An intelligent detection method and system for decoration quality

Abstract

This application relates to the field of image processing technology, and specifically relates to an intelligent detection method and system for decoration quality. The method includes the steps of: preprocessing and tagging historical images of the wall surface, where different tags represent different defects; setting a scale, taking any pixel point as the central pixel point, and using the scale as the radius to obtain a circular trajectory line, and sampling the pixel points on the circular trajectory line to obtain neighborhood sampling points; at any scale, calculating the overall consistency of the central pixel point within the area where any type of defect is located to obtain the corresponding direction set at this scale; calculating the set difference based on the direction sets corresponding to various defects at each scale; training a neural network model to obtain a detection model; inputting a real-time picture into the detection model and outputting the probabilities that the real-time picture belongs to different tags. This application has the effect of improving the detection accuracy of the wall surface.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and particularly relates to an intelligent detection method and system for decoration quality. Background Art

[0002] During the interior decoration process, factors such as wall coatings (e.g., insufficient adhesion of inferior putty) and changes in temperature and humidity of the external environment can cause surface defects on the decorated walls. The specific defect types can be cracks, bulges, scratches, etc. The occurrence of the above defects will, on the one hand, affect the flatness and aesthetics of the wall surface, and on the other hand, there is also a risk of wall coating peeling off, increasing the subsequent maintenance cost. Therefore, the detection of walls after decoration or during the decoration process is an important link in the decoration process. Traditional wall defect detection mostly relies on manual observation, which is, on the one hand, easily affected by subjective factors, and on the other hand, it will also increase the labor burden of the staff and is not conducive to improving efficiency. In order to reduce the labor burden of the staff, in related technologies, the walls after decoration can be detected through image recognition and deep learning algorithms. For example, a wall crack diagnosis method, device, computer device, and storage medium disclosed in the patent application document with the publication number CN118154604A. It first obtains the image information of the wall, and then identifies the wall image to identify the cracks on the wall surface.

[0003] LBP (Local Binary Patterns) is an algorithm for image texture analysis and feature extraction. The extraction of wall defect information can be achieved through LBP. However, due to the strong directionality of the wall surface features. Traditional LBP extracts features based on the gray value relationship between each pixel and its neighboring pixels. During the extraction process, the importance of neighboring pixels is the same and the scale is single, resulting in the single-scale LBP features obtained ignoring the direction of the features, and there is a large deviation in the effective extraction range, causing low accuracy in the recognition of decorated walls. Summary of the Invention

[0004] In order to improve the accuracy of defect recognition of decorated walls, this application provides an intelligent detection method and system for decoration quality.

[0005] In the first aspect, this application provides an intelligent detection method for decoration quality, adopting the following technical solution:

[0006] An intelligent detection method for decoration quality includes the steps of: preprocessing and labeling the historical images of the wall, where different labels represent different defects;

[0007] Set the scale, take any pixel point as the central pixel point, and use the scale as the radius to obtain a circular trajectory line, and sample the pixel points on the circular trajectory line to obtain neighborhood sampling points;

[0008] At any scale, calculate the overall consistency of the central pixel points in the area where defects of any type are located to obtain the corresponding direction set at that scale; calculate the set difference based on the direction sets corresponding to various defects at each scale;

[0009] Train a neural network model to obtain a detection model; input a real-time image into the detection model and output the probabilities of the real-time image belonging to different labels;

[0010] Among them, the calculation formula for the probabilities of the real-time image belonging to different labels is: ; represents the probability that the real-time image belongs to this label, is the number of selected scales, represents the set difference between the direction sets corresponding to different defects under the feature map obtained based on scale acquired, and represents the probabilities of historical images output during the training process of the neural network model belonging to different labels; when the probability meets the preset threshold corresponding to any label, it indicates that there are defects corresponding to the label on the wall.

[0011] In this application, the direction consistency between the central pixel points and the neighboring pixel points is calculated, and the eigenvalue of the central pixel is determined using the direction consistency as the weight, making the feature extraction of the image more accurate and comprehensive, reducing background noise interference. At the same time, a direction set is constructed based on the overall consistency of the central sampling points in the defect area, and the set difference between different direction sets is calculated; the greater the set difference between the direction sets corresponding to different defects at the same scale, the stronger the ability to distinguish defect types at that scale. Using this feature, weights are assigned to different scales. Combining the weights of different scales, the probabilities of the real-time image of the wall belonging to different defects are calculated to improve the accuracy of decoration quality detection.

