Metal sawtooth generation quality detection method and system based on deep learning

Abstract

This invention discloses a deep learning-based method and system for detecting the generation quality of metal sawtooth serrations, belonging to the field of image data processing. The method includes: acquiring a high-dynamic backlight image of a sawtooth-pressed component and inputting it into a feature extraction network composed of a backbone convolutional network and a physical prior guidance layer; truncating high-frequency main contours using background entropy and hard attention, and enhancing burrs through coupled logarithmic and exponential nonlinear activation; extracting connected components from the enhanced feature map, and constructing an adaptive logistic gating attenuation layer based on the ratio of spatial morphological divergence to topological clustering density; extracting tangentials from the endpoints of residual discontinuous defects and determining homologous breakpoint pairs based on Euclidean distance and yaw angle, and optimizing the generation of a bridging tensor using bending energy functional as an unsupervised loss to repair pseudo-fractures. This method achieves enhancement, cleaning, and topological closure of micro-defects, improving the completeness and detection accuracy of sawtooth generation quality assessment.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method and system for detecting the quality of metal jagged edges generated based on deep learning. Background Technology

[0002] In the manufacturing of ultra-densely bonded multilayer silicon steel sheet components for micro stepper motors, traditional sawtooth inspection is often limited to the measurement of macroscopic geometric dimensions. However, the quality bottleneck in the sawtooth generation process actually lies in the tearing and adhesion burrs that easily occur at the edges during pressing and cutting. These microscopic defects can easily affect stator-rotor interference and motor performance.

[0003] High dynamic range backlight transmission imaging is used to obtain defect information of the pressed parts. When conventional deep learning convolutional neural networks are introduced to extract these micro defects, the network often faces a serious data collapse problem: the jagged macro edges strongly block the backlight, which manifests as a high-frequency step gradient signal; while micro burrs caused by poor generation only cause a gradual decrease in transmittance.

[0004] This extreme contrast difference inevitably causes the network's forward activation to be captured by the jagged high-frequency macroscopic contours, and the microscopic burr signals are completely submerged in the feature map, making it extremely easy to generate gradient vanishing during training.

[0005] Furthermore, even if an attention mechanism is added to the network, the imaging speckle and sensor background noise will be amplified simultaneously, generating a large number of sheet-like artifacts. Moreover, the microscopic tearing defects of the jagged edges are physically continuous, but due to the weak feature response, they are easily degraded into discrete and discontinuous response fragments in the feature map after extraction and denoising, forming topological breaks in the feature map.

[0006] This topological break in the feature map causes traditional deep learning networks' bounding boxes or segmentation masks to fail to output geometric results that conform to metallic coherence.

[0007] Therefore, there is an urgent need for a deep learning detection method constrained by smoothness priors (physically equivalent to the minimum bending energy constraint of metal tearing) and morphological methods, which can selectively excite weak features, deterministically clean artifacts, and perform topological reconstruction on discontinuous defects in feature maps to solve the problem of microscopic quality assessment in the serration generation process. Summary of the Invention

