A visual-based blade spraying uniformity identification method

Abstract

The application discloses a kind of based on vision's blade spraying uniformity identification method, belong to image processing technical field.The application is under two kinds of light intensity acquisition blade image, obtain medium light image and high light image;Extract brightness and carry out pixel level difference, obtain two brightness difference graphs;Luminance difference graph is handled using filter core, and structural uneven intensity graph is generated according to maximum and minimum response value difference;Pixel classification is carried out to structural uneven intensity graph, distinguish smooth point and structural mutation point, and carry out binary and OR operation, obtain structure co-display graph, and structural extraction graph is obtained by morphological operation;According to structural uneven intensity graph, generate unevenness weight graph;Extract structural extraction graph and unevenness weight graph feature, carry out fusion splicing, and generate spraying uniformity score.The method utilizes the visual difference under different illumination conditions, realizes the high-precision identification of blade spraying uniformity.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a vision-based method for identifying the uniformity of leaf coating. Background Technology

[0002] As a core component of hydroelectric power generation equipment, the uniformity of the wear-resistant coating sprayed on the surface of turbine blades directly determines the blade's service life and the unit's operational stability. Therefore, accurate inspection of the coating quality is crucial. Among existing inspection methods, manual visual assessment remains the mainstream approach. However, this method is highly dependent on the experience of the inspectors, resulting in significant subjective bias. It not only struggles to identify minute coating defects at the millimeter level (such as localized missed spraying or thin spraying), but also suffers from low efficiency and poor consistency, failing to meet the needs of industrial-scale batch inspection. While traditional non-destructive testing technologies such as eddy current testing can detect defects such as delamination and inclusions within the coating, they are limited by reliance on imported equipment and high inspection costs. More importantly, their ability to visualize the surface coating uniformity is insufficient, failing to quickly output intuitive uniformity assessment results and providing direct data support for adjusting coating process parameters.

[0003] To address the aforementioned pain points, image recognition technology, due to its advantages of low cost and fast detection speed, has gradually been applied to the field of turbine blade coating inspection. However, in practical applications, the surface differences of turbine blades after coating are small, making it difficult to directly identify the texture of uneven coating through images, resulting in low accuracy in identifying the uniformity of blade coating. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides a vision-based method for identifying the uniformity of blade spraying, which solves the problem of low accuracy in identifying the uniformity of blade spraying in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a vision-based method for identifying the uniformity of blade coating, comprising the following steps:

[0006] Leaf images were acquired under two supplemental lighting intensities to obtain medium and high light images. Brightness was extracted and pixel-level difference was performed to obtain two brightness difference images.

[0007] Each brightness difference map is processed using a set of filter kernels, and two structural inhomogeneity intensity maps are obtained based on the difference between the maximum and minimum response values.

[0008] The pixels in each structural inhomogeneity intensity map are classified to obtain smooth points and structural abrupt points.

[0009] Based on the smoothing point and the structural abrupt change point, the two structural inhomogeneity strengths are compared. Figure TwoValue, and the "or" operation, get structure co-show map, and then morphological operation, get structure extraction map;

[0010] According to two structure uneven intensity map, generate uneven weight map;

[0011] Respectively on structure extraction map and uneven weight map extraction features, and feature fusion and splicing, based on splicing features generate spraying uniformity score.

[0012] Further, the process of obtaining two brightness difference maps includes:

[0013] The neutral and high light image is mapped to the HSV color space respectively, and the brightness channel is extracted to obtain the medium brightness image and the high brightness image;

[0014] The medium brightness image and the standard medium brightness image are subtracted in brightness according to pixel points to obtain the first brightness difference map;

[0015] The high brightness image and the standard high brightness image are subtracted in brightness according to pixel points to obtain the second brightness difference map.

[0016] Further, the process of obtaining two structure uneven intensity maps includes:

[0017] A group of filter kernels with different scale parameters are constructed;

[0018] Each filter kernel is convolved with each brightness difference map to obtain a group of response maps;

[0019] For a group of response maps of the same brightness difference map, the maximum response value and the minimum response value in the filter response are selected at the same pixel point;

[0020] At the same pixel point, the maximum response value and the minimum response value in the filter response are subtracted to obtain a response difference value, and each response difference value belonging to the first brightness difference map constitutes the first structure uneven intensity map, and each response difference value belonging to the second brightness difference map constitutes the second structure uneven intensity map.