[0012] Optionally, the calculation formula for the set difference is: ; represents the set difference between different defects at one scale, represents the chi-square test value between the direction sets of the th defect and the th defect; is a linear normalization function; is the number of defect types.

[0013] In the formula, the chi-square test value represents the difference between different direction sets. Different sets correspond to different defects. The greater the difference, the stronger the ability to distinguish defects. Sum the chi-square test values between different direction sets. The larger the sum, the greater the difference between the sets. Therefore, the features extracted at this scale can better distinguish various defect types, enabling weighting of different scales.

[0014] Optionally, the method for constructing the direction set includes: selecting multiple scales for the defect region and calculating the overall consistency of each pixel point within the defect region at multiple scales; the overall consistency of multiple pixel points in the defect region constitutes the direction set.

[0015] Optionally, the steps for calculating the overall consistency are: calculating the direction consistency between the central sampling point and corresponding different neighborhood sampling points; calculating the overall consistency of the central pixel point based on the direction consistency between the central pixel point and multiple neighborhood sampling points;

[0016] The formula for the overall consistency is: ; In the formula, represents the overall consistency of the central pixel point, represents the number of neighborhood sampling points, represents the pixel point and its th neighborhood sampling point's direction consistency.

[0017] One central pixel point corresponds to multiple neighborhood sampling points, and there is one direction consistency between one central pixel point and one neighborhood sampling point. Calculate the direction consistency between the central sampling point and different neighborhood sampling points, and then process the direction consistency between multiple neighborhood sampling points and the central sampling point to obtain the overall trend of the central pixel point and multiple neighborhood sampling points, improving the robustness of the detection.

[0018] Optionally, the steps for calculating the direction consistency include: calculating the gradient directions of the neighborhood sampling point and the central pixel point respectively; calculating the direction consistency based on the gradient directions of the neighborhood sampling point and the central pixel point;

[0019] The formula for the direction consistency is ;

[0020] represents the direction consistency between the central pixel point and the th neighborhood sampling point, represents the exponential function with as the base, represents the gradient direction of the central pixel point, represents the th neighborhood sampling point's gradient direction, Represents the standard deviation of the gradient direction.

[0021] Calculate the consistency of the gradient directions of the central pixel point and the neighborhood sampling points. The more consistent the gradient directions of the two are, the more consistent the directions of the textures represented by the two are, enhancing the description of the directional texture of the wall surface and improving the accuracy of the detection of the decorated wall surface.

[0022] Optionally, the steps for calculating the respective gradient directions of the neighborhood sampling points and the central pixel point include:

[0023] According to the gradients of the neighboring pixels of the neighborhood sampling points, calculate the horizontal gradient and the vertical gradient of each neighborhood sampling point. The calculation method of the vertical gradient is the same as that of the horizontal gradient; calculate the gradient direction of the neighborhood sampling point through the horizontal gradient and the vertical gradient of the neighborhood sampling point; in the same way as the calculation method of the gradient direction of the neighborhood sampling point, obtain the gradient direction of the central pixel point;

[0024] The formula for calculating the horizontal gradient of the neighborhood sampling point is: ; In the formula: is the horizontal gradient of the neighborhood sampling point, is the horizontal gradient of the th neighboring pixel of the neighborhood sampling point, is the difference in the horizontal coordinates between the neighborhood sampling point and its th neighboring pixel; is the number of neighboring pixels of the selected neighborhood sampling point; is the difference in pixel values between the neighborhood sampling point and its th neighboring pixel.

[0025] Obtain the horizontal gradient and the vertical gradient through interpolation of the pixels around the neighborhood sampling point. The change in pixel values in the image is usually the combined effect of multiple neighboring pixels, and as the weight coefficient measures the contribution of different neighboring pixels to the gradient of the neighborhood sampling point, ensuring that the neighboring pixels closer to the neighborhood sampling point and with larger pixel value differences contribute more to the gradient, improving the accuracy of gradient calculation and further enhancing the robustness of subsequent feature extraction.