[0008] To solve the above-mentioned technical problems, the present invention provides a method, the technical solution of which is as follows: In a first aspect, the present invention provides a method for detecting the quality of metal sawtooth generation based on deep learning, comprising: A high dynamic backlight transmission image of the sawtooth press-fitted part is acquired and input into a deep learning-based feature extraction network. The feature extraction network includes a backbone convolutional module and a physical prior guidance layer. The physical prior guidance layer includes a feature mapping layer, an artifact suppression layer, and a topology reconstruction output layer. In the feature mapping layer, the initial feature map output by the backbone convolution module is used to construct an extreme value suppression hard attention mask using the background transmission benchmark prior factor to truncate the feature transmission of the serrated macro high-frequency main contour, thereby obtaining a residual weak defect feature map; and a logarithmic and exponential coupled nonlinear mapping activation is applied to the residual weak defect feature map to generate a reinforced adhesion response feature map. In the artifact suppression layer, activated connected components are extracted from the enhanced adhesion response feature map, the spatial morphological divergence factor and topological clustering density index of each connected component are calculated, an adaptive decay tensor is generated through the logistic gating function, and a deterministic spatial feature deactivation operation is performed on the feature map to suppress artifact regions where the ratio of the topological clustering density index to the spatial morphological divergence factor exceeds a preset critical distribution benchmark value while preserving microscopic spur features, thus obtaining an artifact suppression feature map; In the topology reconstruction output layer, the local tangential gradient direction vectors of the discontinuous endpoints of the artifact suppression feature map are extracted. Based on the spatial Euclidean distance constraint and the vector yaw angle difference coefficient, the same source discontinuity pairs are determined, and an unsupervised loss function based on the bending energy functional is constructed for optimization and solution. A smooth bridging tensor is generated to complete the topology structure of the defect contour and complete the sawtooth generation quality detection.

[0009] Preferably, the feature transfer of the truncated jagged macroscopic high-frequency main contour includes: comparing the gradient response of each pixel with a preset saturated pixel truncating threshold, determining pixels whose gradient response exceeds the saturated pixel truncating threshold as strong light interference of the jagged macroscopic contour, and triggering the network hard attention mechanism to force the forward activation response value of the pixel in the feature transfer network to 0.

[0010] Preferably, the logarithmic and exponential coupled nonlinear mapping activation includes: statistically analyzing the gray-level probability distribution in the neighborhood of each pixel using a sliding window and calculating the local information entropy; generating local weights for nonlinear mapping activation based on the local information entropy distribution; performing nonlinear stretching on the response value of the residual weak defect feature map using a logarithmic function and sharpening the local gradient using an exponential function; and multiplying the stretching and sharpening results by the local weights to obtain the final enhanced adhesion response feature map; wherein the intensity of logarithmic stretching is controlled by a first nonlinear mapping adjustment coefficient, and the intensity of exponential sharpening is controlled by a second nonlinear mapping adjustment coefficient.

[0011] Preferably, obtaining the final enhanced adhesion response feature map includes: adding 1 to the ratio of the response value of each pixel in the residual weak defect feature map to the background transmission reference prior factor, and then taking the natural logarithm to obtain a logarithmic stretching term; performing an exponential operation on the product of the local gradient of the pixel and the second nonlinear mapping adjustment coefficient to obtain an exponential sharpening term; and multiplying the local weight, the first nonlinear mapping adjustment coefficient, the logarithmic stretching term, and the exponential sharpening term to obtain the pixel value of the enhanced adhesion response feature map, thus obtaining the enhanced adhesion response feature map.

[0012] Preferably, the spatial morphological divergence factor is the ratio of the largest eigenvalue to the second largest eigenvalue of the pixel coordinate covariance matrix of the connected domain; the topological clustering density index is the ratio of the actual active pixel area of ​​the connected domain to the geometric area of ​​the smallest circumscribed convex hull.

[0013] Preferably, the step of generating an adaptive decay tensor through a logistic gating function includes: calculating the ratio of the topological clustering density index to the spatial morphological divergence factor; subtracting a preset critical distribution benchmark value from the ratio; multiplying the ratio by a preset activation function decay slope factor; and then calculating the inhibition factor in the decay tensor through a logistic function, such that the inhibition factor rapidly approaches 1 when the ratio exceeds the preset critical distribution benchmark value; and multiplying the feature map element-wise with the complement of the inhibition factor to perform deterministic spatial feature deactivation.

[0014] Preferably, the determination of the same breakpoint pair based on the spatial Euclidean distance constraint and the vector yaw angle difference coefficient includes: calculating the spatial Euclidean distance between the two discontinuous endpoints; only when the spatial Euclidean distance is less than or equal to a preset search radius threshold, further calculating the inner product inverse cosine of the local tangential gradient direction vectors of the two discontinuous endpoints to obtain the vector yaw angle difference coefficient; comparing the vector yaw angle difference coefficient with a preset small tolerance confidence region boundary; if it does not exceed the boundary, it is determined to be a same breakpoint pair.