[0021] Further, the process of obtaining smooth points and structure mutation points includes:

[0022] Taking any unclassified pixel point on each structure uneven intensity map as the center, the variance of each response difference value in the neighborhood range of the center pixel point is calculated;

[0023] When the variance is less than a threshold value, the corresponding center pixel point and each pixel point in the neighborhood range are classified into a class;

[0024] Select a pixel point not as the center in the class as a new center, calculate the variance of each response difference value in the neighborhood range of the new center pixel point, when the variance is less than the threshold, the pixel points in the neighborhood range of the new center pixel point are classified into the same class, until the pixel points in the class cannot be expanded, the next class is classified, until all pixel points on the structure uneven intensity map are traversed;

[0025] Name the pixel points in each class as smooth points, and name the pixel points not classified into the class as structure mutation points.

[0026] Further, the process of obtaining the structure co-expression map includes:

[0027] In the first structure uneven intensity map, set the smooth points to 0 and the structure mutation points to 1 to obtain the first structure mutation binary map;

[0028] In the second structure uneven intensity map, set the smooth points to 0 and the structure mutation points to 1 to obtain the second structure mutation binary map;

[0029] For the same pixel point position, take the "or" operation of the first structure mutation binary map and the second structure mutation binary map to obtain the structure co-expression map.

[0030] Further, the process of obtaining the structure extraction map includes:

[0031] Perform an opening operation on the structure co-expression map to obtain an image after the opening operation;

[0032] Perform a closing operation on the image after the opening operation to obtain the structure extraction map.

[0033] Further, the process of generating the unevenness weight map includes:

[0034] The ratio of each response difference value in each structure uneven intensity map to the maximum response difference value in the corresponding structure uneven intensity map is taken as the relative unevenness;

[0035] The relative unevennesses in the two structure uneven intensity maps belonging to the same pixel point are added to obtain the unevenness weight;

[0036] The unevenness weights of all pixel points form the unevenness weight map.

[0037] Further, the process of generating the spraying uniformity score includes:

[0038] Features are extracted from the structure extraction map and the unevenness weight map respectively, and feature fusion and splicing are performed to obtain spliced features;

[0039] A CNN network is used to perform deep feature extraction on the spliced features, and a fully connected layer is used to generate the spraying uniformity score.

[0040] Further, the process of obtaining the spliced features comprises:

[0041] The first feature extraction unit is used to extract features from the structure extraction map, so as to obtain the saliency response features and the global statistical features of the structure extraction map;

[0042] The second feature extraction unit is used to extract features from the unevenness weight map, so as to obtain the saliency response features and the global statistical features of the unevenness weight map;

[0043] The saliency response features of the structure extraction map and the saliency response features of the unevenness weight map are multiplied element by element, so as to obtain the saliency response enhanced features;

[0044] The global statistical features of the structure extraction map and the global statistical features of the unevenness weight map are multiplied element by element, so as to obtain the global statistical enhanced features;

[0045] The saliency response enhanced features and the global statistical enhanced features are spliced by using the Concat layer, so as to obtain the spliced features.

[0046] Further, the first feature extraction unit and the second feature extraction unit have the same structure, and each of the first feature extraction unit and the second feature extraction unit comprises: a first convolutional layer, a second convolutional layer, a third convolutional layer, a maximum pooling layer and an average pooling layer;

[0047] The input end of the first convolutional layer is used as the input end of the first feature extraction unit and the second feature extraction unit, the output end of the first convolutional layer is connected with the input end of the second convolutional layer, the output end of the second convolutional layer is connected with the input end of the third convolutional layer, the output end of the third convolutional layer is connected with the input end of the maximum pooling layer and the input end of the average pooling layer respectively, the output end of the maximum pooling layer is used as the first output end of the first feature extraction unit and the second feature extraction unit, and the output end of the average pooling layer is used as the second output end of the first feature extraction unit and the second feature extraction unit.

