Surface defect detection method

Abstract

To provide a surface defect detection method that can determine abnormal points in an image of an object to be detected in an intuitive, high-speed and accurate manner.SOLUTION: A method comprises the steps of: obtaining an image of an object to be detected; inputting the image of the object to be detected into a pre-trained neural network model and obtaining a plurality of detection boxes described with detection scores; mapping each of the detection scores described on the plurality of detection boxes into a two-dimensional array; fitting the detection scores and then substituting the fitted detection scores to an RGB polynomial regression formula; mapping each detection score described in the two-dimensional array to a three-dimensional RGB value using the RGB polynomial regression formula to which the detection scores are substituted; generating a heatmap based on the three-dimensional RGB value, and superposing the heatmap onto the image of the object to be detected so as to display abnormal points displayed in the heatmap.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to the technical field of image detection, and in particular to a method for detecting surface defects. [Background technology]

[0002] A heatmap is a visualization tool that uses color to represent data density. It displays data in a two-dimensional space in the form of hot points with different colors, and is used to show the distribution and concentration of data. This helps observe trends and patterns in the data. A heatmap is generated by mapping numerical data to colors. Typically, larger data values ​​are represented by darker colors, and smaller data values ​​are represented by lighter colors. It can be applied to detecting anomaly and fault types in various products.

[0003] Currently, the common method for detecting wafer anomalies is to present the results using detection boxes, which is a good method, but if there are too many detection boxes, it is quite difficult to view the detection results.If multiple detection boxes appear on one diagram and the detection boxes overlap, such a detection result diagram will be quite complicated and it will be difficult to judge the detection results. Summary of the Invention [Problem to be solved by the invention]

[0004] In view of this, the present invention proposes a surface defect detection method that aims to solve the problem that the accuracy of conventional surface defect detection methods is not high. [Means for solving the problem]

[0005] The present invention provides a surface defect detection method, comprising: acquiring an image of a test object, inputting the image of the test object into a pre-trained neural network model, and obtaining a plurality of detection boxes with detection scores; Mapping the detection scores written in the plurality of detection boxes into a two-dimensional array, respectively; fitting the detection scores and substituting the fitted detection scores into an RGB polynomial regression equation; Mapping each detection score described in the two-dimensional array to a three-dimensional RGB numerical value using an RGB polynomial regression equation into which the detection scores are substituted; generating a heat map based on the three-dimensional RGB numerical values, and superimposing the heat map on the image of the detected object to present abnormal points displayed in the heat map.

[0006] Furthermore, in the surface defect detection method, the method of training the neural network model includes: capturing images of the detected object having anomalies, each of which is a multi-channel RGB image; manually marking anomalies in each of the object images and surrounding the anomalies in the image using detection boxes, each of which contains the anomalies of the object; constructing a dataset by screening images containing outliers and dividing the dataset into a training set and a test set; The method includes constructing a neural network model based on the EfficientNet-SSD algorithm and training it by inputting image data in a training set to generate a neural network model based on the EfficientNet-SSD algorithm, using the autonomous learning ability of the EfficientNet-SSD model to perform feature extraction on the image to obtain a plurality of feature images with different resolutions, generating a plurality of detection boxes with different resolutions on these feature images based on the plurality of feature images with different resolutions, outputting the probability that each pixel point in the plurality of detection boxes with different resolutions is an abnormal point, and evaluating the neural network model based on the EfficientNet-SSD algorithm using image data in a test set to obtain an optimized neural network model.

[0007] The surface defect detection method further includes optimizing the neural network model by using the generated EfficientNet-SSD algorithm-based neural network model to adjust the positions of the generated detection boxes with different combinations of resolution and position so that the detection boxes are closer to the true boxes marked with abnormal points, thereby obtaining an optimized EfficientNet-SSD algorithm-based neural network model, and evaluating the optimized EfficientNet-SSD algorithm-based neural network model using image data from a test set to obtain the optimized neural network model.

[0008] Furthermore, in the surface defect detection method, the step of mapping the detection scores written in the plurality of detection boxes to a two-dimensional array includes: Establishing a two-dimensional matrix of the same size according to the size of the original detected image; Substituting a detection score for each pixel point in the detection box into a corresponding element of the two-dimensional matrix according to the coordinates of each pixel point in the detection box; Repeating the above steps includes multiplying the detection score of each pixel point in each detection box by the corresponding element of the two-dimensional matrix, and thus each element value in the two-dimensional matrix is ​​the cumulative sum of the detection scores of the corresponding pixel points in each detection box.

