High-precision edge extraction method based on image automatic processing

Abstract

The application relates to the field of machine vision and image recognition, and discloses a high-precision edge extraction method based on image automatic processing, which comprises the following steps: a standardized image system is constructed, and an initial standardized wood template image is acquired through the standardized image system; the initial standardized wood template image is subjected to multimodal fusion pretreatment; an Otsu algorithm is used to determine a high threshold value and a low threshold value, an initial edge image and an edge gradient image are generated after edge detection of a binary image; then, the initial edge image is input into a pre-trained lightweight adaptive U-Net model, and an edge probability image is output; the edge probability image is sequentially subjected to morphological opening operation denoising processing, B-spline curve smoothing fitting processing and coordinate system calibration processing, and wood template edge coordinates are obtained; the average error of the wood template edge coordinates and standard template edge coordinates is calculated. The application realizes accurate conversion of wood template edge from an image to actual physical coordinates.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine vision and image recognition, and particularly relates to a high-precision edge extraction method based on image automatic processing. BACKGROUND

[0002] In building construction, wooden templates are used for concrete pouring and forming, and the edge precision directly determines the dimensional accuracy of the concrete member. If there is deviation in the edge, it is easy to cause concrete leakage and unqualified member forming. In the field of furniture manufacturing, the straightness of the wooden template edge is a key factor affecting the quality of plate splicing and edge sealing. The traditional manual measurement method is low in efficiency (single template measurement time ≥ 3 minutes) and large in error (average error ± 0.5 mm), and is affected by human experience, which is difficult to meet the requirements of "millimeter-level precision" and "batch automatic detection" in modern production.

[0003] The existing image recognition technology faces three core problems in wooden template edge extraction: firstly, the wooden template surface has natural wood grain, knots, color difference and other texture characteristics, which easily forms a pseudo-edge and interferes with the detection of the real edge; secondly, the light conditions in the industrial scene are complex, and problems such as shadows and strong light reflections cause the contrast of the edge region to decrease, making it difficult to identify weak edge signals; thirdly, traditional edge detection algorithms (such as Canny algorithm and Sobel algorithm) use fixed thresholds and single filtering methods, which cannot adaptively adjust parameters to match wooden templates of different materials and different defect states, making it difficult to balance edge integrity and noise suppression. If the threshold is too high, weak edges are easily missed; if the threshold is too low, pseudo-edge noise cannot be effectively filtered, and edge positioning accuracy cannot be broken through ± 0.3 mm.

[0004] Some existing technologies attempt to introduce deep learning models (such as U-Net and FCN) to improve edge extraction accuracy, but have two major defects: first, the preprocessing process is not optimized for the texture characteristics of wooden templates, and the original image is directly input into the model, which causes the model to process a large amount of redundant texture information, resulting in low inference efficiency (single frame processing time ≥ 80 ms); second, the model training uses a general loss function, which does not consider the imbalance problem of samples with low edge pixel ratio of wooden templates, and the intersection over union (IoU) of edge detection is generally lower than 0.8, making it difficult to meet the high-precision detection requirements.

[0005] Therefore, there is an urgent need for an edge extraction method that can adaptively eliminate the interference of wooden template texture, enhance weak edge signals, and balance accuracy and efficiency to solve the deficiencies of existing technologies. SUMMARY

[0006] The application provides a high-precision edge extraction method based on image automatic processing, realizes automatic and high-precision extraction of wood template edges, reduces edge positioning error, improves processing speed, has strong anti-interference ability, and is suitable for wood template detection scenes of different materials, illumination and defect types.

[0007] The application is implemented by the following technical solutions:

[0008] A high-precision edge extraction method based on image automatic processing comprises the following steps:

[0009] A standardized image system is constructed, and an initial standardized wood template image is acquired through the standardized image system;

[0010] The initial standardized wood template image is subjected to multi-modal fusion preprocessing, which comprises sequentially performing adaptive gamma illumination normalization processing, texture direction weighted mean filtering processing, Laplacian pyramid edge enhancement processing and Sauvola local adaptive binaryzation processing on the initial standardized wood template to obtain a binaryzation image, wherein the adaptive gamma illumination normalization processing comprises four sub-steps of image gray feature analysis, texture complexity calculation and basic gamma value determination, edge region priority adjustment and gamma value optimization, gamma value smooth transition and illumination normalization execution.

[0011] An Otsu algorithm is used to determine a high threshold value and a low threshold value After edge detection is performed on the binaryzation image, disconnected edge segments with a distance of px and consistent gradient directions are linearly connected to generate a preliminary edge image and an edge gradient image; then the preliminary edge image is input into a pre-trained lightweight adaptive U-Net model to output an edge probability map; the lightweight adaptive U-Net model uses a depth separable convolution to construct an encoder, fuses features through a skip connection, and is trained by using a dynamic weight joint loss function, wherein 、 is a positive number.

[0012] The edge probability map is sequentially subjected to morphological opening operation denoising processing, B-spline curve smooth fitting processing and coordinate system calibration processing to obtain wood template edge coordinates.

[0013] The average error of the wood template edge coordinates and standard template edge coordinates is calculated, if the average error is not greater than a first threshold value, the edge coordinates and an edge label visual image are output, if the average error is greater than the first threshold value or the edge detection intersection over union is less than a second threshold value, a retake instruction or an alarm signal is triggered.

[0014] As an optimization, the specific process of constructing a standardized image system and acquiring an initial standardized wood template image through the standardized image system is as follows:

[0015] Constructing a standardized image acquisition system comprising a pixel industrial camera, a fixed-focus lens, a ring-shaped light source, and a standard checkerboard calibration component;

[0016] Setting the shooting parameters of object distance, shutter speed, and ISO;

[0017] Calibrating the camera intrinsic parameters using a standard checkerboard before each batch of detection to obtain an intrinsic parameter matrix K;

[0018] Shooting a wooden template image to obtain an initial standardized wooden template image.

[0019] As an optimization, the specific process of sequentially performing adaptive gamma illumination normalization on the initial standardized wooden template is as follows:

[0020] Step A, image gray feature analysis:

[0021] Step A1, calculating the gray histogram of the initial standardized wooden template image, counting the distribution of gray values of all pixels in the initial standardized wooden template image, and determining the % quantile and % quantile , dividing the initial standardized wooden template image into dark regions, normal regions, and bright regions, wherein the pixel gray values of the dark regions are , the pixel gray values of the normal regions are , and the pixel gray values of the bright regions are , , positive and ;

[0022] Step A2, taking 3x3 as a first local window, traversing each pixel of the initial standardized wooden template image, and using a Sobel operator to calculate the x-direction brightness gradient and the y-direction brightness gradient of the first local window, respectively, and calculating the brightness gradient value of each window by the formula ;

[0023] Step A3, according to the distribution characteristics of the brightness gradient value , all brightness gradient values The gradient is divided into low gradient area, medium gradient area and high gradient area, and the gradient area is superimposed with the dark area, normal area and bright area divided in step A1 to form 9 fine partitions, namely dark-low gradient area, dark-medium gradient area, dark-high gradient area, normal-low gradient area, normal-medium gradient area, normal-high gradient area, bright-low gradient area, bright-medium gradient area and bright-high gradient area, wherein the luminance gradient value of the low gradient area is , the luminance gradient value of the medium gradient area is , and the luminance gradient value of the high gradient area is , wherein , is a positive number, and ;

[0024] Step B, texture complexity calculation and basic gamma value determination:

[0025] Step B1, taking 5*5 as the second local window, traversing each pixel in the 9 fine partitions, calculating the texture contrast Contrast and texture entropy value Entropy in each second local window through the gray level co-occurrence matrix GLCM, wherein the texture contrast Contrast is used to reflect the clarity of the texture in the second local window, and the texture entropy value Entropy is used to reflect the random complexity of the texture in the second local window;

[0026] Step B2, the texture complexity coefficient of each second local window is calculated by the formula , and the second local window of is divided into a low texture area, the second local window of is divided into a medium texture area, and the second local window of is divided into a high texture area, wherein , is a positive number, and ;

[0027] Step B3, according to the corresponding relationship between the fine partition and the texture complexity partition, the basic gamma value of each pixel is determined;

[0028] Step B4, the determined basic gamma value is subjected to value range constraint, if , the basic gamma value is set to ; if , the basic gamma value is set to , , is a positive number, and ; ;

[0029] Step C, edge region priority adjustment and gamma value optimization:

[0030] Step C1, call the initial standardization wood template image to generate edge gradient map by improving Canny initial edge detection method, extract the pixels with top 20% gradient value as edge region pixels, and mark the edge region in the high gradient area in the 9 fine partitions;

[0031] Step C2, according to the brightness area to which the pixels in the edge region of the high gradient area belong, the brightness gradient value is fine tuned: if it belongs to the dark area, the brightness gradient value is increased by 0.1 to obtain the adjusted gamma value ; if it belongs to the bright area, the brightness gradient value is reduced by 0.1 to obtain the adjusted gamma value ; the high gradient area pixels of the non-edge region and other fine partition pixels, the adjusted gamma value ;

[0032] Step C3, calculate the edge signal-to-noise ratio SNR of the adjusted 9 fine partitions, wherein, represents the average gray value of the edge region pixels, is the average gray value of the noise pixels in the non-edge region; if , the adjusted gamma value is determined as the final gamma value of the current region ; if , adjust repeatedly in the way of increasing in dark area / decreasing in bright area with a step of 0.05 until or touch and boundary, is the set threshold value;

[0033] Step D: gamma value smooth transition and illumination normalization execution:

[0034] Step D1, take 7x7 as the smoothing window, traverse each pixel of the initial standardization wood template image, and perform weighted average calculation on the final gamma value of all pixels in the smoothing window, wherein the window center pixel weight is set as , and the total weight of the remaining 48 pixels in the window is , to obtain the smoothed final gamma value , ;

[0035] Step D2: Verify the smoothed adjacent pixels Differences, if differences Then, the weights of the smoothing window are readjusted. Specifically, the weight of the center pixel of the smoothing window is increased to 0.6, and the total weight of the non-center pixels is 0.4. Smoothing is then performed again until adjacent pixels... The difference is not greater than the preset difference threshold;

[0036] Step D3: For each pixel of the initial standardized wooden template image, use the formula... Perform a gamma transform, where Initial standardized wooden template image pixels grayscale value, pixels after gamma transformation The grayscale value is used to complete adaptive gamma illumination normalization processing and output an illumination-normalized wooden template image.