[0048] The present application has the following advantages:

[0049] 1. The present application is aimed at the problem that the overall difference of the surface of the water turbine blade after spraying is small, and the uneven texture is easy to be covered. The unevenness intensity map is constructed by using "double light illumination difference + multi-scale filter kernel" in the present application. The brightness difference of the uneven area is amplified by using medium and high light intensity, and the "actual spraying and standard state deviation" is highlighted by pixel level difference. The texture features are amplified by using "maximum-minimum response difference", the problem of "uneven texture and background feature confusion" in traditional image recognition is solved, and the abnormal area is highlighted.

[0050] 2、The application obtains the suspected uneven spraying texture through the "or" operation and morphological processing of the structure co-plot, and then generates an unevenness weight map, thereby reducing the probability of misjudging the inherent characteristics of the blade as a spraying defect, and solving the problem of low recognition accuracy in the prior art.

[0051] 3、The application converts the spraying uniformity into a quantifiable score through feature fusion of the structure extraction map and the unevenness weight map, and can intuitively reflect the overall spraying quality of the blade.

[0052] 4、The application is based on visual image acquisition and processing, does not need to rely on complex non-destructive testing equipment, has high automation degree in the detection process, can quickly complete the blade spraying uniformity evaluation, is suitable for batch detection scenes on the production line, reduces the detection cost and time cost while ensuring the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure One It is a flowchart of a visual-based blade spraying uniformity recognition method;

[0054] Figure Two It is a flowchart of the process of obtaining the splicing feature;

[0055] Figure Three It is a structural schematic diagram of the first feature extraction unit and the second feature extraction unit. DETAILED DESCRIPTION

[0056] The specific embodiments of the application are described below to facilitate those skilled in the art to understand the application, but it should be clear that the application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the application defined and determined by the appended claims, and all applications utilizing the concept of the application are within the scope of protection.

[0057] As shown in Figure One A visual-based blade spraying uniformity recognition method includes the following steps:

[0058] Obtain blade images under two light supplement intensity conditions to obtain neutral and high light images, extract brightness, perform pixel-level difference, and obtain two brightness difference maps;

[0059] Each brightness difference map is processed using a set of filter kernels, and based on the difference between the maximum response value and the minimum response value, two structure unevenness intensity maps are obtained.

[0060] Classify the pixel points in each structure unevenness intensity map to obtain smooth points and structure mutation points.

[0061] According to the smooth points and the structural mutation points, two structural uneven intensity maps are valued and subjected to OR operation to obtain a structural co-display map, and then subjected to morphological operation to obtain a structural extraction map. Figure Two

[0062] According to the two structural uneven intensity maps, an unevenness weight map is generated.

[0063] Features are extracted from the structural extraction map and the unevenness weight map respectively, and feature fusion and splicing are performed, and a spraying uniformity score is generated based on the spliced features.

[0064] In the embodiment, the process of obtaining two luminance difference maps includes:

[0065] The medium and high illumination images are respectively mapped to an HSV color space, and a luminance channel is extracted to obtain a medium luminance image and a high luminance image.

[0066] The medium luminance image is subjected to luminance subtraction with a standard medium luminance image pixel by pixel to obtain a first luminance difference map.

[0067] The high luminance image is subjected to luminance subtraction with a standard high luminance image pixel by pixel to obtain a second luminance difference map.

[0068] In the embodiment, the two light supplement intensity conditions can be specifically set as medium illumination intensity 500 lux-800 lux and high illumination intensity 1200 lux-1500 lux. After the surface of the water turbine blade is sprayed, there are still fine coating textures and blade structures (such as casting lines). The medium illumination intensity can ensure the image definition while avoiding overexposure of the coating caused by strong light, so that the HSV luminance channel can accurately reflect the gradual difference in coating thickness. Under high illumination, the reflection characteristics of the uneven spraying area (such as the thin coating area and the missed spraying point) and the uniform area will produce more significant contrast - the area with thick coating has lower brightness due to enhanced diffuse reflection, and the area with thin coating has higher brightness due to exposure of the blade base color.

[0069] In the embodiment, the medium and high illumination images are the same in other shooting parameters except for the light supplement intensity conditions.