[0015] Preferably, the optimization solution using the unsupervised loss function includes: constructing a bending energy loss function consisting of a weighted sum of a bending stiffness penalty term and a tensile tension penalty term, using the endpoint coordinates of the same breakpoint pair and the local tangential gradient direction vector as boundary constraints, and generating a bridging compensation fitting path by minimizing the bending energy loss function.

[0016] Preferably, the bending stiffness penalty term is a weighted average of the square of the second derivative of the fitted path, and the tensile tension penalty term is a weighted average of the square of the first derivative of the fitted path; the boundary constraints require that the coordinates of the fitted path coincide at both endpoints and that the tangents are parallel to the corresponding local tangential gradient direction vectors.

[0017] Secondly, this invention provides a deep learning-based metal sawtooth generation quality inspection system, employing the following technical solution: A deep learning-based metal sawing quality inspection system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the deep learning-based metal sawing quality inspection method described above is implemented.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing a background transmission baseline prior factor and a saturation pixel cutoff threshold, a hard attention suppression mask for the macroscopic high-frequency contour of jagged edges is constructed, which cuts off the gradient interference caused by strong light masking. Then, a nonlinear activation layer coupled with logarithmic stretching and exponential sharpening is constructed, which enables the micro-burrs associated with jagged edges with extremely low values ​​to obtain specific gains in the deep feature map. This solves the problem that conventional networks are prone to getting trapped in local optima when extracting weak signals, and reduces the risk of missing micro-defects at the jagged edge.

[0019] 2. By utilizing the geometric dual-dimensional discrimination mechanism extracted by the network in the feature tensor space, an adaptive decay tensor is generated for each candidate region through the logistic gating function. This tensor is equivalent to the deterministic spatial feature deactivation operation based on physical prior calculation, which enables the full speckle artifacts with excessive geometric ratios to be suppressed by the network depth, while slender and irregular defects that conform to the mechanical properties of metal tearing are preserved. This maximizes the purity of the feature matrix without introducing randomness and reduces the risk of misjudgment when detecting micro-defects with jagged edges.

[0020] 3. By extracting the endpoint tangential tensor to determine the homology relationship, and using the bending energy functional that characterizes the smoothness of the curve as the loss function for unsupervised optimization, this reconstruction mechanism utilizes the smoothness prior inherent in metal deformation mechanics to overcome the limitations of defect breakage in the feature map. It smoothly reconstructs the originally discontinuous pixels in the feature map into a defect contour that conforms to continuity, reducing the measurement error caused by feature map breakage and providing highly reliable data support for the accurate evaluation of sawtooth generation quality. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the implementation of the deep learning-based metal sawtooth generation quality inspection method according to an embodiment of the present invention. Detailed Implementation

[0022] Step S1: Acquire the image and construct a deep learning-based jagged edge feature extraction network.

[0023] Acquire high dynamic range backlit transmission images of the pressed parts. In the quality inspection of metal serration generation, the imaging consists of a large-area uniform array LED backlight, a telecentric imaging lens, and a high dynamic range CMOS industrial camera.

[0024] Camera exposure time set The obtained high dynamic range backlight transmission image is saved as a two-dimensional grayscale tensor. The grayscale value range to The camera exposure time is a hyperparameter that can be adjusted by the implementer according to the specific implementation scenario.

[0025] To overcome the shortcomings of conventional convolutional neural networks in weak texture extraction, a forward activation gated channel directly guided by physical priors is used to capture microscopic tearing and adhesion defects that accompany the sawtooth generation process. The sawtooth edge feature extraction network is a convolutional neural network guided by physical prior activation.