[0026] Optionally, the pixel values of the neighborhood sampling points are calculated using the bilinear interpolation method.

[0027] Optionally, during the training process of the neural network model, input the historical images and the corresponding feature maps at different scales, and output the probabilities of the historical images belonging to different labels; the loss function for training is the Huber loss, and the condition for stopping training is: reaching the preset number of training times or the value of the loss function being less than the preset threshold.

[0028] Optionally, the The method for obtaining the feature map includes: calculating the feature values of each pixel point in the historical image to obtain the corresponding feature map; The method for calculating the feature values is: ;

[0029] Among them, is the feature value of the central pixel point, is the number of neighborhood sampling points, is the direction consistency between the central pixel point and the th neighborhood sampling point, is the th neighborhood sampling point pixel, is the pixel of the central pixel point, .

[0030] Taking the direction consistency between the central pixel point and the neighborhood sampling points as a condition to extract feature values, and dividing weights according to the contributions of different neighborhood point pairs to the feature values, so as to better capture the directional texture features in the image.

[0031] In a second aspect, the present application provides an intelligent detection system for decoration quality, adopting the following technical solutions:

[0032] An intelligent detection system for decoration quality includes: a processor and a memory, and the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent detection method for decoration quality is adopted.

[0033] Generating a computer program for the above-mentioned intelligent detection method for decoration quality and storing it in the memory to be loaded and executed by the processor. Thus, making a system according to the memory and the processor is convenient to use.

[0034] The present application has the following technical effects:

[0035] Using different neighborhood radii to extract features at different scales, reducing the omission of texture details in the image. According to the set difference between the direction sets of different defects at different scales, dividing weights for different scales, and calculating the probabilities of multiple defects in the real-time image according to the weights. In the present application, the application of weights comprehensively considers the differences between the features extracted at different scales, determines the importance of the features extracted at different scales, reduces the error from the true value, and improves the detection accuracy of the decoration wall surface. Description of the Drawings

[0036] Figure 1 It is the flowchart of a method for intelligent detection of decoration quality in this application.

[0037] Figure 2 It is the flowchart of step S3 in a method for intelligent detection of decoration quality in this application.

[0038] Figure 3 It is the flowchart of step S4 in a method for intelligent detection of decoration quality in this application.

[0039] Figure 4 It is the flowchart of step S5 in a method for intelligent detection of decoration quality in this application. Specific implementation mode

[0040] This application embodiment discloses a method for intelligent detection of decoration quality, which extracts LBP feature values of images at multiple scales. The direction consistency between the central pixel point and its neighborhood pixel points in the defect areas corresponding to different defects at the same scale constitutes a direction set. Weights are set for different scales according to the set differences between the direction sets corresponding to different defects at the same scale. After the real-time image is input into the neural network model, the probability that the real-time image belongs to different defects is calculated in combination with the scale weights, thereby improving the accuracy of intelligent detection of decoration quality. Refer to Figure 1 , the method for intelligent detection of decoration quality mainly includes steps S1 - S6.

[0041] S1: Preprocess and label the historical images of the wall surface, where different labels represent different defects.

[0042] In step S1, a high-resolution camera is used to collect historical images during different wall surface decorations. Gaussian filtering is used to remove the noise in the images, and the images are grayscaled. Each historical image is labeled by those skilled in the art (cracks, paint scratches, stains, bulges, normal label, and all except normal are defects).

[0043] S2: Set the scale. Taking any pixel point as the central pixel point and the scale as the radius, a circular trajectory line is obtained, and the pixel points on the circular trajectory line are sampled to obtain neighborhood sampling points.

[0044] The number of neighborhood sampling points is . In this embodiment, bilinear interpolation (a method for interpolating discrete data in a two-dimensional space. It estimates the value of any point by weighted averaging the values of four nearest neighbor points) is used to estimate the pixel values of neighborhood sampling points, so as to facilitate the calculation of the values of neighborhood sampling points.