[0070] The texture formed by uneven spraying (such as the accumulation lines caused by local over-thickness and the mottled marks formed by missed spraying) has a core performance of abnormal rules of luminance distribution. The present application extracts the luminance channel (V channel) through the HSV color space, which can completely eliminate the interference of hue (H) and saturation (S), and only retains the luminance change information directly related to the texture, so that the subsequent analysis can completely focus on the luminance distribution characteristics of the uneven texture.

[0071] ​The brightness of the sprayed uniform area will present a stable and regular distribution (consistent with the standard image) under medium / high light, and uneven textures (such as accumulation, missing spraying) will break this regularity, resulting in a significant deviation between the actual brightness and the standard image. By pixel-level difference between the "actual brightness image-standard brightness image", this deviation can be directly converted into light and dark changes in the difference image - for example, the accumulation texture area has thick coating and weak reflection, the actual brightness is lower than the standard image, and after difference, it presents a dark area; the missing spraying texture area has exposed base color and strong reflection, the actual brightness is higher than the standard image, and after difference, it presents a bright area. This difference method can specifically amplify the outline and details of uneven textures, solving the problem of "uneven textures and normal surface features being confused" when directly analyzing the original image.

[0072] Under medium light, the brightness changes more gently, which is suitable for capturing large-area and gradual uneven textures (such as stripes formed by slow changes in coating thickness); under high light, the brightness is more sensitive to fine structures, which can highlight small-scale and abrupt uneven textures (such as spots formed by pinholes and local missing spraying). The combination of the two forms complementary difference images that can completely cover all types of texture patterns that may be produced by uneven spraying, ensuring that no key features are missed during subsequent filter processing and structure extraction, and laying a foundation for accurate texture recognition.

[0073] The standard medium brightness image and the standard high brightness image are pre-stored blade images with uniform spraying.

[0074] In this embodiment, the process of obtaining two structure unevenness intensity images includes:

[0075] A set of filter kernels with different scale parameters is constructed;

[0076] Each filter kernel is convolved with each brightness difference image to obtain a set of response images;

[0077] For a set of response images of the same brightness difference image, the maximum response value and the minimum response value in the filter response are selected at the same pixel point;

[0078] At the same pixel point, the maximum response value and the minimum response value in the filter response are subtracted to obtain a response difference value. The response difference values belonging to the first brightness difference image are arranged according to the corresponding pixel points to form a first structure unevenness intensity image, and the response difference values belonging to the second brightness difference image are arranged according to the corresponding pixel points to form a second structure unevenness intensity image.

[0079] In this embodiment, the expression of the filter kernel is:

[0080] ,

[0081] where G σ(x,y) is a filter kernel, e is a natural constant, (x,y) is a pixel position, and σ is a scale parameter.

[0082] When a set of filter kernels is used to process the same image, multiple response values are generated at the same pixel position.

[0083] When σ is 1, 2, 4, and 8 in turn, a set of filter kernels with different scale parameters is: G1(x,y), G2(x,y), G4(x,y), and G8(x,y).

[0084] The luminance distribution of the uniform spraying area is flat, and the response values of different scale filters are small; and the non-uniform texture area has a luminance mutation (such as the boundary between the normal area and the non-spraying area, and the ups and downs of the accumulated texture), and the response values of different scale filters are significantly different - the edge has a small scale kernel response and a large scale kernel response, or vice versa. Through the difference calculation of the "maximum response value - minimum response value", the difference can be directly amplified, so that the non-uniform texture area presents a high difference value (high contrast) in the structure uneven intensity graph, and the uniform area presents a low difference value (low contrast), completely separating the "non-uniform texture" and "uniform background", and solving the problem of "non-prominent texture features" in the original response graph.

[0085] In the embodiment, the process of obtaining the smooth points and the structure mutation points includes:

[0086] Taking any unclassified pixel point on each structure uneven intensity graph as the center, the variance of each response difference value in the neighborhood range of the center pixel point is calculated;

[0087] When the variance is less than a threshold value, the corresponding center pixel point and each pixel point in the neighborhood range are classified into a class;

[0088] A pixel point in the class that has not been centered is selected as a new center, the variance of each response difference value in the neighborhood range of the new center pixel point is calculated, and when the variance is less than the threshold value, the pixel points in the neighborhood range of the new center pixel point are classified into the same class, until the pixel points in the class cannot be expanded, the classification of the next class is performed, and all pixel points on the structure uneven intensity graph are traversed.

