Metal surface drilling positioning method based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method for drilling and locating holes on metal surfaces, comprising the following steps: acquiring multiple frames of images of the metal surface to construct an original multi-frame image sequence; performing weighted fusion on the original multi-frame image sequence to obtain a fused image, a specular mask, and an edge map; constructing an improved HorNet backbone network to output multi-scale texture feature maps; inputting the multi-scale texture feature maps into a texture-coordinate joint head to generate a texture saliency map and initial drilling coordinate parameters; inputting the initial drilling coordinate parameters into a projection-aware coordinate transformer to obtain drilling positioning coordinates in the machine tool workpiece coordinate system; constructing a multi-task loss function and performing end-to-end training; inputting the fused image into the trained and converged improved HorNet network to output the drilling positioning coordinates. This invention can achieve high-precision drilling and locating under complex metal surface conditions and is suitable for intelligent manufacturing and CNC machining scenarios.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and computer vision technology, and in particular to a method for locating drill holes on metal surfaces based on artificial intelligence. Background Technology

[0002] With the increasing demand for high-precision automated positioning in intelligent manufacturing and CNC machining, visual positioning technology for drilling positions on metal surfaces has attracted widespread attention. Existing drilling positioning methods mostly rely on single-frame images or traditional edge / corner detection and general convolutional networks for coordinate inference. Some methods use simple image averaging or filtering to fuse multiple frames and map image coordinates to the machine tool coordinate system using camera calibration. However, these methods generally suffer from the following problems in practical applications:

[0003] The prevalence of smooth, brushed, and highly reflective metal surfaces makes traditional edge / corner detection or general vision models sensitive to lighting and reflection, and difficult to stably extract features under weak textures. Acquisition, calibration, and equipment operation are affected by vibration, temperature drift, oil contamination, and changes in ambient light, causing the geometric relationship between imaging and equipment to drift easily, and making it difficult for image coordinates and machine tool workpiece coordinates to remain consistent over a long period. Coordinate transformation is mostly based on ideal geometric assumptions, and installation errors and posture deviations are difficult to be dynamically corrected. The accumulation of system errors makes it difficult to guarantee millimeter-level accuracy. There is a lack of a unified optimization goal and evaluation mechanism that takes into account both texture stability and positioning accuracy, and the algorithm's generalization and robustness are insufficient across different materials and processing states. Positioning results usually lack confidence and anomaly self-checking, making it difficult to suppress false detections and missed detections in a timely manner, and still relying on manual verification and trial drilling on site. Insufficient adaptation to the real-time cycle time of CNC systems results in high algorithm latency and resource consumption, high maintenance and calibration costs, and difficulty in achieving stable mass production under complex working conditions.

[0004] Therefore, how to provide an artificial intelligence-based method for drilling and locating holes on metal surfaces is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an artificial intelligence-based method for locating drill holes on metal surfaces. This invention comprehensively employs multi-frame image weighted fusion, an improved HorNet deep network guided by specular mask and edge map priors, a texture-coordinate joint head, and a projection-aware coordinate transformer. It has the advantages of high positioning accuracy under weak texture and high reflectivity conditions, convenient integration with machine tool coordinate systems, and good efficiency and stability in CNC machining processes.

[0006] The method for locating drill holes on a metal surface based on artificial intelligence according to an embodiment of the present invention includes the following steps:

[0007] Multiple frames of images of the metal surface were acquired and preprocessed to obtain the original multi-frame image sequence;

[0008] The original multi-frame image sequence is weighted and fused to obtain a fused image, and a mirror mask and edge map are generated at the same time.

[0009] An improved HorNet backbone framework is constructed, which includes recursively gated convolutional units, adaptive order control units, complex wavelet gated units, and large kernel dilated convolutional units.

[0010] The system receives a fused image, a mirror mask, and an edge map as input. These are then combined into multi-channel input data through channel stitching. A normalization method is used to standardize the pixel values ​​to obtain standardized input data.

[0011] The recursive gated convolutional unit performs recursive convolution on standardized input data to extract local features. Each convolution is performed by convolution operation and linear weighted summation to obtain the convolution result. After the convolution calculation, the mirror mask and edge map are introduced into the gate function as bias information. The gate function uses the Sigmoid function to perform a nonlinear transformation on the convolution result and the value after weighted bias of the mirror mask and edge map. Controlled features are generated between the convolution output and the input by element-wise multiplication, and the first feature map is output.

[0012] The adaptive step control unit uses the feature gradient analysis method to calculate the gradient change magnitude between two adjacent recursive convolution results on the first feature map and compares it with a preset threshold. When the change magnitude exceeds the threshold, it continues to iterate the convolution; when the change magnitude is lower than the threshold, it terminates the iteration and outputs the second feature map.

[0013] The complex wavelet gating unit performs complex wavelet decomposition on the second feature map to obtain multi-scale multi-directional sub-bands containing amplitude and phase information. Based on each multi-scale multi-directional sub-band, the phase consistency index is calculated, and the phase consistency index is superimposed with the input feature through a channel-by-channel weighting method to output a third feature map that simultaneously contains spatial and frequency domain information.

[0014] The large kernel dilated convolution unit performs convolution operation on the third feature map. During the convolution process, multiple dilation rates with different values ​​are introduced. The depth-separable convolution method is combined to reduce the parameter scale and maintain feature continuity. The convolution output is downsampled to generate feature maps of multiple scales. The multi-scale texture feature maps are then merged through channel stitching.

[0015] Input the multi-scale texture feature map into the texture-coordinate joint head to generate a texture saliency map and initial coordinate parameters for drilling;

[0016] Input the initial drilling coordinate parameters into the projection sensing coordinate transformer to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system;

[0017] A multi-task loss function is constructed based on borehole positioning coordinates, texture saliency map, borehole initial coordinate parameters, and mirror mask. End-to-end training is performed to update the parameters of the improved HorNet backbone network and texture-coordinate joint head.

[0018] The improved HorNet network, trained and converged by inputting fused images, combines texture saliency maps, initial borehole coordinate parameters, and a projection-aware coordinate transformer to output the final borehole positioning coordinates.

[0019] Optionally, the acquisition of multiple frames of images of the metal surface and the preprocessing to obtain the original multi-frame image sequence specifically include:

[0020] Acquire multiple frames of surface images covering the borehole area;

[0021] Workpiece pose data is collected. The workpiece pose data is obtained by the encoder and position sensor of the machine tool control system to obtain the position parameters and attitude parameters of the workpiece in the machine tool coordinate system, or by obtaining the spatial pose information of the workpiece through a three-dimensional measurement device.