[0037] As an optimization, in B3, the base gamma value for each pixel... The specific implementation method for determining the rules is as follows:

[0038] Dark areas - low gradient regions: If located in a low-texture area, If it is in the medium texture area, If it is in a high-texture area, ;

[0039] Dark areas - medium gradient regions: If located in low-texture areas, If it is in the medium texture area, If it is in a high-texture area, ;

[0040] Dark areas - high gradient regions: If located in low-texture areas, If it is in the medium texture area, If it is in a high-texture area, ;

[0041] Highlights - Low Gradient Areas: If located in a low-texture area, If it is in the medium texture area, If it is in a high-texture area, ;

[0042] Highlights - Medium Gradient Area: If located in a low-texture area, If it is in the medium texture area, If it is in a high-texture area, ;

[0043] Highlights - High Gradient Areas: If located in a low-texture area, If it is in the medium texture area, If it is in a high-texture area, ;

[0044] Normal part-low gradient region, normal part-medium gradient region, normal part-high gradient region: All are 1.0.

[0045] As optimization, the specific process of performing texture direction weighted mean filtering processing, Laplacian pyramid edge enhancement processing and Sauvola local adaptive binarization processing is as follows:

[0046] Texture direction weighted mean filtering processing: the texture main direction of the wood template in the illumination normalized wood template image is determined by a Sobel operator, and a filter core matched with the texture main direction is used to perform weighted mean filtering on the illumination normalized wood template image to obtain a filtered image, wherein the filter core of the texture direction weighted mean filtering is an ellipse, the long axis is parallel to the texture main direction, the length of the long axis is 8px, and the length of the short axis is 3px; the weight formula of the weighted mean filtering is ; in the formula, d is the distance of the pixel from the center of the filter core, is the half length of the long axis of the filter core;

[0047] Laplacian pyramid edge enhancement processing: the filtered image is decomposed into a multi-layer Laplacian pyramid, contrast stretching and threshold enhancement are respectively performed on different levels, and an edge enhancement image is reconstructed;

[0048] Sauvola local adaptive binarization processing: the mean value and standard deviation of the pixels in a third local window of a preset size in the edge enhancement image are calculated, the binarization threshold is determined based on the mean value and the standard deviation, and a binarization image is output, and the calculation formula is: , in the formula, is the mean value of the pixels in the third local window, is the standard deviation of the pixels in the third local window, k and R are constants, and , , for example, , .

[0049] As optimization, the lightweight adaptive U-Net model includes an encoder and a decoder, the encoder includes four down-sampling modules, each down-sampling module is composed of a depth separable convolution layer, a batch normalization layer and a ReLU activation layer, the decoder includes four up-sampling modules, and each up-sampling module fuses features through adaptive weight skip connection with the down-sampling module of the corresponding level of the encoder;

[0050] The process of training the lightweight adaptive U-Net model is as follows:

[0051] A set of 10,000 initial standardized wood template image plates of different materials and different defect types are constructed, and are divided into a training set, a validation set and a test set in a ratio of 8:1:1; an Adam optimizer is used, with an initial learning rate set to , the learning rate is attenuated to a set multiple of the previous round every 20 rounds, and the model is iteratively trained for 100 rounds until the intersection over union of the edge detection of the validation set is greater than a set threshold, the dynamic weight joint loss function formula of the lightweight adaptive U-Net model is , wherein, is the cross-entropy loss, is the Dice coefficient, and , wherein A is the predicted edge pixel set, B is the real edge pixel set, is the dynamic adjustment coefficient; and The initial value of is 0.6, and when the lightweight adaptive U-Net model is trained, the value is adjusted according to the proportion p of edge pixels in the validation set every 10 iterations: when , when , .

[0052] As an optimization, the edge probability map is sequentially subjected to morphological opening operation denoising processing, B-spline curve smoothing fitting processing and coordinate system calibration processing to obtain the specific process of the wood template edge coordinates:

[0053] Morphological opening operation denoising processing: a rectangular structural element is used to first erode and then dilate the edge probability map to remove isolated noise blocks and obtain denoised edge pixel points;

[0054] B-spline curve smoothing fitting processing: the denoised edge pixel points are grouped, and a B-spline curve is used for fitting to obtain a fitting curve to control the fitting error;

[0055] Coordinate system calibration: based on the standardized image system, the edge coordinates of the fitting curve in the image coordinate system are converted into world coordinate system coordinates, thereby outputting the wood template edge coordinates.

[0056] As an optimization, the size of the rectangular structural element used in the morphological opening operation is 3x3, the erosion operation removes isolated noise blocks with an area , and the dilation operation restores the main contour of the edge;

[0057] The B-spline curve is a cubic curve, the edge pixel points are grouped according to a set number, and the fitting error is not greater than a set error threshold; if the average distance between the fitting curve and the original edge point is greater than a set distance threshold, control points are added to the local edge segment formed by the fitting group and its left and right adjacent fitting groups to adjust the fitting curve.

[0058] As optimization, the conversion formula of the coordinate system calibration is: , , wherein, is the edge coordinate in the image coordinate system, is the edge coordinate in the world coordinate system, K is the camera intrinsic matrix in the normalized image system, , is the principal point coordinate of the camera, and Z is the object distance.

[0059] As optimization, the intrinsic matrix K is specifically represented as:

[0060] , wherein, , is the focal length in the x and y directions, , is the principal point coordinate of the camera.

[0061] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0062] The present application realizes the accurate conversion of the wood template edge from the image to the actual physical coordinate through the whole process of adaptive gamma normalization, texture filtering to eliminate interference, Laplace enhancement of multi-scale edge, Otsu+Canny preliminary screening+lightweight U-Net fine-tuning, B-spline fitting error control, and coordinate calibration to physical size, which meets the industrial detection precision requirements.

[0063] The present application has targeted design for each link in view of the problems of uneven illumination of the wood template (gamma partition adaptation), complex / missing texture (texture filtering+texture-free Otsu bottoming), and scarring interference (scar identification skipping over-optimization), and can stably output the edge result whether it is a smooth plate, a textured plate or a defective plate. DETAILED DESCRIPTION

[0064] The accompanying drawings, which are included to provide a further understanding of the embodiments of the present application and constitute a part of the application, illustrate embodiments of the present application and do not limit the present application. In the drawings:

[0065] Figure 1 is a flowchart of the high-precision edge extraction method based on image automatic processing according to the present application. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below with reference to the embodiments and drawings, and the illustrative embodiments of the present application and the description thereof are only used to explain the present application and do not limit the present application.

[0067] The present application provides a high-precision edge extraction method based on image automatic processing, as shown inFigure 1 The method comprises the following steps:

[0068] S1, a standardized image system is constructed, and an initial standardized wood template image is acquired through the standardized image system;

[0069] S2, the initial standardized wood template image is subjected to multi-modal fusion preprocessing, including sequentially performing adaptive gamma light normalization processing, texture direction weighted mean filtering processing, Laplacian pyramid edge enhancement processing and Sauvola local adaptive binaryzation processing on the initial standardized wood template to obtain a binaryzation image, wherein the adaptive gamma light normalization comprises four sub-steps of image gray feature analysis, texture complexity calculation and basic gamma value determination, edge region priority adjustment and gamma value optimization, gamma value smooth transition and light normalization execution;

[0070] S3, an Otsu algorithm is used to determine a high threshold value and a low threshold value After edge detection is performed on the binaryzation image, disconnected edge segments with a distance of <3px and consistent gradient directions are linearly connected to generate a preliminary edge image and an edge gradient image; then the preliminary edge image is input into a pre-trained lightweight adaptive U-Net model to output an edge probability map; the lightweight adaptive U-Net model adopts a depth separable convolution to construct an encoder, fuses features through a skip connection, and is trained by using a dynamic weight joint loss function;

[0071] S4, the edge probability map is sequentially subjected to morphological opening operation denoising processing, B-spline curve smooth fitting processing and coordinate system calibration processing to obtain wood template edge coordinates;

[0072] S5, the average error of the wood template edge coordinates and the standard template edge coordinates is calculated, if the average error is not greater than a first threshold value, the edge coordinates and an edge label visualization image are output; if the average error is greater than the first threshold value or the edge detection intersection over union is less than a second threshold value, a retake instruction or an alarm signal is triggered. The first threshold value and the second threshold value can be selected according to actual conditions, which will not be described here.

