illumination intensity is expressed as follows: In the formula, To compensate for the light intensity, The light intensity to be compensated; Temperature gradient calculations are performed on the infrared thermal imaging data to obtain the temperature gradient of the bubble region, wherein the temperature gradient of the bubble region is represented as follows: In the formula, The temperature gradient in the bubble region. The partial derivative of temperature in the x-direction. This is the temperature partial derivative in the y-direction; The temperature gradient in the bubble region is subjected to multi-scale fusion processing to obtain multi-scale temperature gradient features, wherein the multi-scale temperature gradient features are represented as follows: In the formula, It features multi-scale temperature gradient characteristics. For scale indexing, The total number of scales. Let be the weight coefficient for the s-th scale. Let be the temperature gradient after Gaussian filtering at the s-th scale.

4. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 3, characterized in that, The step of extracting spatial texture features and material composition features from the preprocessed image data, and enhancing the feature representation of the spatial texture features and material composition features using an attention module, includes: The compensated light intensity Input the residual network to perform feature extraction, and obtain the spatial texture features. ; The near-infrared image Input the residual network to perform feature extraction, and obtain the material composition features. ; The channel attention weight matrix is ​​generated through the channel attention submodule. The spatial attention weight matrix is ​​generated through the spatial attention submodule. ; According to the channel attention weight matrix The spatial attention submodule generates a spatial attention weight matrix. For the spatial texture features and the material composition characteristics Feature enhancement processing is performed separately to obtain enhanced spatial texture features. and the compositional characteristics of the reinforced material .

5. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 1, characterized in that, The feature fusion process includes: Spatial texture features The visible light branch feature activation degree is calculated, and the visible light branch feature activation degree is represented as follows: In the formula, Avis represents the activation degree of the visible light branch feature. The length of the feature vector. It is the Sigmoid activation function. For the transpose of the learnable weight vector, Let i be the i-th element of the spatial texture feature; Material composition characteristics Near-infrared branch feature activations are calculated to obtain the near-infrared branch feature activations, which are represented as follows: In the formula, Near-infrared branch feature activation, The i-th element is a characteristic of the material composition. Based on the activation values ​​of the visible light branch and the near-infrared branch, dynamic weights are assigned to obtain the dynamic weights α and β of the visible light branch and the near-infrared branch, respectively. The dynamic weight assignment is represented as follows: Spatial texture features Material composition characteristics Channel attention fusion is performed to obtain channel attention fusion features, which are represented as follows: In the formula, For channel attention fusion features, Output weights for the convolutional block attention module in the visible light branch. Output weights for the grouped sparse attention module of the near-infrared branch; Spatial texture features Material composition characteristics Spatial attention fusion is performed to obtain spatial attention fusion features, which are represented as follows: In the formula, For spatial attention fusion features, Output weights for the convolutional block attention module of the near-infrared branch. Output weights for the grouped sparse attention module of the visible light branch; The channel attention fusion features and spatial attention fusion features are fused to obtain the final fused features, which are represented as follows: In the formula, For the final fusion features, These are the weighting coefficients for thermal imaging features. Features of infrared thermal imaging; Based on the final fusion features, the total number of bubbles n and the actual physical area S within the detection region are calculated. Then, the bubble density is determined based on these two parameters, where the bubble density is expressed as follows: In the formula, Bubble density, The total number of bubbles in the detection area. This represents the actual physical area of ​​the detection zone.

6. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 1, characterized in that, The step of performing pixel-level segmentation of the fused features to obtain a bubble segmentation mask includes: Fusion features Input the feature pyramid network, replace the third to fifth convolutional layers of the feature pyramid network with trident modules, and perform feature processing using dilated convolutions with dilation rates d=2, 4, and 8 to obtain multi-scale features. The dilated convolution is represented as follows: In the formula, For dilated convolution at pixels The output at that location, For fusion features In pixels Pixel value at that location, The kernel size is [size]. For the convolution kernel at position The weight of the position, For expansion rate, The coordinates of the sampled pixels from the dilated convolution; Candidate bubble regions are obtained by aligning regions of interest with multi-scale features to extract candidate region features. Calculate the centroid coordinates of the candidate bubble region, where the centroid coordinates are represented as follows: In the formula, Using the centroid coordinates, Let's say the pixel area of ​​the bubble. These are the pixel values ​​of the bubble mask; Calculate the radial distance between the centroid coordinates and the pixels within the candidate bubble region to obtain the radial distance from the pixel to the centroid. The radial distance from the pixel to the centroid is represented as follows: In the formula, For pixels to the centroid coordinates The radial distance; A constraint loss is constructed for the radial distance, resulting in a radial distance correction loss, which is expressed as follows: In the formula, For radial distance correction loss, Let be the expected function. The equivalent radius of the bubble is... ; Based on the radial distance correction loss, the masking loss and radial distance correction loss of the standard masked region convolutional neural network are weighted and summed to obtain the total loss for bubble segmentation. The total loss for bubble segmentation is expressed as follows: In the formula, The total loss for bubble segmentation, For mask loss, The weighting coefficients for radial distance correction loss; The candidate bubble region is segmented and optimized based on the total loss of bubble segmentation to obtain the bubble segmentation mask.

7. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 1, characterized in that, The extraction of multi-dimensional bubble feature parameters from the bubble segmentation mask includes: The reflectance of the defective and non-defective regions corresponding to the bubble segmentation mask is extracted to obtain the reflectance difference in the visible light band. Near-infrared reflectance is extracted from the defective and non-defective regions corresponding to the bubble segmentation mask to obtain the difference in near-infrared reflectance. Pixel statistics are performed on the bubble segmentation mask to obtain the bubble pixel area; The bubble volume is calculated by combining the bubble pixel area, coating thickness, and spatial resolution, and the equivalent diameter of the bubble is obtained from the bubble volume. Based on the bubble pixel area, the equivalent circumference is calculated. Edge detection is performed on the bubble segmentation mask to extract edge pixels and calculate the circumference, thus obtaining the actual circumference of the bubble. The roundness of the bubble is then obtained based on the actual circumference.

8. The method for detecting bubble defects in radiation-cooled coatings based on machine vision as described in claim 1, characterized in that, The prediction of performance loss of the radiation-cooled coating by combining the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density includes: The solar reflectance loss is obtained based on the solar reflectance of the defect-free coating, the total area of ​​the inspection area, the actual total physical area of ​​all bubbles, and the percentage decrease in reflectance in the visible light band. The solar reflectance loss is expressed as follows: In the formula, For the loss of solar reflectivity, For a defect-free coating, solar reflectivity The total area of ​​the detection area. This represents the actual total physical area of ​​all bubbles. This represents the percentage decrease in reflectivity in the visible light band. Based on the mechanical strength of the defect-free coating, material sensitivity coefficient, defect distribution index, and bubble density, the mechanical performance loss is obtained, whereby the mechanical performance loss is expressed as follows: In the formula, For mechanical property loss, denoted as σ0, where σ0 is the mechanical strength of the defect-free coating, m is the material sensitivity coefficient, and p is the defect distribution index. Based on the salt spray resistance time, corrosion rate coefficient, bubble density, ambient temperature, and salt spray concentration of the defect-free coating, the loss of salt spray corrosion resistance is obtained. The loss of salt spray corrosion resistance is expressed as follows: In the formula, To compensate for the loss of salt spray corrosion resistance, Salt spray resistance time for defect-free coatings The corrosion rate coefficient is... For ambient temperature, This refers to the salt spray concentration.

9. A machine vision-based system for detecting bubble defects in radiation-cooled coatings, characterized in that, include: An acquisition unit is used to acquire image data of the radiation-cooled coating, the image data including visible light images, near-infrared images and infrared thermal imaging data; The preprocessing unit is used to perform environmentally adaptive preprocessing on the visible light image, the near-infrared image, and the infrared thermal imaging data to correct the effects of illumination fluctuations and construct multi-scale temperature gradient features to obtain preprocessed image data. The feature extraction unit is used to extract spatial texture features and material composition features from the preprocessed image data through a dual-branch deep learning network, and to enhance the feature representation of the spatial texture features and material composition features by combining an attention module and performing feature fusion to obtain fused features and bubble density. A segmentation unit is used to perform pixel-level segmentation of the fused features to obtain a bubble segmentation mask; The prediction unit is used to extract multi-dimensional bubble feature parameters from the bubble segmentation mask, and combine the multi-dimensional bubble feature parameters, the multi-scale temperature gradient features, and the bubble density to predict the performance loss of the radiation-cooled coating, so as to obtain the coating defect detection result.

Images