[0022] Camera calibration data is collected. The camera calibration data is obtained by image acquisition and calibration algorithm of calibration board, including camera intrinsic parameter matrix and extrinsic parameter matrix.

[0023] Perform geometric distortion correction on multiple frames of surface images;

[0024] Spatial registration of multiple frames of surface images based on feature point matching;

[0025] Time synchronization of multiple surface images is performed based on a unified timestamp.

[0026] Optionally, the step of performing weighted fusion on the original multi-frame image sequence to obtain a fused image, and simultaneously generating a mirror mask and an edge map, specifically includes:

[0027] For the original multi-frame image sequence, a local window is established at the pixel level. The sliding window method is used to calculate the gray-level variance of each frame image within the window, and the gray-level variance results are converted into pixel weights through a normalization method.

[0028] The pixel value of each frame image is multiplied with the corresponding pixel weight, and accumulation and normalization operations are performed at the same pixel position. A pixel-wise weighting method based on local contrast is used to generate a fused image.

[0029] The difference between the local maximum and minimum gray values ​​is calculated using the brightness difference method for the gray channels of the fused image, and the difference result is binarized using a segmentation method with a set threshold to obtain the mirror mask of the specular reflection area.

[0030] The Sobel operator method is used to calculate the gradient components in the horizontal and vertical directions of the fused image, and the gradient magnitude is obtained by the sum of squares and square root method. Then, the gradient magnitude is used to extract edge pixels by threshold segmentation, and the edge pixels are connected to form an edge map by combining connectivity analysis method.

[0031] Optionally, the improved HorNet backbone network specifically includes:

[0032] An improved HorNet backbone framework is constructed, which includes recursively gated convolutional units, adaptive order control units, complex wavelet gated units, and large kernel dilated convolutional units.

[0033] The system receives a fused image, a mirror mask, and an edge map as input. These are then combined into multi-channel input data through channel stitching. A normalization method is used to standardize the pixel values ​​to obtain standardized input data.

[0034] The recursive gated convolutional unit performs recursive convolution on standardized input data to extract local features. Each convolution is performed by convolution operation and linear weighted summation to obtain the convolution result. After the convolution calculation, the mirror mask and edge map are introduced into the gate function as bias information. The gate function uses the Sigmoid function to perform a nonlinear transformation on the convolution result and the value after weighted bias of the mirror mask and edge map. Controlled features are generated between the convolution output and the input by element-wise multiplication, and the first feature map is output.

[0035] The adaptive step control unit uses the feature gradient analysis method to calculate the gradient change magnitude between two adjacent recursive convolution results on the first feature map and compares it with a preset threshold. When the change magnitude exceeds the threshold, it continues to iterate the convolution; when the change magnitude is lower than the threshold, it terminates the iteration and outputs the second feature map.

[0036] The complex wavelet gating unit performs complex wavelet decomposition on the second feature map to obtain multi-scale multi-directional sub-bands containing amplitude and phase information. Based on each multi-scale multi-directional sub-band, the phase consistency index is calculated, and the phase consistency index is superimposed with the input feature through a channel-by-channel weighting method to output a third feature map that simultaneously contains spatial and frequency domain information.

[0037] The large-kernel dilated convolutional unit performs convolution operations on the third feature map. During the convolution process, multiple dilation rates with different values ​​are introduced. The depth-separable convolution method is combined to reduce the parameter scale while maintaining feature continuity. The convolution output is downsampled to generate feature maps of multiple scales, and then merged through channel stitching to obtain multi-scale texture feature maps.

[0038] Optionally, the step of inputting the multi-scale texture feature map into the texture-coordinate joint head to generate a texture saliency map and initial borehole coordinate parameters specifically includes:

[0039] Construct a texture-coordinate joint head that includes feature mapping units, saliency analysis units, and coordinate regression units;

[0040] The feature mapping unit performs a channel-wise weighted summation operation on the multi-scale texture feature map to obtain preliminary convolutional features. The preliminary convolutional features are then processed by a non-linear activation method, which includes performing a modified linear unit (ReLU) operation on each pixel position and batch normalizing to obtain the first mapping features.

[0041] The saliency analysis unit uses channel attention and spatial attention methods to weight the first mapping features, calculates the texture saliency response map, extracts high saliency regions through threshold segmentation, and outputs the texture saliency map.

[0042] The coordinate regression unit receives the first mapping feature and the texture saliency map as joint input, uses a fully connected mapping method to convert the input features into a coordinate parameter vector, and uses the least squares regression method to predict the initial coordinate parameters of the borehole center point, and outputs the initial coordinate parameters of the borehole.

[0043] Optionally, the step of inputting the initial drilling coordinate parameters into the projection sensing coordinate transformer to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system specifically includes:

[0044] Construct a projection-sensing coordinate transform that includes a coordinate input unit, a projection mapping unit, and a coordinate correction unit;

[0045] The coordinate input unit uses a data synchronization method to align the initial coordinate parameters of drilling, workpiece pose data and camera calibration data under the same time reference, and forms a joint coordinate input vector through a coordinate stitching method.

[0046] The projection mapping unit receives the joint coordinate input vector and uses a perspective projection transformation method to map the initial drilling coordinate parameters from the image coordinate system to the workpiece space coordinate system. The perspective projection transformation method includes using the camera intrinsic parameter matrix to perform a mapping operation from two-dimensional pixel coordinates to normalized camera coordinates, and combining the extrinsic parameter matrix to map the camera coordinate system to the machine tool workpiece coordinate system, and outputting the initial projection coordinates.

[0047] The coordinate correction unit receives the preliminary projected coordinates, performs a spatial alignment method in conjunction with the workpiece pose data, performs rotation and translation correction on the coordinate positions, and uses a numerical optimization method to iteratively minimize the residual error, outputting standardized drilling positioning coordinates.