[0073] Next, the aforementioned five steps will be introduced one by one.

[0074] In S1, the specific process of constructing a standardized image system and acquiring an initial standardized wood template image through the standardized image system is as follows:

[0075] S1.1, a standardized image system comprising a pixel industrial camera, a fixed-focus lens, a ring-shaped light source and a 0.01mm-precision standard chessboard calibration assembly is constructed;

[0076] S1.2, shooting parameters of object distance, shutter speed and ISO are set;

[0077] S1.3, before each batch detection, a standard chessboard is used to calibrate the camera intrinsic parameters to obtain an intrinsic matrix K;

[0078] S1.4, an initial standardized wooden template image is obtained by shooting the wooden template.

[0079] For example, a standardized image acquisition system is constructed, which includes a 2000-megapixel industrial camera, an 8-mm fixed-focus lens, a 5500-K ring light source, and a 0.01-mm precision standard chessboard calibration component; the shooting parameters are fixed at a focal length of 300 mm, a shutter speed of 1 / 500 s, and an ISO of 100. Before each detection, the standard chessboard is automatically shot to verify the deviation of the intrinsic matrix (if the deviation is greater than 1%, calibration is triggered).

[0080] In some embodiments, the intrinsic matrix K is specifically represented as:

[0081] wherein, , is the focal length in the x and y directions, , is the principal point coordinate of the camera.

[0082] The intrinsic matrix K is a matrix that describes the internal inherent parameters of the camera, and reflects the geometric transformation relationship when the camera projects a three-dimensional scene point to a two-dimensional image plane. , is the equivalent focal length of the camera in the x and y directions, which determines the scaling degree of the scene point on the image; , is the principal point coordinate of the camera, i.e., the coordinate of the intersection of the camera optical axis and the image plane, which is used to determine the origin offset of the image coordinate system. Knowing the intrinsic matrix, the pixel coordinates on the image plane can be converted into coordinates in the camera coordinate system, which is the basis for conversion from image pixels to actual physical space measurement.

[0083] The acquisition system constructed by including a specific industrial camera, a lens, a light source, and a calibration component can provide a stable and unified hardware basis for wooden template image acquisition, reduce the image quality fluctuations caused by hardware differences (such as different imaging characteristics of different cameras), and ensure that the images acquired each time are comparable and analyzable in terms of basic conditions such as resolution and lighting. For example, a 2000-megapixel industrial camera can provide high-resolution images, clearly capturing the details of the wooden template, which is beneficial for subsequent edge extraction and other processing.

[0084] Setting fixed shooting parameters such as object distance, shutter speed, and ISO can prevent changes in image brightness and sharpness caused by arbitrary changes in shooting parameters. For example, a fixed object distance of 300mm ensures that the wooden template is displayed in a stable proportion in the image, a fixed shutter speed of 1 / 500s reduces blur caused by possible slight vibrations of the wooden template, and a fixed ISO of 100 controls the image noise level, resulting in stable image quality.

[0085] Before each batch of testing, standard checkerboard calibration is used to promptly correct intrinsic parameter offsets caused by camera usage time and environmental changes (such as lens deformation due to temperature), ensuring the accuracy of the intrinsic parameter matrix and thus making image imaging more precise. This provides reliable image data for subsequent image-based operations such as size measurement and edge extraction, reducing the transmission of errors caused by inaccurate camera parameters to subsequent processing stages.

[0086] Next, we will analyze the technical solutions in step S2 in detail.

[0087] S2.1 Adaptive gamma illumination normalization processing includes four sub-steps: image grayscale feature analysis, texture complexity calculation and basic gamma value determination, edge region priority adjustment and gamma value optimization, gamma value smooth transition and illumination normalization execution.

[0088] Step A: Image grayscale feature analysis:

[0089] Step A1: Calculate the grayscale histogram of the initial standardized wooden template image, statistically analyze the grayscale value distribution of all pixels in the initial standardized wooden template image, and determine the 10th percentile of the grayscale values. and 90th percentile The initial standardized wooden template image is divided into dark areas, normal areas, and bright areas, wherein the pixel grayscale values ​​of the dark areas are... The pixel grayscale value of the normal region The pixel grayscale value of the bright area ;

[0090] Step A2: Using a 3×3 area as the first local window, traverse each pixel of the initial standardized wooden template image and calculate the brightness gradient in the x-direction within the first local window using the Sobel operator. brightness gradient in the y-direction Through formula The brightness gradient value of each window is calculated. ;

[0091] Step A3: Based on the brightness gradient value Based on the distribution characteristics, all brightness gradient values ​​of the initial normalized wooden template image are... The gradient is divided into low gradient area, medium gradient area and high gradient area, and the gradient partition is superimposed with the dark area, normal area and bright area divided in step A1 to form 9 fine partitions, namely dark-low gradient area, dark-medium gradient area, dark-high gradient area, normal-low gradient area, normal-medium gradient area, normal-high gradient area, bright-low gradient area, bright-medium gradient area and bright-high gradient area, wherein the luminance gradient value of the low gradient area is , the luminance gradient value of the medium gradient area is , and the luminance gradient value of the high gradient area is .

[0092] Step A4, traversing the 9 fine partitions generated in step A3, performing the following operations on the high gradient area

[0093] Step A4.1, using a contour detection algorithm (such as findContours of OpenCV) to identify continuous connected domains, and calculating the area S=long x wide of the circumscribed rectangle of the connected domain;

[0094] A4.2 If S>20x20 pixels (determined as a large area high gradient area), the connected domain image is intercepted, and a 3x3 local window is used to calculate the texture entropy Entropy (the calculation logic is the same as the gray level co-occurrence matrix in subsequent step B1) in advance;

[0095] A4.3 If the texture entropy Entropy>90 (determined as a high texture complexity scar suspected area), mark the area as a stutter mark area, and in the texture complexity calculation and basic gamma value determination in subsequent step B, skip the local optimization of the area and directly assign the basic gamma value (preserve the original gray scale to avoid excessive adjustment).

[0096] Step B, texture complexity calculation and basic gamma value determination:

[0097] Step B1, using a 5x5 second local window, traversing each pixel in the 9 fine partitions of the non-stutter mark area, calculating the texture contrast Contrast and texture entropy value Entropy in each second local window through a gray level co-occurrence matrix GLCM, wherein the texture contrast Contrast is used to reflect the clarity of the texture in the second local window, and the texture entropy value Entropy is used to reflect the random complexity of the texture in the second local window;

[0098] Step B2, calculating the texture complexity coefficient of each second local window through the formula , and dividing the second local window of into a low texture area, ​The second local window of the dark part is divided into a low-texture area, The second local window of the dark part is divided into a high-texture area;

[0099] Step B3, determining the base gamma value of each pixel according to the correspondence between the fine partition and the texture complexity partition ;

[0100] determining the base gamma value of each pixel The specific rules are as follows:

[0101] Dark part-low gradient area: if in the low-texture area, ; if in the medium-texture area, ; if in the high-texture area, ;

[0102] Dark part-medium gradient area: if in the low-texture area, ; if in the medium-texture area, ; if in the high-texture area, ;

[0103] Dark part-high gradient area: if in the low-texture area, ; if in the medium-texture area, ; if in the high-texture area, ;

[0104] Bright part-low gradient area: if in the low-texture area, ; if in the medium-texture area, ; if in the high-texture area, ;

[0105] Bright part-medium gradient area: if in the low-texture area, ; if in the medium-texture area, ; if in the high-texture area, ;

[0106] Bright part-high gradient area: if in the low-texture area, ; if in the medium-texture area, ; if in the high-texture area, ;

[0107] Normal part-low gradient area, normal part-medium gradient area, and normal part-high gradient area: are all 1.0;

[0108] Step B4, performing value range constraint on the determined base gamma value If the base gamma value is less than 0.4, the base gamma value is set to 0.4; if the base gamma value is greater than 2.0, the base gamma value is set to 2.0.

[0109] Step C, edge region priority adjustment and gamma value optimization:

[0110] Step C1, call the initial standardization wood template image to generate edge gradient map by improving Canny initial edge detection method, extract the top 20% gradient value pixels as edge region pixels, mark the edge region in the high gradient area of 9 fine partitions;

[0111] Step C2, according to the brightness area to which the edge region pixels in the high gradient area belong, the brightness gradient value is fine tuned: if it belongs to the dark area, the brightness gradient value is increased by 0.1 to obtain the adjusted gamma value ; if it belongs to the bright area, the brightness gradient value is reduced by 0.1 to obtain the adjusted gamma value ; the high gradient area pixels of non-edge region and other fine partition pixels, the adjusted gamma value ;

[0112] Step C3, calculate the edge signal-to-noise ratio SNR of the adjusted 9 fine partitions, wherein, represents the average gray value of the edge region pixels, is the average gray value of the noise pixels in the non-edge region; if , the adjusted gamma value is determined as the final gamma value of the current region ; if , adjust repeatedly in the way of increasing in dark area / decreasing in bright area with a step of 0.05 until or touch the 0.4 / 2.0 boundary.