[0026] The backbone convolutional module uses the first three convolutional blocks of a ResNet-18 pre-trained on ImageNet as a feature extractor, removes the last global average pooling layer and fully connected layer, and outputs a feature map with a spatial scale of 1 / 8 of the input.

[0027] Following the backbone convolutional module, a 1×1 convolution is appended to reduce the number of channels to 64, yielding the initial feature map input to the physical prior guidance layer. .

[0028] To ensure that the feature map contains channels with physical transmittance response, an auxiliary regression task is introduced during the training phase: the first channel of the 1×1 convolution output is normalized to the [0,1] interval using the Sigmoid function, denoted as... .

[0029] The input high dynamic range backlight transmission image is simultaneously grayscaled and normalized to the [0,1] interval, denoted as... ; with mean square error As an auxiliary loss function, it is used in conjunction with the segmentation loss function of the backbone network. Joint optimization is performed using weighted summation.

[0030] After training, It approximates the physical transmittance distribution of the input image, so it can be directly used as a channel with physical transmittance response during the inference stage without having to recalculate the auxiliary loss.

[0031] Step S2: In the network feature extraction layer, a hard attention mask for extreme value suppression is constructed using background prior factors, and a coupled nonlinear activation function is applied to generate a feature map of enhanced adhesion response.

[0032] First, the background transmission baseline prior factor is obtained by high-frequency sampling and fitting under the no-load transmission state. The empirical value is Among them, the background transmission reference prior factor is an empirical value, which can be adjusted by the implementer according to the specific implementation scenario, so that the background transmission reference prior factor can be used as the background reference value.

[0033] from The channel with physical transmittance response is selected and normalized to the [0,1] interval so that it can be compared with the background transmittance reference prior factor. Direct comparison, this channel is denoted as .

[0034] The high-frequency main profile of the sawtooth is in The high step gradient is manifested in the deep network, and therefore, in order to truncate the propagation of high-frequency contours in the deep network, calculation is required. Gradient magnitude of each pixel Set the saturated pixel cutoff threshold The empirical value is 0.12, which can be adjusted by the implementer according to the specific implementation scenario.

[0035] when When the network determines that the pixel is a jagged macroscopic contour disturbance, it triggers a hard attention mechanism, forcibly setting the activation value of the pixel in feature propagation to 0, thus obtaining a residual weak defect feature map. .

[0036] After that, The Sobel operator is used to calculate the first-order partial derivatives in the horizontal and vertical directions for each pixel, resulting in a local gradient vector. The magnitude of this vector is then taken as the local gradient magnitude. .

[0037] Before performing subsequent nonlinear mapping activation, the gray-level probability distribution in the local neighborhood is statistically analyzed using a sliding window, and the local information entropy is calculated to obtain the local constraint weights used for activation mapping. Its value is the normalized ratio of the neighborhood information entropy of the pixel to the maximum possible entropy.

[0038] Subsequently, the feature map is input to a custom nonlinear coupled activation layer, whose forward mapping activation formula is:

[0039] In the formula, the logarithmic term nonlinearly stretches the extremely low light transmittance response of the near-zero region, and the exponential term performs gradient sharpening on the slight fluctuations.

[0040] To control tensile strength, an empirical value is used. , Control the sharpening level, using an empirical value. , and These are all hyperparameter empirical values, which can be adjusted by the implementer according to the specific implementation scenario.

[0041] After activation by this layer, the network outputs a feature map of enhanced adhesion response. Selective positive excitation of micro-burrs on the edge of the serrated tooth was achieved in the feature space.

[0042] Step S3: In the network artifact suppression layer, a decay tensor is generated based on the two-dimensional physical constraints through the logistic gating function, and deterministic spatial feature deactivation is performed.

[0043] Nonlinear activation layers, while stimulating weak signals, also amplify speckle noise, forming patchy artifacts.