[0048] Optionally, the step of constructing a multi-task loss function based on borehole positioning coordinates, texture saliency map, initial borehole coordinate parameters, and specular mask, performing end-to-end training, and updating the parameters of the improved HorNet backbone network and texture-coordinate joint head specifically includes:

[0049] Construct a multi-task loss function based on borehole positioning coordinates, texture saliency map, borehole initial coordinate parameters, and mirror mask;

[0050] The multi-task loss function includes calculating the error between the borehole positioning coordinates and the initial borehole coordinate parameters using the Euclidean distance calculation method or the least squares method, calculating the matching degree between the texture saliency map and the specular mask using the cross-entropy method or the similarity measurement method, and a regularization loss function that uses the parameter norm constraint method to suppress overfitting. The weighted combination yields the joint optimization objective.

[0051] The multi-scale texture feature map and the initial coordinate parameters of the borehole are used as inputs. The forward propagation process is performed to generate the borehole positioning coordinates and texture saliency map. The borehole positioning coordinates and the initial coordinate parameters of the borehole, the texture saliency map and the mirror mask are respectively input into the multi-task loss function for error calculation.

[0052] The gradient is calculated using the backpropagation method, and the parameters of the improved HorNet backbone network and texture-coordinate joint head are updated using an iterative optimization method.

[0053] Optionally, the improved HorNet network, after training and convergence, is fed into the fused image input, and combined with the texture saliency map, initial borehole coordinate parameters, and a projection-aware coordinate transformer to output the final borehole positioning coordinates, specifically including:

[0054] The improved HorNet network, after training and convergence by fusing image inputs, is subjected to recursive gated convolution, adaptive order control, complex wavelet gate, and large kernel dilated convolution operations to output multi-scale texture feature maps.

[0055] The multi-scale texture feature map is input into the texture-coordinate joint header, and feature mapping, saliency analysis and coordinate regression operations are performed to obtain the texture saliency map and the initial coordinate parameters of the borehole.

[0056] The initial drilling coordinate parameters are aligned and stitched together with the workpiece pose data and camera calibration data in the coordinate input unit, and then input into the projection sensing coordinate transformer to perform projection mapping and coordinate correction, so as to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system.

[0057] The saliency of the borehole positioning coordinates is checked by combining the texture saliency map. When the borehole positioning coordinates fall within the saliency region, they are output as valid results. When the borehole positioning coordinates deviate from the saliency region, the correction process is triggered, and the network parameters after training convergence are called to recalculate.

[0058] The final output is the borehole positioning coordinates after significance verification and coordinate correction.

[0059] The beneficial effects of this invention are:

[0060] This invention improves the stability and detail fidelity of the original sequence by weighted fusion of multiple frames of images combined with geometric distortion correction, spatial registration, and temporal synchronization. An improved HorNet is embedded with specular masks and edge maps as gating bias information to distinguish between reflective pseudo-edges and real textures in the spatial-frequency domain and with a large receptive field, outputting robust multi-scale texture features. Texture-coordinate joint head-linked saliency analysis and coordinate regression reduce initial coordinate drift. A projection-sensing coordinate transformer, combined with camera inputs and outputs, performs perspective mapping, rotation / translation correction, and numerical optimization to reduce systematic errors in image-to-machine tool workpiece coordinates. End-to-end training is achieved using a multi-task loss consisting of coordinate error, saliency, and specular consistency, enhancing generalization and robustness under weak texture and high reflectivity conditions. Saliency verification and backoff correction are superimposed during the inference stage to output more reliable drilling positioning coordinates, seamlessly integrating with CNC machining, improving positioning accuracy, repeatability, and cycle time efficiency while reducing reliance on manual alignment. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of the metal surface drilling positioning method based on artificial intelligence proposed in this invention;

[0063] Figure 2 This is a schematic diagram of the metal surface drilling positioning method based on artificial intelligence proposed in this invention;

[0064] Figure 3 This is a framework diagram of the improved HorNet model in the artificial intelligence-based metal surface drilling positioning method proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1-3 The method for locating drill holes on metal surfaces based on artificial intelligence includes the following steps:

[0067] Multiple frames of images of the metal surface were acquired and preprocessed to obtain the original multi-frame image sequence;

[0068] The original multi-frame image sequence is weighted and fused to obtain a fused image, and a mirror mask and edge map are generated at the same time.

[0069] An improved HorNet backbone framework is constructed, which includes recursively gated convolutional units, adaptive order control units, complex wavelet gated units, and large kernel dilated convolutional units.

[0070] The system receives a fused image, a mirror mask, and an edge map as input. These are then combined into multi-channel input data through channel stitching. A normalization method is used to standardize the pixel values ​​to obtain standardized input data.

[0071] The recursive gated convolutional unit performs recursive convolution on standardized input data to extract local features. Each convolution is performed by convolution operation and linear weighted summation to obtain the convolution result. After the convolution calculation, the mirror mask and edge map are introduced into the gate function as bias information. The gate function uses the Sigmoid function to perform a nonlinear transformation on the convolution result and the value after weighted bias of the mirror mask and edge map. Controlled features are generated between the convolution output and the input by element-wise multiplication, and the first feature map is output.

[0072] The adaptive step control unit uses the feature gradient analysis method to calculate the gradient change magnitude between two adjacent recursive convolution results on the first feature map and compares it with a preset threshold. When the change magnitude exceeds the threshold, it continues to iterate the convolution; when the change magnitude is lower than the threshold, it terminates the iteration and outputs the second feature map.

[0073] The complex wavelet gating unit performs complex wavelet decomposition on the second feature map to obtain multi-scale multi-directional sub-bands containing amplitude and phase information. Based on each multi-scale multi-directional sub-band, the phase consistency index is calculated, and the phase consistency index is superimposed with the input feature through a channel-by-channel weighting method to output a third feature map that simultaneously contains spatial and frequency domain information.

[0074] The large kernel dilated convolution unit performs convolution operation on the third feature map. During the convolution process, multiple dilation rates with different values ​​are introduced. The depth-separable convolution method is combined to reduce the parameter scale and maintain feature continuity. The convolution output is downsampled to generate feature maps of multiple scales. The multi-scale texture feature maps are then merged through channel stitching.

[0075] Input the multi-scale texture feature map into the texture-coordinate joint head to generate a texture saliency map and initial coordinate parameters for drilling;

[0076] Input the initial drilling coordinate parameters into the projection sensing coordinate transformer to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system;

[0077] A multi-task loss function is constructed based on borehole positioning coordinates, texture saliency map, borehole initial coordinate parameters, and mirror mask. End-to-end training is performed to update the parameters of the improved HorNet backbone network and texture-coordinate joint head.