[0089] The pixel points in each class are named as smooth points, and the pixel points that are not classified into the classes are named as structure mutation points.

[0090] The threshold value can be set according to experiments or experience, and in the embodiment, the empirical value of the threshold value is 10.

[0091] The present application aims at each structure uneven intensity map, randomly selects an unclassified pixel point as an initial center, and defines a fixed size neighborhood (such as a 3*3 pixel window). By calculating the variance of all response difference values in the neighborhood of the center, the "intensity consistency" of the local area is quantified - the smaller the variance, the more concentrated the response difference value distribution in the neighborhood, and the flatter the structural features of the corresponding area (consistent with the core attributes of smooth points); the larger the variance, the more intense the local intensity fluctuation, and there may be structural mutations. The calculated neighborhood variance is compared with the preset threshold value, and if the variance is less than the threshold value, it is determined that the center and the neighborhood pixels have "structural consistency", and all pixels in the neighborhood of the center are classified into the same pixel class. In the pixel class formed, select a pixel that is not the center as the "new center", calculate the neighborhood variance of the new center, and if it is still less than the threshold value, the neighborhood pixels of the new center are included in the current pixel class, and the dynamic expansion of the clustering boundary is realized. The iteration process continues until the pixel class cannot include new pixels, and this class is classified.

[0092] The present application constructs two structure uneven intensity maps to highlight the distinction of textures. Therefore, the present application classifies the regions belonging to the same area range and having consistent response difference values into a class, and the brightness and structural response of the uniform spraying region are stable and regular (the response difference value fluctuation is small), while the uneven texture region (such as spraying leakage and accumulation) will produce significant response difference value fluctuation due to brightness mutation. Through clustering of "the same area range and consistent response difference value", the "region with stable features" in the uniform spraying or uneven texture region can be aggregated into a class (smooth point), and the "region with feature mutation" can be separated out (structural mutation point), which is convenient for subsequent acquisition of texture information.

[0093] In the present embodiment, the process of obtaining the structure co-display map includes:

[0094] In the first structure uneven intensity map, the smooth point is set to 0, and the structural mutation point is set to 1, to obtain the first structure mutation binary map;

[0095] In the second structure uneven intensity map, the smooth point is set to 0, and the structural mutation point is set to 1, to obtain the second structure mutation binary map;

[0096] For the same pixel position, the first structure mutation binary map and the second structure mutation binary map are subjected to "or" operation to obtain the structure co-display map.

[0097] In the present embodiment, the process of obtaining the structure extraction map includes:

[0098] The structure co-display map is subjected to opening operation to obtain the image after opening operation;

[0099] The image after the opening operation is closed to obtain a structure extraction image.

[0100] The first and second structure mutation binary images correspond to the spraying uneven texture features under the medium and high light conditions respectively. Through the OR operation, any pixel in the two binary images determined as a structure mutation point can be reserved, that is, the uneven feature of the region is captured at least under one light condition, which can completely cover the differentiated uneven features under the two light conditions (such as the gradual unevenness easily identified under the medium light and the mutation unevenness easily identified under the high light), avoiding the missed judgment due to the single light detection blind area (for example, a missed spraying point only appears in the high light image, and the OR operation can effectively retain it). Meanwhile, the operation can integrate the features of the two detection dimensions, so that the structure co-display image completely presents the global distribution of the spraying unevenness.

[0101] Opening operation: first erosion and then dilation, which can remove isolated small noise points (such as single pixel misjudgment) in the structure co-display image, and disconnect the weak false connection (such as the false connection of uneven regions caused by image noise), avoiding the influence of small interference on the overall structure, and making the outline of uneven texture clearer.

[0102] Closing operation: first dilation and then erosion, which can fill small holes inside the uneven region (such as texture breakage caused by local brightness fluctuation), and connect adjacent scattered uneven regions (such as a missed spraying region being mistakenly divided into multiple blocks), ensuring that the extracted structure completely reflects the actual distribution range of the spraying unevenness.