[0009] Furthermore, in the surface defect detection method, mapping each detection score described in the two-dimensional array to a three-dimensional RGB numerical value includes: The method includes generating an RGB matrix according to the size of the original detected image, and calculating the detection scores of each element in the two-dimensional matrix using R, G, and B polynomial regression equations, respectively, to obtain the numerical values ​​of each RGB channel in the RGB matrix.

[0010] Furthermore, in the surface defect detection method, substituting the fitted detection score into the RGB polynomial regression equation Customizing the relationship between color and detection score, Using the relationship between the detection score and the RGB value of the color, the polynomial regression equation f(x) = a n x n +a n-1 x (n-1) +…+a2x 2 +a1x+a o and solving the polynomial regression equation:

[0011] R=1.06105X 7 -3.12570X 6 +3.40036X 5 -1.68407X 4 +3.86728X 3 -3.68337X 2 +1.01210X 1 +1.38694, G=2.12211X 7 -7.73688X 6 +1.12571X 5 -8.26694X 4 +3.14362X 3 -5.52839X 2 +3.37545X 1 -1.98135, B=-2.12211X 7 +8.29394X 6 -1.29283X 5 +1.01249X 4 -4.07428X 3 +7.41653X 2 -3.57669X 1 +7.92540.

[0012] Furthermore, in the surface defect detection method, the relationship between the color and the detection score is as follows: The RGB values ​​corresponding to a detection score of 1 are 255, 0, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.875 are 255, 127.5, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.75 are 255, 255, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.625 are 127.5, 255, and 0, respectively. The RGB values ​​corresponding to a detection score of 0.5 are 0, 255, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.375 are 0, 127.5, and 127.5, respectively. When the detection score is 0.25, the corresponding RGB values ​​are 0, 0, and 255, respectively. The corresponding RGB values ​​for a detection score of 0.125 are 0, 0, and 127.5, respectively. The RGB values ​​corresponding to a detection score of 0 are 0, 0, 0 respectively.

[0013] Furthermore, in the surface defect detection method, a specific coefficient R is added to the RGB polynomial regression equation into which the fitted detection score is substituted. 2 Score is an index that evaluates the suitability of the model. JPEG2025178108000002.jpg13170

[0014] Furthermore, the surface defect detection method further includes normalizing each detection score written in the two-dimensional array and then mapping it to (0 to 1), and the normalization formula is as follows: JPEG2025178108000003.jpg24170

[0015] Furthermore, in the above-mentioned surface defect detection method, the captured image showing the object to be detected is preprocessed and then input into a neural network model based on a pre-trained EfficientNet-SSD algorithm, and each pixel point in the detection box with multiple combinations of resolution and position generates an abnormal point detection score. [Effects of the Invention]

[0016] In the surface defect detection method of the present invention, an image showing an object to be detected is preprocessed and then input into a pre-trained neural network model. Each pixel point in the detection box for multiple combinations of resolution and position generates a detection score for an abnormality. The detection scores for the corresponding pixel points are then mapped into a two-dimensional array. A polynomial regression equation is used to convert the two-dimensional detection scores into three-dimensional RGB, and the two-dimensional array with the detection scores is mapped into three-dimensional RGB numerical values. This generates a heat map, which is then superimposed on the original image of the object to enable intuitive, rapid, and accurate judgment of abnormalities in the image of the object to be detected. [Brief explanation of the drawings]

[0017] Various other benefits and advantages will become apparent to those skilled in the art upon reading the following detailed description of the preferred embodiments. The drawings are used only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. Furthermore, like parts are designated by like reference numerals throughout the drawings. [Figure 1] 1 is a flowchart of a surface defect detection method according to an embodiment of the present invention. [Figure 2] 1 is a diagram showing the relationship between detection scores and colors defined in a surface defect detection method according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0018] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. While the drawings illustrate exemplary embodiments of the present disclosure, it should be understood that the present disclosure should not be limited to the embodiments set forth herein, but may be embodied in various forms. Rather, these embodiments are provided to provide a more complete understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. Furthermore, where not inconsistent, embodiments and features of the embodiments of the present invention may be combined with each other. Hereinafter, the present invention will be described in detail with reference to the drawings and in connection with the embodiments.