[0078] The improved HorNet network, trained and converged by inputting fused images, combines texture saliency maps, initial borehole coordinate parameters, and a projection-aware coordinate transformer to output the final borehole positioning coordinates.

[0079] In this embodiment, the acquisition of multiple frames of images of the metal surface and the preprocessing to obtain the original multi-frame image sequence specifically include:

[0080] Acquire multiple frames of surface images covering the borehole area;

[0081] Workpiece pose data is collected. The workpiece pose data is obtained by the encoder and position sensor of the machine tool control system to obtain the position parameters and attitude parameters of the workpiece in the machine tool coordinate system, or by obtaining the spatial pose information of the workpiece through a three-dimensional measurement device.

[0082] Camera calibration data is collected. The camera calibration data is obtained by image acquisition and calibration algorithm of calibration board, including camera intrinsic parameter matrix and extrinsic parameter matrix.

[0083] Perform geometric distortion correction on multiple frames of surface images;

[0084] Spatial registration of multiple frames of surface images based on feature point matching;

[0085] Time synchronization of multiple surface images is performed based on a unified timestamp.

[0086] This implementation method acquires multiple frames of surface images covering the drilling area, and combines them with workpiece pose data obtained from machine tool encoders and position sensors, as well as camera intrinsic and extrinsic parameter data obtained from calibration board experiments. It employs processing methods such as geometric distortion correction, feature point matching and registration, and time synchronization to obtain an original multi-frame image sequence with unified spatial coordinates and time reference. This effectively eliminates lens distortion and acquisition errors, ensures the consistency between image information and workpiece coordinates, and improves the reliability and accuracy of subsequent fusion and feature extraction.

[0087] In this embodiment, the step of performing weighted fusion on the original multi-frame image sequence to obtain a fused image, and simultaneously generating a mirror mask and an edge map, specifically includes:

[0088] For the original multi-frame image sequence, a local window is established at the pixel level. The sliding window method is used to calculate the gray-level variance of each frame image within the window, and the gray-level variance results are converted into pixel weights through a normalization method.

[0089] The pixel value of each frame image is multiplied with the corresponding pixel weight, and accumulation and normalization operations are performed at the same pixel position. A pixel-wise weighting method based on local contrast is used to generate a fused image.

[0090] The difference between the local maximum and minimum gray values ​​is calculated using the brightness difference method for the gray channels of the fused image, and the difference result is binarized using a segmentation method with a set threshold to obtain the mirror mask of the specular reflection area.

[0091] The Sobel operator method is used to calculate the gradient components in the horizontal and vertical directions of the fused image, and the gradient magnitude is obtained by the sum of squares and square root method. Then, the gradient magnitude is used to extract edge pixels by threshold segmentation, and the edge pixels are connected to form an edge map by combining connectivity analysis method.

[0092] This implementation generates a fused image by establishing local windows at the pixel level and calculating grayscale variance in the original multi-frame image sequence. At the same time, a specular mask is obtained using the brightness difference thresholding method, and the edge map is extracted by combining the Sobel operator with connectivity analysis. This enhances surface detail features while suppressing random noise, effectively distinguishes reflective areas from real texture areas, ensures that weak textures are not lost and strong reflections are not interfered with during subsequent feature extraction, and improves the stability and accuracy of overall borehole positioning.

[0093] In this embodiment, the improved HorNet backbone network specifically includes:

[0094] An improved HorNet backbone framework is constructed, which includes recursively gated convolutional units, adaptive order control units, complex wavelet gated units, and large kernel dilated convolutional units.

[0095] The system receives a fused image, a mirror mask, and an edge map as input. These are then combined into multi-channel input data through channel stitching. A normalization method is used to standardize the pixel values ​​to obtain standardized input data.

[0096] The recursive gated convolutional unit performs recursive convolution on standardized input data to extract local features. Each convolution is performed by convolution operation and linear weighted summation to obtain the convolution result. After the convolution calculation, the mirror mask and edge map are introduced into the gate function as bias information. The gate function uses the Sigmoid function to perform a nonlinear transformation on the convolution result and the value after weighted bias of the mirror mask and edge map. Controlled features are generated between the convolution output and the input by element-wise multiplication, and the first feature map is output.

[0097] The adaptive step control unit uses the feature gradient analysis method to calculate the gradient change magnitude between two adjacent recursive convolution results on the first feature map and compares it with a preset threshold. When the change magnitude exceeds the threshold, it continues to iterate the convolution; when the change magnitude is lower than the threshold, it terminates the iteration, thereby dynamically determining the number of convolution iterations and outputting the second feature map.

[0098] The complex wavelet gating unit performs complex wavelet decomposition on the second feature map to obtain multi-scale, multi-directional sub-bands containing amplitude and phase information. Based on each multi-scale, multi-directional sub-band, a phase consistency index is calculated, and the phase consistency index is superimposed on the input features through a channel-by-channel weighting method to output a third feature map that simultaneously contains spatial and frequency domain information. The complex wavelet decomposition includes performing convolution operations on the second feature map using real part filters and imaginary part filters respectively to obtain real part coefficients and imaginary part coefficients, combining them into complex coefficients to obtain complex sub-bands, performing downsampling operations on the complex sub-bands to achieve scale decomposition, and repeating the convolution and downsampling process in multiple directions to form a multi-scale, multi-directional complex wavelet decomposition result containing amplitude and phase information.

[0099] The large-kernel dilated convolutional unit performs convolution operations on the third feature map. During the convolution process, multiple dilation rates with different values ​​are introduced. The depth-separable convolution method is combined to reduce the parameter scale while maintaining feature continuity. The convolution output is downsampled to generate feature maps of multiple scales, and then merged through channel stitching to obtain multi-scale texture feature maps.

[0100] This implementation constructs an improved HorNet backbone network that includes recursive gated convolution, adaptive order control, complex wavelet gating, and large-kernel dilated convolution. Specular masking and edge map bias information are introduced into the gating function, enabling the network to suppress invalid responses in specular reflective regions and enhance detail representation in weakly textured regions. Simultaneously, phase consistency and multi-scale dilated convolution ensure global continuity, thereby achieving stable modeling of complex textures on metal surfaces and significantly improving the robustness and accuracy of borehole positioning.