[0044] Network for feature maps The response value is greater than the threshold. Extracting eight-neighbor connected components from pixels. Experience value It can be adjusted by the implementer according to the specific implementation scenario, and geometric indicators in two dimensions can be calculated: One is the spatial morphological divergence factor. The ratio of the variance of the principal inertial axis to the variance of the secondary inertial axis characterizes the degree to which the defect deviates from the elongation of the circular spot.

[0045] The second is the topological aggregation density index. The ratio of the actual active pixel area to the minimum circumscribed convex hull area represents the density of the region filling.

[0046] Among them, spatial morphological divergence factor By extracting the pixel coordinate covariance matrix of the connected components and calculating the eigenvalues ​​of the matrix. , ( ≥ )get, = A larger value indicates a more elongated shape of the region; Topological clustering density index , This represents the number of active pixels within the connected component. This represents the total number of pixels contained in the smallest outer convex hull of the connected component.

[0047] To achieve artifact filtering, the network constructs an adaptive gated decay layer based on the logistic function, outputting decay tensor weights. :

[0048] In the formula, The critical distribution baseline value is empirically selected. , To activate the attenuation slope factor, an empirical value is selected. , and These are all hyperparameter empirical values, which can be adjusted by the implementer according to the specific implementation scenario.

[0049] If the ratio of the connected components exceeds the critical point, then rapidly approaching The network interacts with the decay tensor element-wise with the feature map, i.e., multiplies it by... This is equivalent to a deterministic spatial feature deactivation strictly guided by physical formulas, with deep attenuation of full-bodied sheet artifacts, and only precisely retaining the slender, irregular defect response.

[0050] In practice, the burrs and tears that accompany metal extrusion are usually continuous; however, during the imaging process, due to the extremely narrow width of these micro-burrs, they are limited by the optical imaging limit and will inevitably appear as discrete, discontinuous pseudo-discontinuities on the feature map after cleaning.

[0051] Step S4: In the network topology reconstruction output layer, extract the tangential gradient for the defect breakpoint and construct the bending energy loss function for optimization to reconstruct the topology continuity.

[0052] After artifact suppression, weak spur features may appear as discontinuous fragments in the feature map due to response loss. Given the physical continuity of metal sheet spurs, it is necessary to restore their topological continuity in the feature space. The network undergoes topology reconstruction, and the parser extracts the discontinuous endpoints in the feature map after artifact suppression. and Through local Calculate the normalized local tangential gradient direction vector in the neighborhood and , among which The neighborhood value is an empirical value and can be adjusted by the implementer according to the specific implementation scenario.

[0053] Calculate the spatial Euclidean distance between any two endpoints. ,when Then, the vector yaw angle difference coefficient is further calculated. ,like If the breakpoints are in the same origin, they are considered a pair of breakpoints; otherwise, they are not considered a pair of breakpoints.

[0054] in, An empirical value is used to determine the search radius threshold. , Empirical values ​​are taken for the tolerance confidence region limits. , and These are hyperparameter empirical values ​​that can be adjusted by the implementer according to the specific implementation scenario.

[0055] To smoothly reconstruct fracture spikes, the network constructs an unsupervised loss function based on the bending energy functional. :

[0056] In the formula, The bending stiffness penalty coefficient is empirically taken as [value missing]. , The tensile tension penalty coefficient is empirically taken as [value missing]. ,in, and These are hyperparameter empirical values ​​that can be adjusted by the implementer according to the specific implementation scenario.

[0057] Specifically, the boundary constraints are: , ; , Then, the Gauss-Newton iteration method is used to apply the constrained bending energy loss function to the above-mentioned problem. Minimize the solution to obtain a smooth bridging path that satisfies the constraints.

[0058] The underlying layer of the network uses the Gauss-Newton iterative method to minimize the constrained unsupervised loss function and outputs the optimal path tensor that achieves the minimum value.

[0059] Finally, the bridging tensor is mapped back to the feature layer to complete the breaks in the feature map and output a topologically complete detection result, which is used for the final evaluation of the jagged edge generation quality.