[0019] Referring to FIG. 1, the surface defect detection method according to an embodiment of the present invention includes the following steps.

[0020] Step 1: Obtain an image of the object to be detected, input the image of the object to a pre-trained neural network model, and obtain a number of detection boxes with detection scores.

[0021] Specifically, the captured image containing the object to be detected is preprocessed and then input into a neural network model based on a pre-trained EfficientNet-SSD algorithm, and each pixel point in the detection box for multiple combinations of resolution and position generates an anomaly detection score.

[0022] In this embodiment, the detected object may be a semiconductor device, more specifically, a wafer. The photographed tool may be a high-resolution camera, such as a 25-megapixel camera. Preprocessing may involve cutting the image to a certain resolution, for example, a multi-channel RGB image of 448*448 pixels (or other resolutions such as 800*800). The detection score is the probability that a pixel consisting of any one pixel point and another pixel point in the image is an abnormal point. Here, the combination of resolution and position means that each detection box has a different size (resolution) and a different position in the image. Since the detection boxes are not offset with an interval, each detection box may have overlapping partial areas.

[0023] In this embodiment, each pixel point in the detection box corresponds to one detection score.

[0024] Step 2: The detection scores written in the plurality of detection boxes are each mapped to a two-dimensional array.

[0025] Specifically, in step 1, according to the size of the original detected image (the resolution in the length direction is L, the resolution in the width direction is W, and the image size is L*W), a two-dimensional matrix of the same size is established, and the two-dimensional matrix has L rows and W columns.

[0026] Step 3: According to the coordinates of each pixel point in the detection box, the detection score of each pixel point in the detection box is assigned to the corresponding element of the two-dimensional matrix.

[0027] The above steps are repeated to accumulate the detection scores of each pixel point in each detection box into the corresponding elements of the two-dimensional matrix.

[0028] Thus, each element value in the two-dimensional matrix is ​​the cumulative sum of the detection scores of the corresponding pixel points in each detection box.

[0029] It should be noted that when the detection score value of each pixel point in the detection box is between 0 and 1, the accumulated cumulative sum value is between 0 and N, where N is a known quantity and is the cumulative sum of the detection scores of the detection box under the same pixel. In this way, the detection score data can be further normalized and remapped to between 0 and 1.

[0030] Step S4: fitting the detection scores, and substituting the fitted detection scores into the RGB polynomial regression equation.

[0031] In a specific implementation, an RGB matrix is ​​generated according to the size of the original detected image, and the detection scores of each element in the two-dimensional matrix in step 2 are calculated using R, G, and B polynomial regression equations respectively to obtain the values ​​of each RGB channel in the RGB matrix.

[0032] Specifically, the polynomial regression equation is f(x)=a n x n +a n-1 x (n-1) +…+a2x 2 +a1x+a o is.

[0033] In this embodiment, the polynomial is obtained as follows: The relationship between custom colors and detection scores is shown in Figure 2.

[0034] The RGB values ​​corresponding to a detection score of 1 are 255, 0, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.875 are 255, 127.5, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.75 are 255, 255, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.625 are 127.5, 255, and 0, respectively. The RGB values ​​corresponding to a detection score of 0.5 are 0, 255, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.375 are 0, 127.5, and 127.5, respectively. The corresponding RGB values ​​for a detection score of 0.25 are 0, 0, and 255, respectively. The corresponding RGB values ​​for a detection score of 0.125 are 0, 0, and 127.5, respectively. The RGB values ​​corresponding to a detection score of 0 are 0, 0, 0 respectively.

[0035] Table 1 below shows the results of running the program 20 times using the data shown in Figure 2.

[0036] JPEG2025178108000004.jpg117170

[0037] As can be seen from the table above, when N=7, R2 is already equal to 1, and when N=8~20, it is also equal to 1. However, the more complex the polynomial regression equation, the longer the calculation time of the system, so when N=7, it is considered to be already in the optimal solution. Therefore, when N=7, the solution of polynomial regression can be obtained by the following formula: R=1.06105X 7 -3.12570X 6 +3.40036X 5 -1.68407X 4 +3.86728X 3 -3.68337X 2 +1.01210X 1 +1.38694, G=2.12211X 7 -7.73688X 6 +1.12571X 5 -8.26694X 4 +3.14362X 3 -5.52839X 2 +3.37545X 1 -1.98135, B=-2.12211X 7 +8.29394X 6 -1.29283X 5 +1.01249X 4 -4.07428X 3 +7.41653X 2 -3.57669X 1 +7.92540.