[0101] In this embodiment, the step of inputting the multi-scale texture feature map into the texture-coordinate joint head to generate the texture saliency map and the initial coordinate parameters of the borehole specifically includes:

[0102] Construct a texture-coordinate joint head that includes feature mapping units, saliency analysis units, and coordinate regression units;

[0103] The feature mapping unit performs a channel-wise weighted summation operation on the multi-scale texture feature map to obtain preliminary convolutional features. The preliminary convolutional features are then processed by a nonlinear activation method, which includes performing a modified linear unit ReLU function operation on each pixel position to enhance nonlinear expressive power. Batch normalization is then performed to obtain the first mapping features.

[0104] The saliency analysis unit uses channel attention and spatial attention methods to weight the first mapping features, calculates the texture saliency response map, extracts high saliency regions through threshold segmentation, and outputs the texture saliency map.

[0105] The coordinate regression unit receives the first mapping feature and the texture saliency map as joint input, uses a fully connected mapping method to convert the input features into a coordinate parameter vector, and uses the least squares regression method to predict the initial coordinate parameters of the borehole center point, and outputs the initial coordinate parameters of the borehole.

[0106] This implementation inputs a multi-scale texture feature map into a texture-coordinate joint head, sequentially performs feature mapping, saliency analysis, and coordinate regression, and generates a texture saliency map and initial borehole coordinate parameters. This highlights the features of salient regions and establishes a correspondence with the borehole center position, thereby accurately identifying the target area and outputting the initial coordinate results even under complex texture and reflective interference conditions. This provides reliable input for subsequent coordinate transformation and precise positioning, improving the stability and accuracy of borehole positioning.

[0107] In this embodiment, the step of inputting the initial drilling coordinate parameters into the projection sensing coordinate transformer to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system specifically includes:

[0108] Construct a projection-sensing coordinate transform that includes a coordinate input unit, a projection mapping unit, and a coordinate correction unit;

[0109] The coordinate input unit uses a data synchronization method to align the initial coordinate parameters of drilling, workpiece pose data and camera calibration data under the same time reference, and forms a joint coordinate input vector through a coordinate stitching method.

[0110] The projection mapping unit receives the joint coordinate input vector and uses a perspective projection transformation method to map the initial drilling coordinate parameters from the image coordinate system to the workpiece space coordinate system. The perspective projection transformation method includes using the camera intrinsic parameter matrix to perform a mapping operation from two-dimensional pixel coordinates to normalized camera coordinates, and combining the extrinsic parameter matrix to map the camera coordinate system to the machine tool workpiece coordinate system, and outputting the initial projection coordinates.

[0111] The coordinate correction unit receives the preliminary projected coordinates, performs a spatial alignment method in conjunction with the workpiece pose data, performs rotation and translation correction on the coordinate positions, and uses a numerical optimization method to iteratively minimize the residual error, outputting standardized drilling positioning coordinates.

[0112] This implementation aligns the initial drilling coordinate parameters with the workpiece pose data and camera calibration data under a unified time reference, and performs perspective projection, rotation, and translation correction in the projection sensing coordinate transformer. Combined with numerical optimization iteration to reduce residual errors, it outputs standardized drilling positioning coordinates, thereby achieving accurate mapping from image coordinates to the machine tool workpiece coordinate system. This effectively eliminates clamping deviations and imaging errors, ensuring that the drilling positioning results are highly consistent with the actual machining coordinates, and significantly improving machining accuracy and reliability.

[0113] In this embodiment, the step of constructing a multi-task loss function based on borehole positioning coordinates, texture saliency map, initial borehole coordinate parameters, and specular mask, performing end-to-end training, and updating the parameters of the improved HorNet backbone network and texture-coordinate joint head specifically includes:

[0114] Construct a multi-task loss function based on borehole positioning coordinates, texture saliency map, borehole initial coordinate parameters, and mirror mask;

[0115] The multi-task loss function includes calculating the error between the borehole positioning coordinates and the initial borehole coordinate parameters using the Euclidean distance calculation method or the least squares method, calculating the matching degree between the texture saliency map and the specular mask using the cross-entropy method or the similarity measurement method, and a regularization loss function that uses the parameter norm constraint method to suppress overfitting. The weighted combination yields the joint optimization objective.

[0116] The multi-scale texture feature map and the initial coordinate parameters of the borehole are used as inputs. The forward propagation process is performed to generate the borehole positioning coordinates and texture saliency map. The borehole positioning coordinates and the initial coordinate parameters of the borehole, the texture saliency map and the mirror mask are respectively input into the multi-task loss function for error calculation.

[0117] The gradient is calculated using the backpropagation method, and the parameters of the improved HorNet backbone network and texture-coordinate joint head are updated using an iterative optimization method.

[0118] This implementation constructs a multi-task loss function that includes coordinate regression loss, saliency constraint loss, and regularization loss to jointly optimize borehole positioning coordinates, texture saliency maps, initial borehole coordinate parameters, and specular masks. It employs end-to-end training and backpropagation to iteratively update the parameters of the improved HorNet backbone network and the texture-coordinate joint head. This simultaneously improves coordinate prediction accuracy and saliency consistency, enhances the model's generalization and robustness, and effectively ensures the stability and reliability of borehole positioning.

[0119] In this embodiment, the step of inputting the fused image into the trained and converged improved HorNet network, combining the texture saliency map, the initial borehole coordinate parameters, and the projection-aware coordinate transformer to output the final borehole positioning coordinates specifically includes:

[0120] The improved HorNet network, after training and convergence by fusing image inputs, is subjected to recursive gated convolution, adaptive order control, complex wavelet gate, and large kernel dilated convolution operations to output multi-scale texture feature maps.

[0121] The multi-scale texture feature map is input into the texture-coordinate joint header, and feature mapping, saliency analysis and coordinate regression operations are performed to obtain the texture saliency map and the initial coordinate parameters of the borehole.

[0122] The initial drilling coordinate parameters are aligned and stitched together with the workpiece pose data and camera calibration data in the coordinate input unit, and then input into the projection sensing coordinate transformer to perform projection mapping and coordinate correction, so as to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system.

[0123] The saliency of the borehole positioning coordinates is checked by combining the texture saliency map. When the borehole positioning coordinates fall within the saliency region, they are output as valid results. When the borehole positioning coordinates deviate from the saliency region, the correction process is triggered, and the network parameters after training convergence are called to recalculate.