[0060] During the training phase, 500 images of pressed parts containing various typical serrated defects were collected. Professionals annotated the micro-defects at the pixel level to construct a binary ground truth mask. The 500 images were empirically selected hyperparameter values, which can be adjusted by the implementer according to the specific implementation scenario.

[0061] The training image is input into the backbone convolutional module and the feature mapping layer and artifact suppression layer in the physical prior guidance layer, and the artifact suppression feature map is output.

[0062] The loss function used for training is the binary cross-entropy loss between the artifact suppression feature map and the ground truth mask. Only the parameters of the backbone convolutional module are updated; all coefficients in the physical prior guiding layer are treated as fixed hyperparameters and do not participate in gradient updates.

[0063] The Adam optimizer is used to train the model for 100 epochs with a learning rate of 1e-4 to obtain the final detection model. The learning rate of 1e-4 and epoch=100 are empirical values ​​that can be adjusted by the implementer according to the specific implementation scenario.

[0064] During the inference phase, the weights of the backbone convolutional modules are frozen, the image to be tested is forward-propagated to the artifact suppression feature map, and then it enters the topology reconstruction output layer. Discontinuous endpoints are extracted, homologous breakpoint pairs are determined, and a smooth bridging tensor is generated using the Gauss-Newton method with bending energy functional as unsupervised loss to complete the topology continuity reconstruction and output the final quality detection result.

[0065] This invention also discloses a deep learning-based metal sawtooth generation quality inspection system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the deep learning-based metal sawtooth generation quality inspection method of this invention.