[0045] S3: At any scale, calculate the overall consistency of the central pixel points within the region where defects of any type are located to obtain the corresponding direction set at this scale;

[0046] Calculate the gradient direction of the central pixel point through the horizontal gradient and vertical gradient of the central pixel point, and calculate the gradient direction of the neighborhood sampling points by calculating the horizontal gradient and vertical gradient of the neighborhood sampling points. Calculate the consistency of the gradient directions between the neighborhood sampling points and the central pixel sampling points through the gradient directions of the neighborhood sampling points and the central pixel sampling points. Calculate the overall consistency of the central pixel point through the direction consistency between the central sampling point and multiple neighborhood sampling points. The overall consistency of the central pixel points within the defect regions corresponding to different defects at different scales forms a direction set, and calculate the difference between the direction sets corresponding to different defects at different scales to obtain the set difference.

[0047] In one embodiment, referring to Figure 2 , step 3 includes steps S30 - S36.

[0048] S30: Edge filling of the grayscale image.

[0049] Since the edge pixels of the image are located on the boundary of the image and their neighborhood pixel points are incomplete, the neighborhood of the edge pixels of the image usually only considers one quadrant. For example, if a central pixel point is located in the upper left corner of the image, then its neighborhood pixel points only exist in the lower right of the central pixel point. In this application, the edge filling method is used to assist in obtaining the features (when extracting features, avoid extracting the features of the filled pixel points); thereby improving the accuracy of intelligent detection of decoration quality.

[0050] The edge filling method is copy filling. The main principle of copy filling is to extend the neighborhood of the edge pixel points beyond the image boundary and fill the extended area by copying the values of the edge pixels. Specifically, if the pixel point is located on the upper edge of the image, copy the pixel value of the upper edge to the extended area for filling. If the pixel point is located on the lower edge of the image, copy the pixel value of the lower edge to the extended area for filling, and so on to complete the filling of the left and right edges of the image.

[0051] S31: Calculate the horizontal gradient and vertical gradient of each neighborhood sampling point according to the gradients of the neighboring pixels of the neighborhood sampling points;

[0052] Calculate the horizontal gradient of the neighborhood sampling point based on the gradients of the neighboring pixels of the neighborhood sampling point. The calculation method is:

[0053] ;

[0054] In the formula: is the horizontal gradient of the neighborhood sampling points, is the horizontal gradient of the th neighboring pixel of the neighborhood sampling points, is the difference in the horizontal coordinates between the neighborhood sampling points and its th neighboring pixel (the difference in the vertical coordinates is used for calculating the vertical gradient when calculating); ); is the number of neighboring pixels of the selected neighborhood sampling points; is the difference in pixel values between the neighborhood sampling points and its th neighboring pixel. The horizontal gradient of the neighboring pixel closer to the neighborhood sampling point has a greater impact on the gradient of the neighborhood sampling point.

[0055] In an image, the change in pixel values is usually the result of the combined action of neighboring pixels. By means of weighted summation, the horizontal gradient calculation formula can comprehensively consider the influence of multiple neighboring pixels, thereby improving the accuracy of horizontal gradient calculation.

[0056] In the same way as the calculation method of the horizontal gradient, the vertical gradient is obtained.

[0057] S32: Calculate the gradient direction of the neighborhood sampling points through the horizontal gradient and vertical gradient of the neighborhood sampling points.

[0058] The calculation method of the gradient direction of the neighborhood sampling points is as follows:

[0059] ;

[0060] In the formula, represents the gradient direction of the th neighborhood sampling point; is the arctangent function, which is used to calculate the angle between the gradient vector and the horizontal axis; is the horizontal gradient of the neighborhood sampling point, is the vertical gradient of the neighborhood sampling point.

[0061] S33: In the same way as the calculation method of the gradient direction of the neighborhood sampling points, obtain the gradient direction of the central pixel point; the specific calculation will not be elaborated here.

[0062] S34: Calculate the direction consistency between the gradient direction of the neighborhood sampling points and the gradient direction of the central pixel point.