[0038] The R2Score determines the fit of the model to the data. Specific coefficient (R2Score): R2 is an index that evaluates the fit of the model, showing the degree to which the model can interpret changes in the data. The closer the R2 value is to 1, the better the model fits the data.

[0039] When a formula specific to mapping detection scores to R, G, and B is generated using polynomial regression, the R-Squared calculation formula is as follows:

[0040] JPEG2025178108000005.jpg14170

[0041] In this embodiment, the RGB polynomial regression equation can be used to substitute the fitted detection score to obtain the RGB value corresponding to any pixel point. For example, if the detection score of a pixel point in a detection box is 0.32 (32%), the above R, G, and B polynomial regression equation can be used to convert 0.32 into three-dimensional data with R=0, G=70, and B=195, which can then be displayed or presented in the corresponding color after rendering.

[0042] If the detection score of a pixel point in a detection box is 0.79 (79%), then the above polynomial regression formula will convert 0.79 into three-dimensional data with R=255, G=227, and B=0, and the corresponding color will be displayed after rendering.

[0043] Step S5: Generate a heat map based on the three-dimensional RGB numerical values, and superimpose the heat map on the image of the detected object to present the abnormal points displayed in the heat map.

[0044] Therefore, by using a polynomial regression equation, data with a detection score between 0 and 1 (0-100%) can be converted into the corresponding R, G, and B parameters. In this way, different colors can be displayed for different detection results (e.g., red indicates a high probability that this point is an anomaly), and presented as a heat map.

[0045] As can be seen from the above, in the present invention, the relationship between the abnormal point detection score and RGB is defined to identify the constant parameters of the polynomial regression equation, an equation for mapping the detection score to R, G, and B is generated using the polynomial regression equation, the detection scores within the detection box are accumulated by integrating them, a two-dimensional matrix of the accumulated detection scores is generated, the detection score data is further normalized and remapped between 0 and 1, and then the R, G, and B values ​​are recalculated using the polynomial regression equation, and finally a heat map is generated that is advantageous for identifying abnormal points.

[0046] That is, based on the size of each detection box, the detection scores of the corresponding pixel points are mapped into a two-dimensional array, and a polynomial regression equation is used to convert the two-dimensional detection scores into three-dimensional RGB, so that the two-dimensional array with the detection scores is remapped between 0 and 1 to generate a three-dimensional array of RGB values, which are then converted into a heat map based on the three-dimensional array of RGB values, and the heat map result is superimposed and presented on the original photographed image of the detected object.

[0047] In this way, by observing the heat map, a person can determine abnormal points on the object to be detected based on the heat map.

[0048] As can be seen from the above examples, in the present invention, an image depicting an object to be detected is pre-processed and then input into a pre-trained neural network model. Each pixel point in the detection box for multiple combinations of resolution and position generates an abnormality detection score. The detection scores for the corresponding pixel points are then mapped into a two-dimensional array. A polynomial regression equation is used to convert the two-dimensional detection scores into three-dimensional RGB, and the two-dimensional array with the detection scores is then mapped into three-dimensional RGB numerical values. This generates a heat map, which is then superimposed on the original image of the object to intuitively, quickly, and accurately determine abnormalities in the image of the object to be detected. This is also advantageous for identifying areas with defects on the surface of the object to be detected.

[0049] In the above embodiment, the selected neural network model is EfficientNet-SSD, and the training process of this neural network model based on the EfficientNet-SSD algorithm is as follows:

[0050] Step 1: Take images of the object to be detected with abnormal points, each of which is a multi-channel RGB image.

[0051] Specifically, the object to be detected may be a wafer, and a 25 megapixel camera may be used to capture images of multiple wafers with the same abnormality or multiple wafers with different abnormalities, and then the images may be cut into multi-channel RGB images with a resolution of 448*448 (or other resolutions such as 800*800).

[0052] In a specific implementation, each wafer having an abnormality is placed on a detection table, and a 25 megapixel camera is used to photograph the wafer, and the image is stored in a designated folder.

[0053] Step 2: Manually mark the abnormal points in each of the object images, and use detection boxes to enclose the abnormal points in the images, each of which contains the abnormal points of the object.

[0054] Specifically, the detection box may be selected as a rectangular box, an oval box, or the like.