[0113] Step D: gamma value smooth transition and illumination normalization execution:

[0114] Step D1, take 7x7 as the smoothing window, traverse each pixel of the initial standardization wood template image, and perform weighted average calculation on the final gamma value of all pixels in the smoothing window, wherein the weight of the center pixel of the smoothing window is set to 0.5, and the weight of the remaining 48 pixels in the smoothing window is set to 0.5, to obtain the smoothed final gamma value ;

[0115] Step D2, verify the difference of adjacent pixels after smoothing, if the difference If the difference is greater than 0.1, then the smoothing window weights are readjusted, specifically the center pixel weight of the smoothing window is increased to 0.6 and the non-center pixel weights total 0.4, and the smoothing is performed again until the difference between adjacent pixels is less than or equal to 0.1. the difference is less than or equal to 0.1.

[0116] Step D3, for each pixel of the initial normalized wood template image, the adaptive gamma illumination normalization processing is performed by the formula performing a gamma transformation, wherein is the gray value of the pixel of the initial normalized wood template image, is the gray value of the pixel after the gamma transformation, is the gray value of the pixel after the gamma transformation, the adaptive gamma illumination normalization processing is completed, and an illumination normalized wood template image is output.

[0117] The adaptive gamma illumination normalization processing is achieved through multiple sub-steps, and the following core effects are achieved:

[0118] 1. Precise partition adaptation: first, the image is finely divided into 9 regions according to the gray histogram and the gradient (dark / bright part with low / medium / high gradient), and then combined with the texture complexity to subdivide, so that different characteristic regions (such as dark low texture and bright high gradient) have adaptive gamma values, solving the problem of local overexposure / overexposure caused by uneven wood template illumination and complex texture, making the overall brightness of the image more uniform.

[0119] 2. Edge protection optimization: call the edge gradient map, and fine-tune the gamma value of the high gradient area (edge intensive area) accordingly, avoid over-processing when brightening the dark edge and darkening the bright edge, preserve edge details, improve the accuracy of subsequent edge extraction, and make the edge clearer and more continuous.

[0120] 3. Smooth natural transition: 7x7 window weighted average smoothing of gamma values avoids color blocks caused by partition adjustment, making the image illumination transition natural; verify and adjust the weight to ensure that the difference between adjacent pixels is small, and the output image has good visual effect without obvious light-dark boundaries.

[0121] 4. Lay the foundation for the future: the processed image has uniform illumination, clear edges, and weak texture interference, providing high-quality input for subsequent texture filtering, edge enhancement, and deep learning edge extraction, improving the accuracy and stability of the entire wood template edge extraction scheme, and reducing errors caused by illumination and texture.

[0122] It also needs to be explained that in this step, C1 calls the improved Canny primary edge detection method in step S3.1 to generate an edge gradient image from the initial standardized wood template image, that is, step C1 calls the improved Canny primary edge detection method to perform a simplified version of step 3.1 once using the initial standardized wood template image output in step S1 (without any preprocessing in step S2), only basic Canny edge detection is performed to generate a temporary edge gradient image (without connecting edge segments, only the gradient value of each pixel is obtained), only to provide edge region position reference for step C1, without the need for a high-precision primary edge image, the core is to quickly obtain the basic data of which pixels may be edges; Step C1 calls the temporary edge gradient image generated in the previous stage to complete gamma value optimization and output a light normalization image.

[0123] S2.2, texture direction weighted average filtering processing:

[0124] The texture main direction of the wood template in the light normalization wood template image is determined by the Sobel operator, and a filter kernel matched with the texture main direction is used for weighted average filtering of the light normalization wood template image to obtain a filtered image, wherein the filter kernel of the texture direction weighted average filtering is an ellipse, the long axis is parallel to the texture main direction, the length of the long axis is 8px, and the length of the short axis is 3px; The weight formula of the weighted average filtering is ; In the formula, d is the distance of the pixel from the center of the filter kernel, is the half length of the long axis of the filter kernel.

[0125] This step solves the wood template texture interference problem by the combination strategy of "directional filtering + weighted average", and lays a foundation for subsequent edge extraction, and the specific effects are as follows:

[0126] Accurate suppression of wood grain interference and retention of edge signal:

[0127] The natural wood grain on the surface of the wood template (such as the annual rings of pine wood and the fiber texture of cedar wood) will interfere with edge extraction, and traditional average filtering will blur both the wood grain and the edge. However, this step:

[0128] Directional filter kernel: determine the texture main direction (such as horizontal / vertical distribution of wood grain) by Sobel operator, use an elliptical filter kernel (long axis 8px, short axis 3px) with the long axis parallel to the texture direction, and only perform weighted smoothing on the pixels in the texture direction, so that the wood grain is flattened along its own direction;

[0129] Weighted average formula: weight The texture pixel weight of the filter kernel center is high, and the edge pixel weight is low. In this way, the wood grain will be highlighted and smoothed along the texture direction; while the real edge of the wood template (such as the joint seam and defect edge) will be less affected by the filtering because it is perpendicular to the texture direction, thus achieving the suppression of wood grain without damaging the edge.

[0130] Improve image signal-to-noise ratio and reduce subsequent processing difficulty:

[0131] The irregular distribution of wood grain increases image noise (texture noise), which leads to false edges in subsequent edge detection. This step smoothes the wood grain grayscale fluctuations and reduces the grayscale difference of texture noise (such as the weighted average of dark and light stripes of wood grain).

[0132] Texture direction targeted smoothing: ironing out the grayscale fluctuations of wood grain, reducing the grayscale difference of texture noise (such as the weighted average of dark and light stripes of wood grain);

[0133] Edge preservation: make the grayscale contrast of real edges vs. background in the image more prominent. After processing, the image signal-to-noise ratio (real edge grayscale difference / texture noise grayscale difference) is improved, and subsequent binaryzation and deep learning edge extraction can more accurately identify real edges, reducing the probability of misjudging wood grain as an edge.

[0134] Collaborate with the overall scheme to strengthen edge extraction accuracy:

[0135] This step is a key link in multi-modal preprocessing, connected to adaptive gamma illumination normalization (solving uneven illumination) and Laplacian pyramid edge enhancement (strengthening real edges):

[0136] Front collaboration: Gamma normalization solves the uneven light and dark caused by light, but the texture noise of wood grain still exists; this step relays to suppress texture noise and makes preprocessing more thorough;

[0137] Back collaboration: edge enhancement relies on clear edge-background contrast, and the real edges preserved by this step can make the Laplacian pyramid more accurately enhance edge signals, ultimately improving the accuracy of step 3 (adaptive deep learning edge extraction).

[0138] S2.3, Laplacian pyramid edge enhancement processing:

[0139] The filtered image is decomposed into a multi-layer Laplacian pyramid, and contrast stretching and threshold enhancement are performed on different levels to reconstruct an edge-enhanced image.

[0140] The specific execution process is as follows:

[0141] S2.3.1, image decomposition: construct a Laplacian pyramid:

[0142] Input the image filtered by texture direction weighting, generate multi-layer pyramids by Gaussian blur + down-sampling: first, apply Gaussian blur to the image with different degrees (e.g. 5x5, 9x9 window), then down-sample it by row and column (e.g. from 5472x3648→2736x1824→1368x912…), get 4-5 layers of "Gaussian pyramids";

[0143] Based on the Gaussian pyramids, calculate the Laplacian residual of each layer using the current layer - the up-sampled and Gaussian blurred result of the previous layer (e.g. the nth layer Laplacian image = the nth layer Gaussian image - the up-sampled and blurred image of the (n-1)th layer), construct a 4-layer Laplacian pyramid (corresponding to edge information of different frequencies).

[0144] S2.3.2, hierarchical enhancement: contrast stretching + threshold enhancement:

[0145] Top layer (low-frequency layer, corresponding to large-scale edges): apply 1.5 times contrast stretching to the topmost Laplacian image (e.g. 1368x912 size), map the pixel gray value range from to (adjust the specific coefficients according to the wood template image statistics), strengthen the gray difference of large-scale edges (e.g. the overall outline of the wood template);

[0146] Bottom layer (high-frequency layer, corresponding to small-scale edges): calculate the global gray mean m for the bottommost Laplacian image (original size 5472x3648), set the threshold to 1.2*m, multiply the pixels with gray value greater than the threshold by 1.2 (enhance small-scale edges such as scarring details), and keep the pixels with gray value not greater than the threshold unchanged, highlight the fine edges.

[0147] S2.3.3, image reconstruction: generate edge enhancement map:

[0148] From the top layer, up-sample the enhanced Laplacian image and fuse it with the next layer: first, up-sample the top layer enhancement map to the size of the next layer, and add it pixel by pixel with the next layer enhancement map;

[0149] Repeat the fusion process until the original image size is restored, get the edge enhancement image, which fuses the enhancement results of large-scale outlines and small-scale details, provides clearer edge signals for subsequent binarization and edge extraction;

[0150] Because the edge type of the wood template is complex: there are large-scale edges (corresponding to the top layer of the Laplacian pyramid) such as spliced joints and large cracks, and small-scale edges (corresponding to the bottom layer) such as scars and fine cracks. Through layered enhancement, large-scale edges are more prominent due to contrast stretching, and small-scale edges are clearer due to threshold enhancement, achieving no missing of large defects and identification of small defects, solving the problem of traditional edge enhancement that either misses small defects or misses large defects, achieving multi-scale edge enhancement and covering different defects.