[0063] In one embodiment, the calculation formula for direction consistency is:

[0064] ;

[0065] In the formula, represents the central pixel point and the The directional consistency of adjacent neighborhood sampling points is an exponential function indicating the gradient direction of the central pixel point denotes the th adjacent neighborhood sampling point's gradient direction indicating the standard deviation of the gradient direction

[0066] In the formula represents the absolute value of the difference between the gradient direction of the central pixel point and the th adjacent neighborhood sampling point. This difference value is the core part of the extreme directional consistency. The smaller the difference, the closer the gradient directions of the central pixel point and the adjacent neighborhood sampling points are. The standard deviation reflects the degree of change of the gradient direction in the entire image or a certain area. If the gradient direction changes greatly will have a large value; conversely, if the gradient direction changes little will have a small value. Dividing the difference in gradient direction by the standard deviation can standardize the difference value so that it is not affected by the overall change of the gradient direction in the image. The role of the negative sign is to transform the difference value into a decreasing function. The greater the difference, the smaller the value of the expression; the smaller the difference, the greater the value of the expression. Finally, the exponential function maps the difference value to a weight value between 0 and 1, which is used to adjust the contribution of the adjacent neighborhood sampling points to the eigenvalue of the central pixel point

[0067] In the formula the standard deviation of the gradient direction refers to the standard deviation of the gradient directions of the central pixel point and the adjacent neighborhood sampling points

[0068] In another embodiment, the calculation formula for directional consistency is

[0069] ;

[0070] In the formula represents the directional consistency between the central pixel point and the th adjacent neighborhood sampling point is an exponential function indicating the gradient direction of the central pixel point denotes the th adjacent neighborhood sampling point's gradient direction indicating the standard deviation of the gradient direction

[0071] When calculating the directional consistency in this formula, the square root of the difference between the central pixel point and the adjacent neighborhood sampling points is used. This calculation method is smoother and can better handle the small changes in pixel gradients

[0072] S35: Calculation of overall consistency;

[0073] For each pixel point in the defect area under the same type of defect at the same scale, the overall consistency can be calculated by the following formula:

[0074] ;

[0075] In the formula represents the overall consistency of the central pixel point, represents the number of neighborhood sampling points, represents the directional consistency between the pixel point and its th neighborhood sampling point.

[0076] In another embodiment, the formula for the overall consistency of each pixel point in the defect area under the same type of defect at the same scale is:

[0077] ;

[0078] In the formula, represents the overall consistency of a pixel point, represents the number of neighborhood sampling points, represents the directional consistency between the pixel point and its th neighborhood sampling point. represents the minimum value function, represents the maximum value function; represents the directional consistency between the central pixel point and the neighborhood pixel points.

[0079] S36: Constructing a direction set; At one scale, the overall consistency of each central pixel point in the defect area corresponding to one type of defect forms a direction set.

[0080] The defect area includes multiple pixel points. The overall consistency of each pixel point in the defect area constitutes a direction set. Each type of defect includes a direction set at each scale. Then, multiple direction sets are formed for multiple types of defects at different scales.

[0081] S4: Calculate the set difference according to the direction sets corresponding to various types of defects at each scale.

[0082] Referring to Figure 3 , step S4 includes steps S41 - S42.

[0083] S41: Calculate the chi-square test value between the direction sets corresponding to different defects at the same scale.

[0084] The chi-square test value represents the difference between sets in different directions. In other embodiments, it can also be represented by the KL divergence (a measure of the asymmetry between two probability distributions in information theory) and the within-class variance (which refers to the degree of difference between pixel values within the same class), etc.

[0085] S42: Calculate the set difference;

[0086] Calculate the set difference of different defects at this scale according to the chi-square test values between sets in each direction. The set difference can be calculated by the following formula:

[0087] ;

[0088] represents the set difference in the direction sets corresponding to different defect types at a scale. The larger the set difference, the more important the feature map of this scale is. represents the th defect and the th defect between the chi-square test values of the direction sets; is the number of defect types; is the linear normalization function.

[0089] Calculate the sum of the chi-square test values between all direction sets in. Exemplarily, when there are three types of defects (cracks, scratches, bulges), , then three group differences need to be calculated, namely cracks and bulges, cracks and scratches, and scratches and bulges.