[0055] Step 3: Screen the images containing abnormal points to construct a dataset, and divide the dataset into a training set and a test set. For example, the training set may be 80% of the dataset, and the other 20% of the data in the dataset is the test set.

[0056] Step 4: Construct a neural network model based on the EfficientNet-SSD algorithm and train it using image data from a training set to generate a neural network model based on the EfficientNet-SSD algorithm. The neural network model based on the EfficientNet-SSD algorithm is used to predict a detection score after detecting detection boxes corresponding to each of multiple resolution and position combinations in the input image, where the resolution and position of the detection box in the image are adaptively determined based on the sample set, training set, and training accuracy during the training process of the neural network model based on the EfficientNet-SSD algorithm. After detecting detection boxes corresponding to each resolution and position combination in the image, the detection score is used to indicate the probability that each corresponding pixel point is an abnormal point.

[0057] To make the neural network model more complete, this embodiment may further include step 5, in which the generated EfficientNet-SSD algorithm-based neural network model is used to adjust the positions of the generated detection boxes for different resolution and position combinations, so that these detection boxes are closer to the true boxes marked with abnormal points, thereby optimizing the neural network model, thereby obtaining an optimized EfficientNet-SSD algorithm-based neural network model, and evaluating the optimized EfficientNet-SSD algorithm-based neural network model using image data from a test set, thereby obtaining the optimized neural network model.

[0058] In this embodiment, the neural network model is optimized using the standard momentum optimization algorithm (Momentum) for the optimizer.

[0059] The specific implementation process of the embodiment of the present invention is as follows: The object to be detected is placed on the detection platform, and the shooting path of the detection camera (for example, a 25-megapixel camera can be selected) is set to capture an image of the object to be detected. The image is input into a pre-trained neural network model. The neural network model outputs multiple detection boxes with detection scores, generates a two-dimensional array based on the size of the captured image, maps the detection scores to the two-dimensional array according to the size of the detection boxes, generates a three-dimensional array according to the size of the captured image, obtains a polynomial regression equation to convert the detection scores into three-dimensional RGB, normalizes the two-dimensional array with detection scores, stores the result back into a two-dimensional array, inputs the normalized two-dimensional array of detection scores into a three-dimensional RGB polynomial regression equation to obtain three-dimensional RGB numerical values, stores the result back into the previously generated three-dimensional array, and generates a heat map based on the RGB values ​​of the three-dimensional array, and superimposes the heat map on the original detection image.

[0060] The two-dimensional array normalization formula is as follows: JPEG2025178108000006.jpg25170

[0061] To summarize, in an embodiment of the present invention, an image of an object to be detected is input into a pre-trained neural network model to obtain and define multiple detection boxes each containing a detection score. The detection score of each pixel point in each detection box is then multiplied by the corresponding element of the two-dimensional matrix, so that each element value in the two-dimensional matrix is ​​the cumulative sum of the detection scores of the corresponding pixel points in each detection box. The detection scores are then fitted and substituted into an RGB polynomial regression equation. The RGB polynomial regression equation into which the detection scores have been substituted is used to map each detection score listed in the two-dimensional array into a three-dimensional RGB numerical value. Finally, a heat map is generated that is useful for identifying abnormality locations, making the detection results more intuitive and accurate. Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention intends to include these modifications and variations.

Claims

1. 1. A surface defect detection method, comprising: acquiring an image of a test object, inputting the image of the test object into a pre-trained neural network model, and obtaining a plurality of detection boxes with detection scores; Mapping the detection scores written in the plurality of detection boxes into a two-dimensional array, respectively; fitting the detection scores and substituting the fitted detection scores into an RGB polynomial regression equation; Mapping each detection score described in the two-dimensional array to a three-dimensional RGB numerical value using an RGB polynomial regression equation into which the detection scores are substituted; generating a heat map based on the three-dimensional RGB numerical values, overlaying the heat map on the image of the detected object, and presenting abnormal points displayed in the heat map.