[0151] And the process connects the previous texture direction filtering (suppresses wood grain noise and preserves real edges) process, and the Laplacian pyramid enhancement can accurately amplify the edge signal: the filtered image has weakened the wood grain interference, and the enhancement process only needs to focus on the edge itself, avoiding the noise being enhanced at the same time. For example, the wood template spliced joint has been separated from the wood grain after filtering, and the contrast is further improved after enhancement, and the subsequent binarization can more accurately extract.

[0152] S2.4, Sauvola local adaptive binarization processing:

[0153] The third local window (for example, the local window size of Sauvola local adaptive binarization is 15x15) of a preset size is used to calculate the mean and standard deviation of the pixels in the third local window in the edge enhanced image, and a binarization threshold is determined based on the mean and standard deviation, and a binarization image is output, and the calculation formula is: , wherein is the mean of the pixels in the third local window, is the standard deviation of the pixels in the third local window, , .

[0154] The specific implementation process of S2.4 is as follows:

[0155] S2.4.1, traverse the image to divide the local window: for the wood template image after preprocessing such as Laplacian pyramid edge enhancement, a local window of 15x15 pixels is used to traverse the entire image window by window. The window slides by row or column, and moves one pixel each time (the step size can also be adjusted according to requirements, but is usually 1 pixel to ensure that all areas are covered).

[0156] S2.4.2, calculate the statistics in the window: for each 15x15 local window, calculate the mean and standard deviation of the pixels in the window. The mean is the average of all pixel gray values in the window, reflecting the overall brightness level in the window; the standard deviation reflects the dispersion degree of the pixel gray values in the window, reflecting the local texture and illumination change.

[0157] S2.4.3, Calculate the binary threshold: put the calculated and into the binary threshold formula , to calculate the binary threshold corresponding to the current window . The threshold is dynamically adjusted according to the local features in the window, which can adapt to changes in different areas.

[0158] S2.4.4, Pixel-by-pixel binary processing: for each pixel in the current local window, compare its gray value with the calculated threshold . If the pixel gray value is greater than or equal to , set the pixel to foreground (e.g. gray value 255); if less than , set to background (e.g. gray value 0). After completing the current window processing, continue to process the next window, until the entire image is traversed, and the final binary template image is obtained.

[0159] The surface of the wood template has local uneven illumination and texture changes due to material, defects (such as scars, cracks), etc. Sauvola local adaptive binary processing calculates the mean and standard deviation for each pixel in a 15x15 local window to determine the threshold, which can accurately adapt to the characteristics of different areas of the wood template. Compared with global binary processing, it can avoid the problems of edge blur, detail loss or false edge caused by improper overall threshold, so that the real edges of the wood template (such as joint seams, defect edges) can be more clearly separated from the background, thus achieving the purpose of adapting to the local illumination and texture differences of the wood template.

[0160] At the same time, accurate binary results are the basis for subsequent edge extraction (such as improved Canny edge detection, U-Net model optimization, etc.). The binary image generated by this method can better preserve the edge details of the wood template, reduce noise interference, and enable the subsequent edge extraction step to more accurately identify the real edges of the wood template, improving the accuracy and reliability of the entire edge extraction scheme.

[0161] It should be noted that the improved Canny initial edge detection method is an optimization scheme for wood template images with strong texture interference and uneven illumination. It realizes accurate initial screening edge and gradient information extraction through Otsu automatic threshold adaptation + gradient direction constraint edge connection + edge gradient map output. It is based on the traditional Canny algorithm, with the addition of Otsu automatic double threshold adaptation mechanism, gradient direction consistency constraint weak edge screening strategy, and edge gradient map synchronous output function. It simultaneously inputs the image after illumination normalization and texture filtering preprocessing, realizes the initial edge detection without manual parameter adjustment, anti-wood texture interference, and edge fracture repair, and provides gradient information support for subsequent adaptive gamma optimization and deep learning refinement.

[0162] The specific steps of improving the Canny initial edge detection method are as follows:

[0163] Step 1: Preprocessing adaptation:

[0164] The Sauvola locally adaptive binarization image output in step S2.4 is output (illumination normalization, texture filtering, and edge enhancement have been completed), avoiding the influence of original image noise and illumination interference on edge detection.

[0165] Step 2: Automatic calculation of double threshold value by Otsu algorithm (instead of traditional manual threshold value):

[0166] For the gray level histogram of the binarized image, the Otsu maximum inter-class variance method is used to automatically divide the edge pixels and background pixels, and the high threshold value ;

[0167] The low threshold value is calculated according to a fixed ratio (adapting to the edge gray level difference of the wooden template, balancing missed detection and false detection);

[0168] Step 3: Gradient calculation and non-maximum suppression:

[0169] A 3x3 Sobel operator is used to calculate the x-direction gradient , y-direction gradient , and gradient amplitude (reflecting edge strength) is obtained by , and gradient direction (0° / 45° / 90° / 135° four-vectorization) is obtained by ;

[0170] Non-maximum suppression: for each pixel, compare the gradient amplitudes of the pixel and its adjacent two pixels along its gradient direction (such as 90° direction along the vertical direction), and only keep the pixel with the maximum gradient amplitude (remove non-peak points on the edge, making the edge thinner).

[0171] Step 4: Edge connection with gradient direction constraint:

[0172] First, mark strong edge pixels (gradient amplitude , directly judged as real edge) and non-edge pixels (gradient amplitude , directly removed);

[0173] For weak edge pixels (gradient amplitude ), introduce gradient direction consistency constraint:

[0174] If the Euclidean distance between the weak edge pixel and the strong edge pixel is , and the gradient direction angle between them is (judged as the same continuous edge), then the weak edge pixel is upgraded to strong edge;

[0175] Conversely, if weak edge pixels are not connected to strong edge pixels, or if the gradient directions are significantly different (considered as texture interference), then the weak edge pixel is removed.

[0176] Core optimization: Traditional Canny only connects adjacent spaces, while this solution adds gradient direction consistency constraints, effectively filtering out pseudo edges such as wood grain, and at the same time filling in the broken parts of real edges (such as the tiny breaks in the splicing seams of wooden templates).

[0177] Step 5: Output double results

[0178] Output initial edge image: only retain strong edge pixels (real edge contours) as input to the lightweight U-Net model (reducing the model's inference pressure);

[0179] Output edge gradient map: Records the gradient magnitude of each pixel. (Step 3 Calculation Results) are used for subsequent S2.1 Adaptive Gamma Optimization (selecting high gradient edge regions for priority adjustment) and S5 Error Calculation (assisting edge region localization).

[0180] For the wood template detection scenario, the parameters in its formula (window size 15×15, k=0.3, R=128) have been adjusted and adapted to the characteristics of wood template images. They perform well in suppressing wood grain texture interference and highlighting real edges, and can effectively distinguish the background and edge areas of the wood template, providing high-quality binarized images for subsequent processing.

[0181] In S3, a combined strategy of traditional algorithm initial screening and deep learning refinement is used to solve the problem of edge extraction of wooden templates. The specific process of S3 is as follows:

[0182] S3.1, Traditional algorithm for initial screening (generating initial edge image + gradient map):

[0183] S3.1.1. Automatically calculate the high threshold from the binarized image using the Otsu algorithm (maximum inter-class variance method). and low threshold This replaces the traditional manual threshold of Canny and adapts to the brightness differences of wooden templates.

[0184] Statistically analyze the gray-level histogram (pixel gray-level value distribution) of the binarized image, iterate through all possible thresholds, calculate the inter-class variance of the foreground (edge) and background (non-edge), and select the threshold with the largest inter-class variance as the high threshold. Press again Obtain low threshold .

[0185] S3.1.2, calculate the gray gradient (reflecting the degree of gray change) and the gradient direction of each pixel of the binary image with Sobel operator (3x3 window), mark the pixels with gradient value greater than 0.5 as strong edge pixels (determine as real edge); mark the pixels with gradient value less than 0.1 as non-edge pixels and directly eliminate; mark the pixels with gradient value in the range of 0.1-0.5 as weak edge pixels and wait for determination.

[0186] S3.1.3, for the weak edge pixels, check whether they are adjacent to the strong edge pixels (distance < 3px) and the gradient directions are consistent (such as both are horizontal direction): if the conditions are met, upgrade to strong edge pixels and connect into complete edges; if the conditions are not met, eliminate and finally obtain the initial edge image (only keep the outline of real edges).

[0187] S3.1.4, synchronously output the edge gradient image: record the gradient value of each pixel (the result of gradient calculation in step 3.1.3), and the subsequent step C1 (intelligent gamma optimization) will call the image to screen the edge pixels in the high gradient area, if it is to be used as the output of step C1, the input of S3.1 is the output of S1.