[0103] The obtained structure extraction image not only removes irrelevant interference, but also completely retains the spatial form of the spraying uneven texture, improving the accuracy of the final uniformity score.

[0104] In the embodiment, the process of generating the unevenness weight map includes:

[0105] The ratio of each response difference value in each structure uneven intensity image to the maximum response difference value in the corresponding structure uneven intensity image is taken as the relative unevenness.

[0106] The relative unevennesses in the two structure uneven intensity images belonging to the same pixel point are added to obtain the unevenness weight.

[0107] The unevenness weights of all pixel points are arranged according to the corresponding pixel point positions to form the unevenness weight map.

[0108] The present application quantifies the intensity difference of spraying unevenness into a continuous value between 0 and 1 by calculating the relative unevenness through the ratio of the response difference value to the maximum response difference value. The closer the ratio is to 1, the more serious the unevenness of the pixel point (such as a large-scale missed spraying region); and the closer the ratio is to 0, the lighter the unevenness (such as a small range of coating thickness fluctuation).

[0109] The two structural unevenness intensity maps correspond to unevenness characteristics in medium and high light conditions respectively (for example, the medium light map focuses on gradual unevenness, and the high light map focuses on sudden unevenness). By adding the relative unevenness of the same pixel point in the two maps, the unevenness information under the two light conditions can be fused, and the unevenness weight map can reflect the severity of unevenness in different conditions.

[0110] In this embodiment, the process of generating the spray uniformity score includes:

[0111] Features are extracted from the structure extraction map and the unevenness weight map respectively, and feature fusion and splicing are performed to obtain spliced features.

[0112] The CNN network is used for deep feature extraction of the spliced features, and the fully connected layer is used to generate the spray uniformity score.

[0113] In this embodiment, as shown in Figure Three the process of obtaining the spliced features includes:

[0114] The first feature extraction unit is used to extract features from the structure extraction map to obtain the salient response features and the global statistical features of the structure extraction map.

[0115] The second feature extraction unit is used to extract features from the unevenness weight map to obtain the salient response features and the global statistical features of the unevenness weight map.

[0116] The salient response features of the structure extraction map and the salient response features of the unevenness weight map are multiplied element by element to obtain salient response enhanced features.

[0117] The global statistical features of the structure extraction map and the global statistical features of the unevenness weight map are multiplied element by element to obtain global statistical enhanced features.

[0118] The Concat layer is used to splice the salient response enhanced features and the global statistical enhanced features to obtain spliced features.

[0119] Salient response enhanced features: by element-wise multiplication of "structure extraction map salient response features x unevenness weight map salient response features", the "structure morphology" and "unevenness intensity" can be directly associated - for example, the edge features (structure salient) of a certain uneven area are accompanied by high unevenness intensity (weight salient), and after multiplication, the feature value of the area will be greatly enlarged, further highlighting the key uneven details.

[0120] Global statistical enhancement features: By multiplying the elements of "structure global statistical features x weight global statistical features", the "structure distribution rule" and the "intensity average level" can be integrated, for example, the area proportion of uneven regions (structure statistics) and the average intensity of the whole uneven intensity (weight statistics) are multiplied, and after multiplication, the feature value of the region will be greatly enlarged, further highlighting the key uneven details.

[0121] In this embodiment, the uneven texture region presents a high difference value (high contrast) in the structure uneven intensity map, and the uniform region presents a low difference value (low contrast), so the weight of the uneven region is larger, which can improve the attention of the corresponding feature and improve the prediction accuracy.

[0122] In this embodiment, as shown in Figure Three The first feature extraction unit and the second feature extraction unit are the same in structure, and each includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a maximum pooling layer and an average pooling layer.

[0123] The input end of the first convolutional layer is the input end of the first feature extraction unit and the second feature extraction unit, the output end thereof is connected with the input end of the second convolutional layer; the output end of the second convolutional layer is connected with the input end of the third convolutional layer; the output end of the third convolutional layer is connected with the input end of the maximum pooling layer and the input end of the average pooling layer respectively; the output end of the maximum pooling layer is the first output end of the first feature extraction unit and the second feature extraction unit; and the output end of the average pooling layer is the second output end of the first feature extraction unit and the second feature extraction unit.