[0090] If the features at a certain scale can well distinguish all defects, then the sum of the chi-square test values between all defect pairs will be larger, indicating the importance of this scale for defect detection, thus comprehensively reflecting the ability of the features at this scale to distinguish different defects.

[0091] In another embodiment, the formula for calculating the set difference is:

[0092] ;

[0093] represents the set difference in the direction sets corresponding to different defect types at a scale. The larger the set difference, the more important the feature map of this scale is. represents the th defect and the th defect between the chi-square test values of the direction sets; is the number of types of defects; is an exponential function; is a linear normalization function.

[0094] S5: Train the neural network model to obtain a detection model.

[0095] Refer to Figure 4 , and step S5 includes steps S51 - S52.

[0096] S51: Calculate the LPB eigenvalues of each pixel point in the historical image to obtain the feature map corresponding to the historical image;

[0097] The calculation method is:

[0098] ;

[0099] In the formula is the eigenvalue of the central pixel point, is the number of neighborhood sampling points, is the direction consistency between the central pixel point and the th neighborhood sampling point, is the th neighborhood sampling point pixel, is the pixel of the central pixel point, .

[0100] reflects the consistency of the gradient direction between the neighborhood pixel points and the neighborhood sampling points. The higher the direction consistency, the larger the value. represents the relative contribution of this neighborhood sampling point to the eigenvalue. According to the similarity between the neighborhood sampling point and the central pixel point gradient, dynamically adjust the contribution of each neighborhood sampling point to better capture the directional texture features in the image, such as wall cracks, paint scratches, etc.

[0101] S52: Neural network model training.

[0102] During the training process of the neural network model, the input is the historical image and the corresponding feature map, and the output is the probability that the historical image belongs to different labels; for each scale of The feature map corresponds to a set of label probabilities. F1-score is used to evaluate the model performance during training, and Huber loss (Huber Loss is a loss function that combines the advantages of mean square error and mean absolute error, aiming to improve the robustness of the model to outliers) is selected as the model loss function. Training is stopped when the preset number of training times is reached or the loss function value is less than the preset threshold. In this embodiment, the preset number of training times is 1000 times, and the preset threshold of the loss function is 0.01; the detection model is obtained after the training is completed.

[0103] S6: Input the real-time image into the detection model and output the probability that the real-time image belongs to different labels.

[0104] During the decoration detection process, take photos of the wall in real time, input the previous real-time pictures into the detection model, and output the probability that the real-time pictures belong to different labels; the probability of the labels can be calculated by the following formula:

[0105] ;

[0106] In the formula, represents the probability that the real-time image belongs to this label, is the number of scales, Indicated on the scale Obtained The difference in directional consistency of different defects under the feature map, Representation based on scale Obtained The probability of the model outputting the label when the feature map is used as input.

[0107] In scale The greater the difference in the sets of directions corresponding to different defects, the greater the difference in the sets of directions corresponding to different defects. The more important the feature map is for detecting defects, the more Get the scale The weight of The larger the value, the greater the weight of the scale. When the feature map is used as input, the model outputs the probability of belonging to the label, and the total probability is calculated by weight, and the probability of the real-time image belonging to different labels is obtained, thereby considering features of different scales and improving the accuracy of detection.

[0108] An embodiment of the present application also discloses a decoration quality intelligent detection method system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, an decoration quality intelligent detection method according to the present application is implemented.

[0109] The above system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. Their settings and functions are known in the art and will not be described in detail herein.

[0110] The above are all preferred embodiments of this application. The protection scope of this application is not limited thereby. Therefore, all equivalent changes made according to the structure, shape, and principle of this application shall be covered within the protection scope of this application.