[0188] S3.2, deep learning refinement (lightweight U-Net model optimization):

[0189] S3.2.1, the input is the initial edge image generated in stage S3.1 (most of the noise has been filtered out and the basic edge outline is retained), the size is consistent with the original wood template image (such as 5472x3648);

[0190] S3.2.2, design a lightweight U-Net model structure, the lightweight U-Net model includes an encoder and a decoder, the encoder uses a depth separable convolution (decomposes a standard convolution into a channel-wise convolution + a point-wise convolution), reduces the amount of calculation, makes the model lighter (reduces the parameters by more than 70%), from each layer of the encoder, the features are directly transmitted through the skip connection to the decoder (such as the texture details of the first layer are directly transmitted to the last layer), and multi-scale edge information (large outline + small details) is fused; the dynamic weight joint loss function formula of the lightweight adaptive U-Net model is , in the formula, is the cross-entropy loss, is the Dice coefficient, and , wherein A is a set of predicted edge pixels, B is a set of real edge pixels, is a dynamic adjustment coefficient; and ​​​The initial value of the is 0.6, which can balance the classification loss and the edge similarity loss, so that the model focuses on the classification of edge vs. non-edge and ensures that the predicted edge and the real edge have consistent shapes; when training the lightweight adaptive U-Net model, every 10 iterations, the edge pixel ratio p of the validation set is adjusted: when , ; when , .

[0191] The lightweight adaptive U-Net model has the following structural characteristics:

[0192] 1. Lightweight encoder design:

[0193] The encoder includes four down-sampling modules, each of which consists of a depth separable convolution layer, a batch normalization layer, and a ReLU activation layer, wherein:

[0194] The depth separable convolution layer has a convolution kernel size of 3x3, and the number of groups is equal to the number of input feature map channels (i.e., channel-wise convolution), and the point-wise convolution kernel size is 1x1;

[0195] The output channel numbers of the four down-sampling modules are 32, 64, 128, and 256 in sequence (compared with the standard U-Net, the total number of parameters is reduced by ≥70%), and each down-sampling module is followed by a max pooling layer with a step size of 2 (pooling kernel 2x2);

[0196] 2. Adaptive feature fusion mechanism:

[0197] The decoder includes four up-sampling modules, each of which fuses features with the corresponding layer of the encoder through an "adaptive weight skip connection", specifically:

[0198] 2.1, a feature weight calculation unit is set at the skip connection, which is calculated by the formula , wherein, represents the feature map output by the U-Net encoder, with a dimension of (H is the height of the feature map, W is the width, is the output channel number of the encoder), which contains local detailed features of the image (such as wood template edge texture, minor defects); represents the feature map output by the U-Net decoder, with a dimension of (which is consistent with the spatial size of the encoder feature map, is the output channel number of the decoder), which contains the global semantic features of the image (such as the overall structure of the wood template, the region category).

[0199] represents a 1x1 convolution operation, specifically:

[0200] To apply a 1x1 convolution kernel, compress the number of channels from to (intermediate number of channels, such as 32), and the output feature map is ;

[0201] To apply a 1x1 convolution kernel, compress the number of channels from to , and the output feature map is ;

[0202] The role is to unify the number of channels of the encoder and decoder feature maps, and ensure that the subsequent can be added element by element.

[0203] Addition operation: add the encoder feature map and the decoder feature map after 1x1 convolution pixel by pixel and channel by channel to obtain the initial weight signal with a dimension of .

[0204] Sigmoid activation function, formula , map the numerical range of the initial weight signal to [0,1], and output the final dynamic weight matrix .

[0205] is a weight matrix with a dimension of , and each element value represents the fusion ratio of the corresponding position encoder feature and decoder feature (the value closer to 1, the higher the proportion of the encoder feature; the closer to 0, the higher the proportion of the decoder feature).

[0206] 2.2, fused feature , realize adaptive matching of encoder detail features and decoder semantic features;

[0207] 3, dynamic weight joint loss function: the model training adopts a dynamic weight joint loss function: , the weight is adjusted dynamically through the "edge pixel ratio": assuming that the edge pixel ratio in the image is r, then (the edge pixel ratio is low, increase to strengthen the edge classification; the ratio is high, reduce to balance the overall loss);

[0208] 4, inference speed constraint: the model on the NVIDIA Jetson AGX Orin (or equivalent computing device) has an inference time of ≤25ms for a wooden template initial edge image with a resolution of 5472x3648, and the model file size is ≤50MB.

[0209] The training process is to build an initial standardized wood template image set containing 10,000 different materials and different defect types (including pine / spruce / poplar, with knot / split defects), which is divided into a training set, a validation set and a test set according to 8:1:1. The Adam optimizer is used, and the initial learning rate is , and the learning rate is reduced to 0.5 times of the previous round every 20 rounds. The edge detection intersection over union of the validation set is ≥0.89 (an index for measuring the accuracy of edge prediction) after 100 iterations of training. After training, the model is accelerated by TensorRT to ensure that the model runs fast in an industrial scenario (inference time ≤25ms).

[0210] The initial edge image is input into the pre-trained U-Net model, which predicts the probability of being an edge (a value between 0 and 1) for each pixel and outputs an edge probability map. The closer the probability is to 1, the more likely it is to be a real edge (for example, the probability of the center of the joint is 0.95, and the probability of the wood grain area is 0.1). This provides high-precision edge data for step S4 (post-processing optimization).

[0211] The specific effect of S3 is:

[0212] Otsu threshold + Canny detection: Otsu algorithm is used to automatically find high and low thresholds to avoid the subjectivity of manual parameter adjustment, making Canny edge detection more suitable for wood template images and accurately identifying large-scale edges such as joint seams and large cracks, solving the "missed detection / misdetected" problem of traditional threshold methods.

[0213] Broken edge connection: for disconnected edges with a distance of <3px and consistent gradient direction, linear connection is used to complete the connection (such as the fine cracks around the knot of the wood template), making the outline of the initial edge image more complete and providing "continuous basic edges" for subsequent deep learning.

[0214] In step S4, the edge probability map is sequentially subjected to morphological opening denoising processing, B-spline curve smoothing fitting processing and coordinate system calibration processing to obtain the specific process of the wood template edge coordinates as follows:

[0215] S4.1, morphological opening denoising processing:

[0216] The edge probability map is first eroded and then expanded using a rectangular structural element to remove isolated noise blocks, obtaining denoised edge pixels. More specifically, the rectangular structural element used in the morphological opening operation has a size of 3x3, the erosion operation removes isolated noise blocks with an area of , and the expansion operation restores the main outline of the edge.

[0217] Morphological opening operation (first erosion and then dilation) uses morphological transformation of the structural element to remove isolated noise blocks in the edge probability map (such as false edges and small noise points misjudged by the U-Net model) and at the same time to preserve the continuity of the real edge as much as possible.

[0218] The specific process is as follows:

[0219] S4.1.1, select a structural element: adopt a 3x3 rectangular structural element with an implicit industrial adaptation size that can cover the typical noise block size of the wood template, and the structural element is represented as: ; wherein 1 represents the pixel position participating in morphological transformation, and the 3x3 size balances the denoising effect and edge retention degree (a size that is too large will damage the real edge, and a size that is too small will not be able to filter small noise).

[0220] S4.1.2, erosion operation: traverse each pixel (denoted as ) of the edge probability map, and use the structural element B to do sliding window matching: calculate the minimum value of all pixels in the window as the new value of the center pixel of the window after erosion, and the calculation formula is: , denotes the erosion operation, and through this step, isolated noise blocks (small in area and with abrupt gray value) can be "reduced" or even eliminated, and the real edge is only slightly thinned after erosion due to continuous distribution. P is the pixel gray value in the edge probability map.

[0221] S4.1.3, dilation operation: traverse the image after erosion again using the structural element B, calculate the maximum value of all pixels in the window, and use it as the new value of the center pixel of the window after dilation, and the calculation formula is: , denotes the dilation operation, and through this step, the real edge is "restored" to its original width after being thinned by erosion, and isolated noise blocks are eliminated by erosion and cannot be restored by dilation, thus achieving "denoising while preserving edges".

[0222] S4.1.4, after the opening operation of erosion and then dilation, the edge pixel points after denoising are obtained, which basically correspond to the real edge of the wood template (noise blocks are filtered and edge continuity is preserved).

[0223] S4.2, B-spline curve smoothing fitting processing:

[0224] The denoised edge pixel points are grouped, and a B-spline curve is used for fitting to obtain a fitting curve to control the fitting error. More specifically, the B-spline curve is a cubic curve, the edge pixel points are grouped by 10-15 points per group, and the fitting error is ≤0.1 mm; if the average distance between the fitting curve and the original edge point is >0.5px, control points are added to the local edge segment formed by the fitting group and its left and right adjacent fitting groups to adjust the fitting curve.

[0225] The B-spline curve fits the edge pixel points by segmentation, approximates the discrete edge with a mathematical curve, achieves the effect of "smoothing edge sawtooth and suppressing local fluctuation", and ensures the accuracy by controlling the fitting error (≤0.1 mm).

[0226] The specific process is as follows:

[0227] S4.2.1, edge pixel grouping: the denoised edge pixel points are grouped according to spatial continuity, and each group contains 10-15 pixels (implicit grouping rule, balancing calculation amount and fitting accuracy). Grouping logic: traverse the pixels in the order of the image coordinate system , calculate the Euclidean distance between the current pixel and the previous pixel: , if (experience value, corresponding to the actual size of the wooden template ≤0.275 mm, which can be regarded as a continuous edge), it belongs to the same group; otherwise, a new group is started.