[0124] The final output is the borehole positioning coordinates after significance verification and coordinate correction.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to an automated drilling and positioning process for a metal casing part of an aero-engine. This type of part is made of high-strength aluminum alloy, and after sandblasting and polishing, it exhibits a significant specular reflection effect, while some areas show almost no clear texture features. Traditional drilling positioning methods based on single-frame images often result in positioning deviations under these conditions, causing the difference between the drilling coordinates and the designed position to exceed 0.5 mm, failing to meet the ±0.25 mm tolerance standard required for aero-engine assembly.

[0127] In this embodiment, the experimental system includes a five-axis CNC machine tool, an industrial camera with a resolution of 2448×2048 pixels, an industrial lens with a focal length of 25 mm, and a high-precision calibration plate that works in conjunction with the machine tool's motion. The industrial camera is fixed near the machine tool spindle and is synchronously triggered with the camera acquisition system through the machine tool control system, thereby acquiring multiple frames of surface images covering the target area before the drilling operation. In the experiment, 15 frames of surface images are acquired for each drilling area at a frame rate of 30fps. Simultaneously, the workpiece's pose data is acquired through the machine tool's built-in encoder and position sensor. The camera's intrinsic and extrinsic parameters are obtained experimentally through the calibration plate to ensure the accurate transformation relationship between the camera coordinate system and the machine tool / workpiece coordinate system.

[0128] The acquired multi-frame surface images first undergo geometric distortion correction to eliminate barrel or pincushion distortion introduced by the lens. Then, a feature point matching method is used to spatially register the image sequence, ensuring that each frame is perfectly aligned in space. Since the camera and machine tool control system are time-synchronized, the image sequence can be temporally aligned using a unified timestamp. After these preprocessing steps, a raw multi-frame image sequence with a unified spatial and temporal reference is obtained.

[0129] The original multi-frame image sequence is fused into a fused image using a pixel-by-pixel weighted fusion method. Specifically, a local window is established at each pixel location, and the gray-level variance of each frame image within that window is calculated. A larger variance indicates that the region contains more texture information. The variance is converted into weights through normalization, obtaining the weight value of each frame image at that pixel location. The pixel value of each frame is then multiplied by its corresponding weight, and this multiplication is performed across all frames, finally normalized to obtain the fused image. Simultaneously, a brightness difference method is applied to the gray-level channels of the fused image to calculate the difference between local maximum and minimum values. After setting a threshold, the difference results are binarized to obtain a mirror mask for the mirror region. The Sobel operator is then used to calculate the horizontal and vertical gradient components of the fused image. The square root of the sum of squares yields the gradient magnitude. Combined with threshold segmentation and connectivity analysis, an edge map reflecting the boundary information of the workpiece surface is formed. The fused image, mirror mask, and edges are then analyzed. Figure 1 And serve as subsequent network input.

[0130] The improved HorNet backbone network in this embodiment consists of four units: recursive gated convolution, adaptive order control, complex wavelet gating, and large-kernel dilated convolution. The recursive gated convolution unit receives a concatenated input of the fused image, a specular mask, and an edge map. It extracts local features through convolution operations and incorporates bias information from the specular mask and edge map into the gating function, suppressing ineffective reflection signals in specular regions and enhancing useful features in weakly textured regions. The adaptive order control unit dynamically judges the gradient change magnitude of the feature map in the recursive convolution. When the change is large, it increases the number of convolution iterations to extract richer details; when the change is small, it terminates the iteration early to reduce computation. The complex wavelet gating unit performs complex wavelet decomposition on the feature map, generating multi-scale, multi-directional subbands containing amplitude and phase information. The multi-scale results are superimposed on the original features using a phase consistency index, thereby enhancing sensitivity to directional and periodic textures. Large-kernel dilated convolutional units expand the receptive field while maintaining feature continuity, reduce parameter scale through depthwise separable convolution, and ultimately output multi-scale texture feature maps.

[0131] The multi-scale texture feature map is input to the texture-coordinate joint header. The feature mapping unit extracts compact feature representations using convolution and nonlinear activation functions, and enhances numerical stability through batch normalization. The saliency analysis unit combines channel attention and spatial attention mechanisms to generate a texture saliency map, highlighting the salient features of the borehole region. The coordinate regression unit receives the texture saliency map and feature mapping results, and outputs the initial coordinate parameters of the borehole using a fully connected layer and least squares regression.

[0132] The initial drilling coordinate parameters are further input into the projection sensing coordinate transformer. This module combines workpiece pose data and camera calibration data, first using an intrinsic parameter matrix to map the initial drilling coordinate parameters from image pixel coordinates to normalized camera coordinates, and then using an extrinsic parameter matrix to transform them to the machine tool workpiece coordinate system. Subsequently, rotation and translation corrections are performed, and residual errors are reduced through iterative numerical optimization methods, finally outputting standardized drilling positioning coordinates.

[0133] During the training phase, this invention constructs a multi-task loss function based on borehole positioning coordinates, texture saliency maps, initial borehole coordinate parameters, and a specular mask. This function consists of three parts: first, a coordinate regression loss, which calculates the difference between the borehole positioning coordinates and the initial borehole coordinate parameters using Euclidean distance or least squares methods; second, a saliency constraint loss, which calculates the consistency between the texture saliency map and the specular mask using cross-entropy or similarity metrics; and third, a regularization loss, which uses parameter norm constraints to limit network complexity and avoid overfitting. These three loss components are weighted and combined to form a joint optimization objective. Through end-to-end training, borehole positioning coordinates and texture saliency maps are generated using forward propagation, and then the parameters of the improved HorNet backbone network and the texture-coordinate joint head are updated through backpropagation and iterative optimization. After several rounds of training iterations, the network reaches convergence and can stably output high-precision results.

[0134] During the inference phase, the improved HorNet network, trained and converged from the fused image input, yields multi-scale texture feature maps. These maps are then processed by a texture-coordinate joint head to generate a texture saliency map and initial drilling coordinate parameters. Combining workpiece pose data and camera calibration data, the drilling positioning coordinates are output via a projection-sensory coordinate transformer. A saliency check is then performed to determine if the drilling positioning coordinates are located in a highly saliency region of the texture saliency map. If not, a correction process is triggered for recalculation, ensuring the reliability of the final result. The final output is the drilling positioning coordinates after saliency check and coordinate correction.