[0124] In this embodiment, the convolution kernel size of the first convolutional layer, the second convolutional layer and the third convolutional layer is 3x3.

[0125] In this embodiment, the spraying uniformity score ranges from 0 to 10, 9-10: no visible uneven texture on the surface of the blade; 7-8: only sporadic uneven texture is detected on the edge of the blade, the area of a single uneven region is less than 1 cm2, 5-6: a small amount of dispersed uneven texture exists in the core area of the blade, the area of a single uneven region is 1-3 cm2, 3-4: continuous uneven texture (such as thick coating stripes with a length of more than 5 cm, or a spraying omission band) appears in the core area of the blade, the area of a single uneven region is 3-5 cm2, and 0-3: a large area of the surface of the blade is uneven, and a continuous spraying omission area (with an area of more than 5 cm2) exists.

[0126] In this embodiment, the blade image is directly input into the CNN network as a comparative technical solution, as shown in Table 1, compared with the traditional single CNN method, the accuracy of the present application is improved by 13.7 percentage points, the false alarm rate is reduced by about 74%, the score standard deviation is reduced by 70%, and the F1 value is improved to 0.97, which shows that the present application significantly improves the recognition accuracy and robustness.

[0127] Table 1 Comparison table of experimental effects of traditional single CNN method and the method of the present application

[0128]

[0129] The present application aims at the problem that the overall difference of the sprayed surface of the water turbine blade is small, and the uneven texture is easy to be covered. The uneven intensity map is constructed by "dual light illumination difference + multi-scale filter kernel" structure: first, the brightness difference of the uneven spraying area is amplified by medium and high two light intensity, and then the "deviation of actual spraying and standard state" is highlighted by pixel-level difference; subsequently, the texture features are amplified by "maximum-minimum response difference", the problem of "uneven texture and background feature confusion" in traditional image recognition is solved, and the abnormal area is highlighted.

[0130] The present application obtains the suspected uneven spraying texture through "or" operation and morphological processing of the structure co-display map, and then generates the unevenness weight map. The structure extraction map and the unevenness weight map are integrated to reduce the probability of misjudging the inherent characteristics of the blade as spraying defects, and solve the problem of low recognition accuracy in the prior art.

[0131] The present application converts the spraying uniformity into a quantifiable score through feature fusion of the structure extraction map and the unevenness weight map, and can intuitively reflect the overall spraying quality of the blade.

[0132] The present application is based on visual image acquisition and processing, does not need to rely on complex nondestructive testing equipment, has high degree of automation in the detection process, can quickly complete the spraying uniformity evaluation of the blade, is suitable for batch detection scene on the production line, reduces the detection cost and time cost while ensuring the accuracy.

[0133] The above is only the preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A vision-based blade spray uniformity identification method, characterized in that, The method comprises the following steps: Obtain leaf images under two light supplement intensity conditions to obtain medium and high light images, extract brightness, perform pixel-level difference, and obtain two brightness difference images; Map the medium and high light images to an HSV color space respectively, extract the brightness channel, and obtain a medium brightness image and a high brightness image; Subtract the brightness of the medium brightness image from that of a standard medium brightness image pixel by pixel to obtain a first brightness difference image; Subtract the brightness of the high brightness image from that of a standard high brightness image pixel by pixel to obtain a second brightness difference image; Process each brightness difference image using a set of filter kernels, and obtain two structure unevenness intensity images based on the difference between the maximum response value and the minimum response value; Construct a set of filter kernels with different scale parameters; Perform convolution operation on each filter kernel and each brightness difference image to obtain a set of response images; For the same brightness difference image, select the maximum response value and the minimum response value in the filter response at the same pixel point; Subtract the maximum response value from the minimum response value in the filter response at the same pixel point to obtain a response difference value, and construct a first structure unevenness intensity image from the response difference values belonging to the first brightness difference image, and construct a second structure unevenness intensity image from the response difference values belonging to the second brightness difference image; Classify the pixel points in each structure unevenness intensity image to obtain smooth points and structure mutation points; Binaryzation the two structure unevenness intensity images according to the smooth points and the structure mutation points, perform "or" operation, obtain a structure co-expression image, and perform morphological operation to obtain a structure extraction image; Generate unevenness weight images according to the two structure unevenness intensity images; Extract features from the structure extraction image and the unevenness weight image respectively, perform feature fusion and splicing, generate a spraying uniformity score based on the spliced features.