[0228] S4.2.2, cubic B-spline curve fitting: for each group of pixel points (denoted as ), a cubic B-spline curve is used for fitting, and the steps are as follows:

[0229] S4.2.2.1, define node vector: let the number of nodes be m=n+4 (the number of nodes of a cubic B-spline curve is a rule to ensure that the curve is connected to the pixel points at the beginning and end), and the node vector , where , , and the intermediate nodes are uniformly distributed ;

[0230] S4.2.2.2, calculate the basis function: the cubic B-spline basis function is calculated by the recursive formula , ;

[0231] S4.2.2.3, fitting curve equation: the curve parameter , and the coordinates of the fitting curve are: , ; is the coordinate of the i-th pixel; , Let be the coordinates of the fitted curve corresponding to the i-th pixel.

[0232] S4.2.3 Fitting Error Control: Calculate the average distance between the fitted curve and the original edge pixels (i.e., fitting error): If the fitting error is >0.5px (corresponding to the actual size >0.0275mm), then add control points to the group of pixels (such as inserting a new pixel at the position with the largest error and regrouping for fitting) until the error is ≤0.5px, and finally ensure that the overall fitting error is ≤0.1mm.

[0233] S4.3 Coordinate System Calibration:

[0234] Based on the standardized image system, the edge coordinates of the fitted curve in the image coordinate system are converted to world coordinate system coordinates, thereby outputting the edge coordinates of the wooden template. More specifically, the conversion formula for coordinate system calibration is as follows: , In the formula, These are the edge coordinates in the image coordinate system. Here, K represents the edge coordinates in the world coordinate system, and K is the camera intrinsic parameter matrix in the normalized image system. , Z represents the principal point coordinates of the camera, and Z represents the object distance. After transformation, the edge coordinates are linearly corrected.

[0235] By using the camera intrinsic parameter matrix and object distance, the image coordinate system (pixel coordinates) is converted into the world coordinate system (physical size coordinates), thus realizing the quantization from "image edge" to "actual edge of wooden template".

[0236] The specific process is as follows:

[0237] S4.3.1 Obtain camera intrinsic parameter object distance: Based on the calibration results of step S1 (standardized image acquisition system), call the stored intrinsic parameter matrix K and object distance Z=300mm; ,in, , The focal length (in pixels) is the focal length in the x and y directions. , The coordinates of the camera's principal point (in pixels).

[0238] S4.3.2, Image coordinates to camera coordinates: For the edge pixels of the B-spline fitted curve (image coordinates (u,v)), first convert them to normalized coordinates in the camera coordinate system. : ;in, It is the inverse of the intrinsic parameter matrix, calculated as follows: .

[0239] S4.3.3, Convert normalized camera coordinates to world coordinate system coordinates with object distance Z = 300mm : , , ; since the wooden template is a plane, fixed as the object distance, only need to calculate ;

[0240] S4.3.4, Traverse all edge points of the fitting curve, complete the coordinate conversion, and output the world coordinate system coordinates of the wooden template edge (unit: mm), which can be directly used for industrial detection (such as size measurement, defect positioning).

[0241] The specific implementation process of S5 is as follows:

[0242] S5.1, Obtain the edge coordinates to be compared:

[0243] S5.1.1, Extract the detection edge coordinates: From the output results of step S4, obtain the edge coordinate data of the current wooden template, which is expressed in the world coordinate system (unit: mm) and contains the actual physical positions of all edge points of the wooden template;

[0244] S5.1.2, Call the standard template coordinates: From the system database, call out the standard template edge coordinates that completely match the current detection wooden template model. The standard coordinates are pre-established by high-precision laser scanning, with an error controlled within 0.01mm, serving as the reference for quality judgment;

[0245] S5.1.3, Coordinate point matching: Adopt the "nearest neighbor matching" method to one-to-one correspond each point of the detection edge with the nearest point in the standard template edge, ensuring that each pair of comparison points belongs to the same position of the wooden template (such as the same section of the same joint seam);

[0246] S5.2, Core index calculation:

[0247] S5.2.1, Calculate the average error: For each pair of matched coordinate points (detection points and standard points), calculate the straight-line distance (Euclidean distance) between the two points to obtain the error value of a single point. Add the error values of all points and divide by the total number of points to obtain the average error, which reflects the deviation degree of the overall wooden template edge from the standard template.

[0248] S5.2.2, Calculate the Intersection over Union (IoU) of the detected edge: generate a "detected edge area" with a width of 0.275mm (corresponding to 5px in the image) as the boundary of the detected edge coordinates; generate a "standard edge area" with the same width as the boundary of the standard template coordinates; calculate the ratio of the overlapping area (intersection) to the total area after merging (union) of the two areas, and get the IoU, which reflects the consistency of the overall shape of the edge (the closer the value is to 1, the better the consistency).

[0249] S5.3, Threshold judgment and result processing:

[0250] S5.3.1, Set the judgment threshold: the first threshold (average error threshold): set to 0.1mm according to the industrial detection accuracy requirement, which is the maximum overall deviation allowed for the wood template edge; the second threshold (IoU threshold): set to 0.85, which is the minimum matching degree requirement for the edge shape and the standard template.

[0251] S5.3.2, Processing of qualified cases (average error ≤0.1mm and IoU ≥0.85):

[0252] Generate edge labeling visualization image: draw the detected edge with red lines and the standard edge with green lines on the original wood template image (initial standardized wood template image), and highlight the areas with local error exceeding 0.05mm with yellow color;

[0253] Image labeling information: label the average error value, detection time, wood template model, etc. in the corners of the image, and add a qualified identification;

[0254] Output and storage: the image is displayed in real time on the screen of the detection equipment, and is saved to the system database at the same time, associated with the unique number of the wood template, for subsequent traceability.

[0255] S5.3.3, Abnormal case processing (average error >0.1mm or IoU <0.85):

[0256] Priority trigger for re-shooting instruction: automatically send a signal to the image acquisition system to adjust the light source brightness (±10%) and camera exposure parameters (±5%), re-shoot the wood template image and repeat the detection process of S1-S4 (maximum 2 retries);

[0257] If it is still abnormal after re-shooting, trigger an alarm signal: the device beeper sounds (for 3 seconds), the screen displays a red warning of detection abnormality, and sends an alarm information to the production management system (specify the type of abnormality: error exceeds limit or shape mismatch);

[0258] Process pause: the detection system pauses the current batch processing and waits for the operator to check (such as whether the wood template is placed upside down, whether the equipment is offset, etc.).

[0259] S5.4, Data recording and feedback: regardless of whether the detection result is qualified or not, the system will record the following information: average error value, intersection over union, detection time, equipment parameters, etc., and associate the data with the template number for archiving. Qualified data is used for production quality statistics, and abnormal data is used for equipment maintenance and algorithm optimization, forming a complete quality closed loop.