[0135] To compare the results, this embodiment selected 10 drilling points on the workpiece for the experiment. The traditional method based on a single frame image and the method of the present invention were used for positioning, and the errors in the X and Y directions were recorded.

[0136] Table 1. Comparison of Drilling Positioning Errors (Unit: mm)

[0137] Drill hole number Traditional method error X The error Y of the traditional method Error X of the method of this invention Error Y of the method of this invention 1 0.42 0.37 0.12 0.1 2 0.55 0.41 0.18 0.14 3 0.61 0.52 0.2 0.15 4 0.44 0.39 0.11 0.09 5 0.58 0.47 0.19 0.12 6 0.63 0.56 0.21 0.17 7 0.47 0.42 0.14 0.11 8 0.53 0.48 0.17 0.13 9 0.46 0.4 0.13 0.1 10 0.6 0.49 0.22 0.15

[0138] Analysis of the data in Table 1 reveals that the traditional method has an average positioning error of 0.53 mm in the X direction and 0.45 mm in the Y direction, with a maximum error of 0.63 mm. In contrast, the method of this invention has an average positioning error of only 0.17 mm in the X direction and 0.13 mm in the Y direction, with a maximum error not exceeding 0.22 mm. Compared to the traditional method, the method of this invention reduces the average error by approximately 68% and the maximum error by approximately 65% ​​across 10 drilling points. Furthermore, the standard deviation of the error of the method of this invention is only 0.04 mm, while that of the traditional method reaches 0.08 mm, indicating that this invention also has a significant advantage in the stability of the output results.

[0139] Regarding processing efficiency, the adaptive order control mechanism introduced in this invention can automatically reduce unnecessary convolution iterations. In experiments, traditional deep networks take an average of 0.85 seconds to process an image of a drilling area, while the method of this invention reduces this to 0.52 seconds, improving processing efficiency by approximately 39%. In actual machine tool processing, after adopting the method of this invention, all drill holes meet the tolerance requirement of ±0.25 mm, while the traditional method has 3 holes that are out of tolerance and require rework. Therefore, this invention reduces the rework rate while ensuring high precision, thus improving overall production stability.

[0140] Comprehensive experimental results demonstrate that this invention, through the introduction of fused image generation, specular mask and edge map constraints, multi-unit collaborative feature extraction of the improved HorNet backbone network, a dual-output mechanism of the texture-coordinate joint head, geometric correction of the projection-aware coordinate transformer, and end-to-end training of the multi-task loss function, not only effectively suppresses specular reflection interference but also enhances the feature representation capability of weakly textured regions. Compared with traditional methods, this invention improves drilling positioning accuracy by approximately 68%, processing efficiency by approximately 39%, and significantly reduces rework rate. This fully demonstrates that this invention has broad application prospects and significant industrial value in scenarios with extremely high drilling accuracy requirements, such as aero-engine housings, automotive parts, and precision molds.

[0141] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for locating drill holes on metal surfaces based on artificial intelligence, characterized in that, Includes the following steps: Multiple frames of images of the metal surface were acquired and preprocessed to obtain the original multi-frame image sequence; The original multi-frame image sequence is weighted and fused to obtain a fused image, and a mirror mask and edge map are generated at the same time. An improved HorNet backbone network was constructed, comprising recursively gated convolutional units, adaptive order control units, complex wavelet gated units, and large-kernel dilated convolutional units. The system receives a fused image, a mirror mask, and an edge map as input. These are then combined into multi-channel input data through channel stitching. A normalization method is used to standardize the pixel values ​​to obtain standardized input data. The recursive gated convolutional unit performs recursive convolution on standardized input data to extract local features. Each convolution is performed by convolution operation and linear weighted summation to obtain the convolution result. After the convolution calculation, the mirror mask and edge map are introduced into the gate function as bias information. The gate function uses the Sigmoid function to perform a nonlinear transformation on the convolution result and the value after weighted bias of the mirror mask and edge map. Controlled features are generated between the convolution output and the input by element-wise multiplication, and the first feature map is output. The adaptive step control unit uses the feature gradient analysis method to calculate the gradient change magnitude between two adjacent recursive convolution results on the first feature map and compares it with a preset threshold. When the change magnitude exceeds the threshold, it continues to iterate the convolution; when the change magnitude is lower than the threshold, it terminates the iteration and outputs the second feature map. The complex wavelet gating unit performs complex wavelet decomposition on the second feature map to obtain multi-scale multi-directional sub-bands containing amplitude and phase information. Based on each multi-scale multi-directional sub-band, the phase consistency index is calculated, and the phase consistency index is superimposed with the input feature through a channel-by-channel weighting method to output a third feature map that simultaneously contains spatial and frequency domain information. The large kernel dilated convolution unit performs convolution operation on the third feature map. During the convolution process, multiple dilation rates with different values ​​are introduced. The depth-separable convolution method is combined to reduce the parameter scale and maintain feature continuity. The convolution output is downsampled to generate feature maps of multiple scales. The multi-scale texture feature maps are then merged through channel stitching. Input the multi-scale texture feature map into the texture-coordinate joint head to generate a texture saliency map and initial coordinate parameters for drilling; Input the initial drilling coordinate parameters into the projection sensing coordinate transformer to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system; A multi-task loss function is constructed based on borehole positioning coordinates, texture saliency map, borehole initial coordinate parameters, and mirror mask. End-to-end training is performed to update the parameters of the improved HorNet backbone network and texture-coordinate joint head. The improved HorNet backbone network, trained and converged by inputting fused images, is combined with texture saliency maps, initial borehole coordinate parameters, and a projection-aware coordinate transformer to output the final borehole positioning coordinates.

2. The method for positioning drill holes on metal surfaces based on artificial intelligence according to claim 1, characterized in that, The process of acquiring and preprocessing multiple frames of images of the metal surface to obtain the original multi-frame image sequence specifically includes: Acquire multiple frames of surface images covering the borehole area; Workpiece pose data is collected. The workpiece pose data is obtained by the encoder and position sensor of the machine tool control system to obtain the position parameters and attitude parameters of the workpiece in the machine tool coordinate system, or by obtaining the spatial pose information of the workpiece through a three-dimensional measurement device. Camera calibration data is collected. The camera calibration data is obtained by image acquisition and calibration algorithm of calibration board, including camera intrinsic parameter matrix and extrinsic parameter matrix. Perform geometric distortion correction on multiple frames of surface images; Spatial registration of multiple frames of surface images based on feature point matching; Time synchronization of multiple surface images is performed based on a unified timestamp.