2. The method for training the neural network model includes: capturing images of the detected object having anomalies, each image being a multi-channel RGB image; manually marking anomalies in each of the object images and surrounding the anomalies in the image using detection boxes, each of which contains the anomalies of the object; constructing a dataset by screening images containing outliers and dividing the dataset into a training set and a test set; 2. The surface defect detection method of claim 1, further comprising: constructing a neural network model based on the EfficientNet-SSD algorithm and training it by inputting image data in a training set to generate a neural network model based on the EfficientNet-SSD algorithm; using the autonomous learning ability of the EfficientNet-SSD model to perform feature extraction on the image to obtain a plurality of feature images with different resolutions; generating a plurality of detection boxes with different resolutions on the feature images based on the plurality of feature images with different resolutions; outputting a probability that each pixel point in the plurality of detection boxes with different resolutions is an abnormal point; and evaluating the neural network model based on the EfficientNet-SSD algorithm using image data in a test set to obtain an optimized neural network model.

3. 3. The surface defect detection method of claim 2, further comprising: optimizing the neural network model by using the generated EfficientNet-SSD algorithm-based neural network model to adjust the positions of the generated detection boxes of different combinations of resolution and position so that the detection boxes are closer to true boxes marked with abnormal points, thereby obtaining an optimized EfficientNet-SSD algorithm-based neural network model; and evaluating the optimized EfficientNet-SSD algorithm-based neural network model using image data of a test set to obtain the optimized neural network model.

4. The step of mapping each of the detection scores described in the plurality of detection boxes to a two-dimensional array includes: Establishing a two-dimensional matrix of the same size according to the size of the original detected image; Substituting a detection score for each pixel point in the detection box into a corresponding element of the two-dimensional matrix according to the coordinates of each pixel point in the detection box; 2. The surface defect detection method according to claim 1, further comprising repeating the above steps and multiplying the detection score of each pixel point in each detection box by a corresponding element of the two-dimensional matrix, wherein each element value in the two-dimensional matrix is ​​a cumulative sum of the detection scores of the corresponding pixel points in each detection box.

5. Mapping each detection score listed in the two-dimensional array to a three-dimensional RGB numerical value includes: The surface defect detection method according to claim 3, further comprising generating an RGB matrix according to the size of the original detected image, and calculating the detection score of each element in the two-dimensional matrix using R, G, and B polynomial regression equations, respectively, to obtain the numerical value of each RGB channel in the RGB matrix.

6. Substituting the fitted detection scores into the RGB polynomial regression equation gives: Customizing the relationship between color and detection score, Using the relationship between the detection score and the RGB value of the color, a polynomial regression equation f(x) = a ｎ x ｎ +a ｎ－１ x （ｎ－１） +...+a ２ x ２ +a １ x + a ｏ and solving the following polynomial regression equation: R=1.06105X ７ -3.12570X ６ +3.40036X ５ -1.68407X ４ +3.86728X ３ -3.68337X ２ +1.01210X １ +1.38694、 G=2.12211X ７ -7.73688X ６ +1.12571X ５ -8.26694X ４ +3.14362X ３ -5.52839X ２ +3.37545X １ -1.98135、 B=-2.12211X ７ +8.29394X ６ -1.29283X ５ +1.01249X ４ -4.07428X ３ +7.41653X ２ -3.57669X １ +7.92540。

7. The relationship between the color and the detection score is as follows: The RGB values ​​corresponding to a detection score of 1 are 255, 0, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.875 are 255, 127.5, and 0, respectively. The RGB values ​​corresponding to a detection score of 0.75 are 255, 255, and 0, respectively. The corresponding RGB values ​​for a detection score of 0.625 are 127.5, 255, and 0, respectively. The RGB values ​​corresponding to a detection score of 0.5 are 0, 255, and 0, respectively. When the detection score is 0.375, the corresponding RGB values ​​are 0, 127.5, and 127.5, respectively. When the detection score is 0.25, the corresponding RGB values ​​are 0, 0, and 255, respectively. When the detection score is 0.125, the corresponding RGB values ​​are 0, 0, and 127.5, respectively.

7. The surface defect detection method according to claim 6, wherein the RGB values ​​corresponding to a detection score of 0 are 0, 0, 0, respectively.

8. In the RGB polynomial regression equation into which the fitted detection score is substituted, a specific coefficient R ２ Score is an index for evaluating the suitability of the model,

9. The method further includes normalizing each detection score described in the two-dimensional array and then mapping it to (0 to 1), wherein the normalization formula is as follows:

10. The surface defect detection method according to claim 1, characterized in that the captured image showing the detected object is pre-processed and then input into a neural network model based on a pre-trained EfficientNet-SSD algorithm, and each pixel point in the detection box for multiple combinations of resolution and position generates an abnormal point detection score.

Images