[0260] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A high-precision edge extraction method based on image automatic processing, characterized in that, The method comprises the following steps: constructing a standardized image system and acquiring an initial standardized wood template image through the standardized image system; performing multi-modal fusion preprocessing on the initial standardized wood template image, which comprises sequentially performing adaptive gamma light normalization processing, texture direction weighted mean filtering processing, Laplacian pyramid edge enhancement processing and Sauvola local adaptive binarization processing on the initial standardized wood template to obtain a binary image, wherein the adaptive gamma light normalization processing comprises four sub-steps of image gray scale feature analysis, texture complexity calculation and basic gamma value determination, edge region priority adjustment and gamma value optimization, gamma value smooth transition and light normalization execution; the specific process of sequentially performing adaptive gamma light normalization processing on the initial standardized wood template is as follows: Step A, image gray scale feature analysis: Step A1, calculate the gray level histogram of the initial normalized wood template image, count the gray value distribution of all pixels in the initial normalized wood template image, determine the gray value quantile and quantile , divide the initial normalized wood template image into dark area, normal area and bright area, wherein the pixel gray value of the dark area is , the pixel gray value of the normal area is , and the pixel gray value of the bright area is , , positive and ; Step A2, taking 3x3 as the first local window, traversing each pixel of the initial normalized wood template image, using Sobel operator to calculate the x direction brightness gradient in the first local window respectively and y direction brightness gradient , the brightness gradient value of each window is calculated by formula ;​ Step A3: Based on the brightness gradient value Based on the distribution characteristics, all brightness gradient values ​​of the initial normalized wooden template image are... The area is divided into low-gradient, medium-gradient, and high-gradient regions. This gradient division is then superimposed on the dark, normal, and bright areas defined in step A1 to form nine finely divided regions: dark-low gradient region, dark-medium gradient region, dark-high gradient region, normal-low gradient region, normal-medium gradient region, normal-high gradient region, bright-low gradient region, bright-medium gradient region, and bright-high gradient region. The brightness gradient value of the low-gradient region... The brightness gradient value of the middle gradient region The brightness gradient value of the high gradient region ,in, , It is a positive number, and ; Step B, texture complexity calculation and basic gamma value determination: Step B1, taking 5*5 as a second local window, traversing each pixel in 9 fine partitions, calculating the texture contrast Contrast and texture entropy value Entropy in each second local window through a gray level co-occurrence matrix GLCM, wherein the texture contrast Contrast is used to reflect the clarity of the texture in the second local window, and the texture entropy value Entropy is used to reflect the random complexity of the texture in the second local window; Step B2, calculating the texture complexity coefficient of each of the second local windows by formula Step B3, dividing the second local window of into a low-texture area, into a medium-texture area, is a positive number, and ;​​​ Step B3, determining the base gamma value of each pixel according to the correspondence between the fine partition and the texture complexity partition ; Step B4, determining the base gamma value , if , the base gamma value is set to ; if , the base gamma value is set to , , is positive, and ; Step C, edge region priority adjustment and gamma value optimization: Step C1, calling the edge gradient image generated by the initial standardized wood template image through the improved Canny initial edge detection method, extracting the pixels with the top 20% gradient values as the edge region pixels, and marking the edge regions in the high gradient area in the 9 fine partitions; Step C2, adjust the luminance gradient value according to the luminance region to which the pixel of the high gradient region in the edge region belongs Step C3, if the pixel belongs to the dark region, increase the luminance gradient value by 0.1 to obtain the adjusted gamma value Step C4, if the pixel belongs to the bright region, decrease the luminance gradient value by 0.1 to obtain the adjusted gamma value Step C5, if the pixel belongs to the non-edge region high gradient region or other fine partition region, the adjusted gamma value is Step C6, if the pixel belongs to the non-edge region high gradient region or other fine partition region, the adjusted gamma value is Step C7, if the pixel belongs to the non-edge region high gradient region or other fine partition region, the adjusted gamma value is Step C8, if the pixel belongs to the non-edge region high gradient region or other fine partition region, the adjusted gamma value is Step C3, calculate the edge SNR of the 9 fine partitions after adjustment, wherein, represents the average gray value of the edge region pixels, is the average gray value of the non-edge region noise pixels; if , the adjusted gamma value is determined ; if , the adjustment is repeated in the way of increasing in dark region / decreasing in bright region with a step of 0.05 , until or the boundary and is touched, is the set threshold value; Step D: gamma value smooth transition and light normalization execution: Step D1. With a 7x7 smoothing window, iterate through each pixel of the initial normalized master template image, and calculate the final gamma value for the center pixel of the smoothing window as the weighted average of all 49 pixels in the window, with the center pixel weighted and the remaining 48 pixels weighted in total to give the smoothed final gamma value , ; Step D2, verifying the difference of the smoothed adjacent pixels If the difference is then readjusting the smoothing window weight until the difference of the adjacent pixels is not greater than a preset difference threshold. Step D3, for each pixel of the initial normalized wood template image, by the formula performing gamma transformation, wherein is the gray value of the pixel of the initial normalized wood template image, is the gray value of the pixel after the gamma transformation, the adaptive gamma illumination normalization processing is completed, and an illumination normalized wood template image is output. The specific process of performing texture direction weighted mean filtering processing, Laplacian pyramid edge enhancement processing and Sauvola local adaptive binarization processing is as follows: The texture direction weighted mean filtering processing: the texture main direction of the wood template in the illumination normalized wood template image is determined by a Sobel operator, a filter core matched with the texture main direction is used to perform weighted mean filtering on the illumination normalized wood template image to obtain a filtered image, wherein the filter core of the texture direction weighted mean filtering is an ellipse, and the long axis is parallel to the texture main direction; the weight formula of the weighted mean filtering is ; in the formula, d is the distance of a pixel from the center of the filter core, is the half length of the long axis of the filter core; Laplacian pyramid edge enhancement processing: decomposing the filtered image into a multi-layer Laplacian pyramid, respectively performing contrast stretching and threshold enhancement on different layers, and reconstructing to obtain an edge enhancement image; Sauvola local adaptive binarization processing: calculating the mean and standard deviation of pixels in a third local window of a preset size in the edge enhanced image, determining a binarization threshold value based on the mean and standard deviation, outputting a binarized image, and the calculation formula is: , wherein, is the mean of pixels in the third local window, is the standard deviation of pixels in the third local window, k and R are constants, and , ; The Otsu algorithm is used to determine the high threshold. With low threshold After performing edge detection on the binarized image, the distance is... Furthermore, disconnected edge segments with consistent gradient directions are linearly connected to generate an initial edge image and an edge gradient map. The initial edge image is then input into a pre-trained lightweight adaptive U-Net model, which outputs an edge probability map. The lightweight adaptive U-Net model uses depthwise separable convolutions to construct an encoder, fuses features through skip connections, and is trained using a dynamic weighted joint loss function. , It is a positive number; performing morphological opening operation denoising processing, B-spline curve smooth fitting processing and coordinate system calibration processing on the edge probability image in sequence to obtain wood template edge coordinates; calculating the average error of the wood template edge coordinates and the standard template edge coordinates, if the average error is not greater than a first threshold value, outputting the edge coordinates and edge label visualization image; if the average error is greater than the first threshold value or the edge detection intersection over union is less than a second threshold value, triggering a retake instruction or an alarm signal.

2. The high-precision edge extraction method based on image automatic processing according to claim 1, characterized in that, The specific process of constructing a standardized image system and acquiring an initial standardized wood template image is as follows: constructing a standardized image acquisition system comprising a pixel industrial camera, a fixed focus lens, a ring light source and a standard checkerboard calibration component; setting the shooting parameters of object distance, shutter speed and ISO; calibrating the camera intrinsic parameters using the standard checkerboard before each batch of detection to obtain an intrinsic matrix K; shooting the wood template image to obtain an initial standardized wood template image.

3. The high-precision edge extraction method based on image automatic processing according to claim 1, characterized in that, B3, the base gamma value of each pixel The specific implementation of the determination rule is that Dark - low gradient region: if in low texture region, ; if in medium texture region, ; if in high texture region, ; Dark - mid gradient zone: if in low texture zone, ; if in mid texture zone, ; if in high texture zone, ; Dark - high gradient zone: if in low texture zone, ; if in medium texture zone, ; if in high texture zone, ; Bright - low gradient region: if in low texture region, ; if in medium texture region, ; if in high texture region, ; bright-to-medium gradient zone: if in low-texture zone, ; if in medium-texture zone, ; if in high-texture zone, ; bright - high gradient zone: if in low texture zone, ; if in medium texture zone, ; if in high texture zone, ; Normal part-low gradient zone, normal part-medium gradient zone, normal part-high gradient zone: were 1.

0.

4. The high-precision edge extraction method based on image automatic processing according to claim 1, characterized in that, The lightweight adaptive U-Net model comprises an encoder and a decoder, the encoder comprises four down-sampling modules, each of the down-sampling modules is composed of a depth separable convolution layer, a batch normalization layer and a ReLU activation layer, the decoder comprises four up-sampling modules, and each of the up-sampling modules is fused with features of a corresponding down-sampling module of the encoder through adaptive weight skip connection; The process of training the lightweight adaptive U-Net model is as follows: An initial standardized wood template image set containing 10,000 images of different materials and different defect types is constructed, which is divided into training set, validation set and test set according to 8:1:1; the Adam optimizer is adopted, the initial learning rate is set, the learning rate is attenuated to the set multiple of the previous round every 20 rounds and iterative training is performed, so that the intersection over union of the edge detection of the validation set is greater than the set threshold, and the dynamic weight joint loss function formula of the lightweight adaptive U-Net model is , wherein, is the cross-entropy loss, is the Dice coefficient, and , wherein A is the predicted edge pixel set, B is the real edge pixel set, is a dynamic adjustment coefficient.

5. The high-precision edge extraction method based on image automatic processing according to claim 1, characterized in that, The specific process of obtaining the wood template edge coordinates by sequentially performing morphological opening operation denoising processing, B-spline curve smoothing fitting processing and coordinate system calibration processing on the edge probability map is as follows: Morphological opening operation denoising processing: a rectangular structural element is used to first erode and then dilate the edge probability map, so as to remove isolated noise blocks and obtain denoised edge pixel points; B-spline curve smoothing fitting processing: the denoised edge pixel points are grouped, and a B-spline curve is used for fitting to obtain a fitting curve, so as to control the fitting error; Coordinate system calibration: based on the standardized image system, the edge coordinates of the fitting curve in the image coordinate system are converted into world coordinate system coordinates, so as to output the wood template edge coordinates.

6. The high-precision edge extraction method based on image automatic processing according to claim 5, characterized in that, The size of the rectangular structure element used in the morphological opening operation is 3x3, the erosion operation removes isolated noise blocks with an area of 9, and the dilation operation restores the edge main body profile. The B-spline curve is a cubic curve, the edge pixel points are grouped according to a set number, and the fitting error is not greater than a set error threshold; If the average distance between the fitting curve and the original edge point is greater than a set distance threshold, control points are added to a local edge segment formed by the fitting group and its left and right adjacent fitting groups to adjust the fitting curve.

7. The high-precision edge extraction method based on image automatic processing according to claim 5, characterized in that, The conversion formula of the coordinate system calibration is: , , wherein, is an edge coordinate in an image coordinate system, is an edge coordinate in a world coordinate system, K is a camera intrinsic matrix in a normalized image system, , is a camera principal point coordinate, and Z is an object distance.

8. The high-precision edge extraction method based on image automatic processing according to claim 7, characterized in that, The intrinsic matrix K is specifically represented as follows: wherein, , is the focal length in x, y direction, , is the principal point coordinate of the camera.

Images