3. The method for positioning drill holes on metal surfaces based on artificial intelligence according to claim 1, characterized in that, The step of performing weighted fusion on the original multi-frame image sequence to obtain a fused image, and simultaneously generating a mirror mask and an edge map, specifically includes: For the original multi-frame image sequence, a local window is established at the pixel level. The sliding window method is used to calculate the gray-level variance of each frame image within the window, and the gray-level variance results are converted into pixel weights through a normalization method. The pixel value of each frame image is multiplied with the corresponding pixel weight, and accumulation and normalization operations are performed at the same pixel position. A pixel-wise weighting method based on local contrast is used to generate a fused image. The difference between the local maximum and minimum gray values ​​is calculated using the brightness difference method for the gray channels of the fused image, and the difference result is binarized using a segmentation method with a set threshold to obtain the mirror mask of the specular reflection area. The Sobel operator method is used to calculate the gradient components in the horizontal and vertical directions of the fused image, and the gradient magnitude is obtained by the sum of squares and square root method. Then, the gradient magnitude is used to extract edge pixels by threshold segmentation, and the edge pixels are connected to form an edge map by combining connectivity analysis method.

4. The method for positioning drill holes on metal surfaces based on artificial intelligence according to claim 1, characterized in that, The step of inputting the multi-scale texture feature map into the texture-coordinate joint head to generate a texture saliency map and initial borehole coordinate parameters specifically includes: Construct a texture-coordinate joint head that includes feature mapping units, saliency analysis units, and coordinate regression units; The feature mapping unit performs a channel-wise weighted summation operation on the multi-scale texture feature map to obtain preliminary convolutional features. The preliminary convolutional features are then processed by a non-linear activation method, which includes performing a modified linear unit (ReLU) operation on each pixel position and batch normalizing to obtain the first mapping features. The saliency analysis unit uses channel attention and spatial attention methods to weight the first mapping features, calculates the texture saliency response map, extracts high saliency regions through threshold segmentation, and outputs the texture saliency map. The coordinate regression unit receives the first mapping feature and the texture saliency map as joint input, uses a fully connected mapping method to convert the input features into a coordinate parameter vector, and uses the least squares regression method to predict the initial coordinate parameters of the borehole center point, and outputs the initial coordinate parameters of the borehole.

5. The method for drilling and positioning holes on a metal surface based on artificial intelligence according to claim 1, characterized in that, The step of inputting the initial drilling coordinate parameters into the projection sensing coordinate transformer to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system specifically includes: Construct a projection-sensing coordinate transform that includes a coordinate input unit, a projection mapping unit, and a coordinate correction unit; The coordinate input unit uses a data synchronization method to align the initial coordinate parameters of drilling, workpiece pose data and camera calibration data under the same time reference, and forms a joint coordinate input vector through a coordinate stitching method. The projection mapping unit receives the joint coordinate input vector and uses a perspective projection transformation method to map the initial drilling coordinate parameters from the image coordinate system to the workpiece space coordinate system. The perspective projection transformation method includes using the camera intrinsic parameter matrix to perform a mapping operation from two-dimensional pixel coordinates to normalized camera coordinates, and combining the extrinsic parameter matrix to map the camera coordinate system to the machine tool workpiece coordinate system, and outputting the initial projection coordinates. The coordinate correction unit receives the preliminary projected coordinates, performs a spatial alignment method in conjunction with the workpiece pose data, performs rotation and translation correction on the coordinate positions, and uses a numerical optimization method to iteratively minimize the residual error, outputting standardized drilling positioning coordinates.

6. The method for drilling and positioning holes on a metal surface based on artificial intelligence according to claim 1, characterized in that, The process involves constructing a multi-task loss function based on borehole positioning coordinates, texture saliency maps, initial borehole coordinate parameters, and a specular mask, performing end-to-end training, and updating the parameters of the improved HorNet backbone network and the texture-coordinate joint head. Specifically, this includes: Construct a multi-task loss function based on borehole positioning coordinates, texture saliency map, borehole initial coordinate parameters, and mirror mask; The multi-task loss function includes calculating the error between the borehole positioning coordinates and the initial borehole coordinate parameters using the Euclidean distance calculation method or the least squares method, calculating the matching degree between the texture saliency map and the specular mask using the cross-entropy method or the similarity measurement method, and a regularization loss function that uses the parameter norm constraint method to suppress overfitting. The weighted combination yields the joint optimization objective. The multi-scale texture feature map and the initial coordinate parameters of the borehole are used as inputs. The forward propagation process is performed to generate the borehole positioning coordinates and texture saliency map. The borehole positioning coordinates and the initial coordinate parameters of the borehole, the texture saliency map and the mirror mask are respectively input into the multi-task loss function for error calculation. The gradient is calculated using the backpropagation method, and the parameters of the improved HorNet backbone network and texture-coordinate joint head are updated using an iterative optimization method.

7. The method for positioning drill holes on metal surfaces based on artificial intelligence according to claim 1, characterized in that, The improved HorNet backbone network, after training and convergence, is input with the fused image. Combining the texture saliency map, initial borehole coordinate parameters, and a projection-aware coordinate transformer, it outputs the final borehole location coordinates. Specifically, this includes: The improved HorNet backbone network, after training and convergence by fusing image inputs, is subjected to recursive gated convolution, adaptive order control, complex wavelet gate, and large kernel dilated convolution operations to output multi-scale texture feature maps. The multi-scale texture feature map is input into the texture-coordinate joint header, and feature mapping, saliency analysis and coordinate regression operations are performed to obtain the texture saliency map and the initial coordinate parameters of the borehole. The initial drilling coordinate parameters are aligned and stitched together with the workpiece pose data and camera calibration data in the coordinate input unit, and then input into the projection sensing coordinate transformer to perform projection mapping and coordinate correction, so as to obtain the drilling positioning coordinates in the machine tool workpiece coordinate system. The saliency of the borehole positioning coordinates is checked by combining the texture saliency map. When the borehole positioning coordinates fall within the saliency region, they are output as valid results. When the borehole positioning coordinates deviate from the saliency region, the correction process is triggered, and the network parameters after training convergence are called to recalculate. The final output is the borehole positioning coordinates after significance verification and coordinate correction.

Images