[0066] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0067] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting the quality of metal sawteeth based on deep learning, characterized in that, include: A high dynamic backlight transmission image of the sawtooth press-fitted part is acquired and input into a deep learning-based feature extraction network. The feature extraction network includes a backbone convolutional module and a physical prior guidance layer. The physical prior guidance layer includes a feature mapping layer, an artifact suppression layer, and a topology reconstruction output layer. In the feature mapping layer, the initial feature map output by the backbone convolution module is used to construct an extreme value suppression hard attention mask using the background transmission benchmark prior factor to truncate the feature transmission of the serrated macro high-frequency main contour, thereby obtaining a residual weak defect feature map; and a logarithmic and exponential coupled nonlinear mapping activation is applied to the residual weak defect feature map to generate a reinforced adhesion response feature map. In the artifact suppression layer, activated connected components are extracted from the enhanced adhesion response feature map, the spatial morphological divergence factor and topological clustering density index of each connected component are calculated, an adaptive decay tensor is generated through the logistic gating function, and a deterministic spatial feature deactivation operation is performed on the feature map to suppress artifact regions where the ratio of the topological clustering density index to the spatial morphological divergence factor exceeds a preset critical distribution benchmark value while preserving microscopic spur features, thus obtaining an artifact suppression feature map; In the topology reconstruction output layer, the local tangential gradient direction vectors of the discontinuous endpoints of the artifact suppression feature map are extracted. Based on the spatial Euclidean distance constraint and the vector yaw angle difference coefficient, the same source discontinuity pairs are determined, and an unsupervised loss function based on the bending energy functional is constructed for optimization and solution. A smooth bridging tensor is generated to complete the topology structure of the defect contour and complete the sawtooth generation quality detection. 2.The deep learning-based metal sawtooth generation quality detection method of claim 1, wherein, The feature transfer of the truncated jagged macroscopic high-frequency main contour includes: comparing the gradient response of each pixel with a preset saturated pixel truncating threshold, determining pixels whose gradient response exceeds the saturated pixel truncating threshold as strong light interference of the jagged macroscopic contour, and triggering the network hard attention mechanism to force the forward activation response value of the pixel in the feature transfer network to 0. 3.The deep learning-based metal sawtooth generation quality detection method of claim 1, wherein, The activation of the logarithmic-exponential coupled nonlinear mapping includes: The gray-level probability distribution in the neighborhood of each pixel is statistically analyzed using a sliding window, and the local information entropy is calculated. Based on the local information entropy distribution, local weights are generated for nonlinear mapping activation. The response value of the residual weak defect feature map is nonlinearly stretched using a logarithmic function and sharpened using an exponential function. The stretching and sharpening results are then multiplied by the local weights to obtain the final enhanced adhesion response feature map. The intensity of logarithmic stretching is controlled by the first nonlinear mapping adjustment coefficient, while the intensity of exponential sharpening is controlled by the second nonlinear mapping adjustment coefficient. 4.The deep learning-based metal sawtooth generation quality detection method of claim 3, wherein, The final enhanced adhesion response feature map obtained includes: The logarithmic stretching term is obtained by adding 1 to the ratio of the response value of each pixel in the residual weak defect feature map to the background transmission reference prior factor, and then taking the natural logarithm. The product of the local gradient of the pixel and the second nonlinear mapping adjustment coefficient is exponentially multiplied to obtain the exponential sharpening term; Multiply the local weight, the first nonlinear mapping adjustment coefficient, the logarithmic stretching term, and the exponential sharpening term to obtain the pixel values ​​of the enhanced adhesion response feature map, thus obtaining the enhanced adhesion response feature map. 5.The deep learning-based metal sawtooth generation quality detection method of claim 1, wherein, The spatial morphological divergence factor is the ratio of the largest eigenvalue to the second largest eigenvalue of the pixel coordinate covariance matrix of the connected domain; the topological clustering density index is the ratio of the actual active pixel area to the geometric area of ​​the smallest circumscribed convex hull of the connected domain. 6.The deep learning-based metal sawtooth generation quality detection method of claim 1, wherein, The generation of the adaptive decay tensor through the logistic gating function includes: The ratio of the topological clustering density index to the spatial morphological divergence factor is calculated. After subtracting the preset critical distribution benchmark value from the ratio, it is multiplied by the preset activation function decay slope factor. Then, the inhibition factor in the decay tensor is calculated by the logistic function, so that when the ratio exceeds the preset critical distribution benchmark value, the inhibition factor quickly approaches 1. The feature map is multiplied element-wise by the complement of the inhibition factor to perform deterministic spatial feature inactivation. 7.The deep learning-based metal sawtooth generation quality detection method of claim 1, wherein, The method for determining pairs of original breakpoints based on spatial Euclidean distance constraints and vector yaw angle difference coefficients includes: Calculate the spatial Euclidean distance between the two discontinuous endpoints. Only when the spatial Euclidean distance is less than or equal to a preset search radius threshold, further calculate the inner product inverse cosine of the local tangential gradient direction vectors of the two discontinuous endpoints to obtain the vector yaw angle difference coefficient. The vector yaw angle difference coefficient is compared with the preset small tolerance confidence domain limit. If it does not exceed the limit, it is determined to be a pair of breakpoints from the same source. 8.The deep learning-based metal sawtooth generation quality detection method of claim 1, wherein, The optimization solution using an unsupervised loss function includes: Using the endpoint coordinates of the same origin breakpoint pair and the local tangential gradient direction vector as boundary constraints, a bending energy loss function is constructed, which is a weighted sum of the bending stiffness penalty term and the tensile tension penalty term. A bridging compensation fitting path is generated by minimizing the bending energy loss function. 9.The deep learning-based metal sawtooth generation quality detection method of claim 8, wherein, The bending stiffness penalty term is a weighted sum of the squares of the second derivative of the fitted path, and the tensile tension penalty term is a weighted sum of the squares of the first derivative of the fitted path; the boundary constraints require that the coordinates of the fitted path coincide at both endpoints and that the tangents are parallel to the corresponding local tangential gradient direction vectors.

10. A deep learning-based metal saw tooth generation quality detection system, characterized by, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the deep learning-based metal sawtooth generation quality inspection method according to any one of claims 1-9.

Images