Claims

1. A decoration quality intelligent detection method, characterized in that: The method comprises the following steps: preprocessing and labeling the historical images of the wall, where different labels represent different defects; Set the scale, take any pixel as the center pixel, take the scale as the radius, get the circular trajectory, sample the pixel points on the circular trajectory to get the neighborhood sampling points; At any scale, calculate the overall consistency of the central pixel points in the area where any type of defect is located to obtain the corresponding direction set at the scale; calculate the set difference based on the direction sets corresponding to each type of defect at each scale; Train the neural network model to obtain the detection model; input the real-time image into the detection model and output the probability that the real-time image belongs to different labels; Among them, the calculation formula for the probability of real-time images belonging to different labels is: ; represents the probability that the real-time image belongs to this label, is the number of selected scales, Indicated on the scale Obtained The set difference between the direction sets corresponding to different defects under the feature map, It indicates the probability that the historical images output during the training of the neural network model belong to different labels. When the probability meets the preset threshold corresponding to any label, it indicates that there is a defect corresponding to the label on the wall. The overall consistency of the central pixel is calculated based on the directional consistency between the central pixel and multiple neighboring sampling points; The step of calculating the directional consistency includes: calculating the gradient directions of the neighborhood sampling points and the central pixel point; calculating the directional consistency according to the gradient directions of the neighborhood sampling points and the central pixel point; The calculation formula for directional consistency is: ; Indicates the center pixel and the The direction consistency of the neighborhood sampling points, Indicates The exponential function with base , Represents the gradient direction of the central pixel, Indicates The gradient direction of the neighborhood sampling points is Indicates the standard deviation of the gradient direction; Historical images corresponding to The method for obtaining the feature map includes: calculating the The feature value is obtained Feature map; The eigenvalue is calculated as: ； in, is the center pixel Eigenvalues, is the number of neighborhood sampling points, is the center pixel and the The direction consistency of the neighborhood sampling points, It is The pixels of the neighboring sampling points, is the pixel at the center pixel, .

2. The intelligent decoration quality detection method according to claim 1 is characterized in that: in, The formula for calculating the set difference is: ; represents the set difference of different defects at a scale, Indicates Defects and The chi-square test value between the direction sets of the defects; is a linear normalization function; is the number of defect types.

3. The intelligent decoration quality detection method according to claim 2 is characterized in that: The method for constructing a direction set includes: selecting multiple scales for the defect area, calculating the overall consistency of each pixel in the defect area at the multiple scales; the overall consistency of multiple pixels in the defect area constitutes a direction set.

4. The intelligent decoration quality detection method according to claim 1 is characterized in that: The calculation steps of the overall consistency are: calculating the directional consistency between the central pixel point and the corresponding different neighborhood sampling points; calculating the overall consistency of the central pixel point according to the directional consistency between the central pixel point and multiple neighborhood sampling points; The overall consistency is calculated as: ; In the formula, Indicates the overall consistency of the central pixel, represents the number of neighborhood sampling points, Represents a pixel and its The direction consistency of the neighborhood sampling points.

5. The intelligent decoration quality detection method according to claim 1 is characterized in that: The steps of calculating the gradient directions of the neighborhood sampling points and the central pixel point include: According to the gradient of the neighboring pixels of the neighborhood sampling point, the horizontal gradient and vertical gradient of each neighborhood sampling point are calculated. The calculation method of the vertical gradient is the same as that of the horizontal gradient. The gradient direction of the neighborhood sampling point is calculated by the horizontal gradient and vertical gradient of the neighborhood sampling point. The gradient direction of the central pixel point is obtained in the same way as that of the gradient direction of the neighborhood sampling point. The calculation formula of the horizontal gradient of the neighborhood sampling point is: ; Where: is the horizontal gradient of the neighborhood sampling point, is the first sampling point in the neighborhood The horizontal gradient of neighboring pixels, is the neighborhood sampling point and its The difference between adjacent pixels in horizontal coordinates; The number of neighboring pixels of the selected neighborhood sampling point; is the neighborhood sampling point and its The pixel value difference between adjacent pixels.

6. The intelligent decoration quality detection method according to claim 1 is characterized in that: The pixel values ​​of the neighborhood sampling points are calculated using the bilinear interpolation method.

7. The intelligent decoration quality detection method according to claim 1 is characterized in that: During the training of the neural network model, historical images and corresponding images at different scales are input. Feature map, output the probability that the historical image belongs to different labels; the loss function of training is Huber loss, and the condition for stopping training is: reaching the preset number of training times or the loss function value is less than the preset threshold.

8. An intelligent decoration quality detection system, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent detection method for decoration quality according to any one of claims 1 to 7 is implemented.

Images