2. The vision-based blade spray uniformity recognition method of claim 1, wherein, The process of obtaining smooth points and structure mutation points comprises: Take any unclassified pixel point in each structure unevenness intensity image as a center, calculate the variance of each response difference value in the neighborhood range of the center pixel point; When the variance is less than a threshold value, classify the center pixel point and the pixel points in the neighborhood range into a class; Select an unclassified pixel point in the class as a new center, calculate the variance of each response difference value in the neighborhood range of the new center pixel point, and when the variance is less than a threshold value, classify the pixel points in the neighborhood range of the new center pixel point into the same class, until the pixel points in the class cannot be expanded, and then classify the next class until all pixel points in the structure unevenness intensity image are traversed; Name the pixel points in each class as smooth points, and name the pixel points not classified into a class as structure mutation points.

3. The vision-based blade spray uniformity identification method of claim 1, wherein, The process of obtaining a structure co-expression image comprises: Set the smooth points to 0 and the structure mutation points to 1 in the first structure unevenness intensity image to obtain a first structure mutation binary image; Set the smooth points to 0 and the structure mutation points to 1 in the second structure unevenness intensity image to obtain a second structure mutation binary image; For the same pixel point position, perform "or" operation on the first structure mutation binary image and the second structure mutation binary image to obtain a structure co-expression image.

4. The vision-based blade spray uniformity identification method of claim 1, wherein, The process of obtaining a structure extraction image comprises: Perform opening operation on the structure co-expression image to obtain an image after opening operation; The image after the opening operation is closed to obtain a structure extraction image.

5. The vision-based blade spray uniformity identification method of claim 1, wherein, The process of generating the non-uniformity weight map includes: The ratio of each response difference value in each structure non-uniformity intensity map to the maximum response difference value in the corresponding structure non-uniformity intensity map is taken as a relative non-uniformity degree; The relative non-uniformity degrees in the two structure non-uniformity intensity maps belonging to the same pixel point are added to obtain a non-uniformity weight; The non-uniformity weights of all pixel points form a non-uniformity weight map.

6. The vision-based blade spray uniformity identification method of claim 1, wherein, The process of generating the spraying uniformity score includes: Features are extracted from the structure extraction image and the non-uniformity weight map respectively, and feature fusion and splicing are performed to obtain spliced features; A CNN network is used to extract deep features from the spliced features, and a full connection layer is used to generate a spraying uniformity score.

7. The vision-based blade spray uniformity identification method of claim 6, wherein, The process of obtaining the spliced features includes: A first feature extraction unit is used to extract features from the structure extraction image to obtain significant response features and global statistical features of the structure extraction image; A second feature extraction unit is used to extract features from the non-uniformity weight map to obtain significant response features and global statistical features of the non-uniformity weight map; The significant response features of the structure extraction image and the significant response features of the non-uniformity weight map are multiplied element by element to obtain significant response enhanced features; The global statistical features of the structure extraction image and the global statistical features of the non-uniformity weight map are multiplied element by element to obtain global statistical enhanced features; A Concat layer is used to splice the significant response enhanced features and the global statistical enhanced features to obtain spliced features.

8. The visual-based blade spray uniformity identification method of claim 7, wherein, The first feature extraction unit and the second feature extraction unit have the same structure and both include a first convolutional layer, a second convolutional layer, a third convolutional layer, a maximum pooling layer and an average pooling layer; The input end of the first convolutional layer is the input end of the first feature extraction unit and the second feature extraction unit, the output end thereof is connected with the input end of the second convolutional layer, the output end of the second convolutional layer is connected with the input end of the third convolutional layer, the output end of the third convolutional layer is connected with the input end of the maximum pooling layer and the input end of the average pooling layer respectively, the output end of the maximum pooling layer is the first output end of the first feature extraction unit and the second feature extraction unit, and the output end of the average pooling layer is the second output end of the first feature extraction unit and the second feature extraction unit.

Images