Intelligent planning method and system for cutting path of wire cutting machine based on image processing

Abstract

The invention relates to the technical field of image processing, and discloses a wire cutting machine cutting path intelligent planning method and system based on image processing. The method comprises the following steps: performing multi-angle image acquisition and preprocessing on a to-be-processed workpiece to obtain a workpiece feature map; inputting the workpiece feature map into a sensing area identification network for processing to obtain a target processing area identification map; performing feature enhancement on the target processing area identification image to obtain an enhanced feature image and an importance weight image; cutting path planning is carried out according to the enhanced feature map and the importance weight map, and an initial cutting path is obtained; and performing multi-objective optimization on the initial cutting path to obtain a target cutting path and a corresponding electric spark parameter configuration table. According to the method, collaborative optimization of the cutting path and the electric spark parameters is achieved, the contradiction among the machining precision, the machining time and the energy consumption is balanced, and the system can achieve the optimal process matching on workpieces made of different materials and with different geometric complexities.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an image processing-based wire cutting machine cutting path intelligent planning method and system. Background Art

[0002] Traditional wire cutting path planning algorithms are no longer able to meet modern manufacturing needs. Traditional algorithms mainly rely on simple edge detection technology and cannot effectively handle the problem of feature extraction of complex workpieces, especially on multi-material structures and variable cross-section workpieces. It is difficult to accurately identify the main and secondary features on the workpiece, resulting in low cutting accuracy.

[0003] In the actual processing process, the traditional wire cutting path planning method fails to fully consider the special geometric features of the workpiece, such as the processing requirements of the target areas such as sharp angles, fine holes, and slits, which makes it easy for the electrode wire to cause vibration and unstable electric spark discharge when cutting in these areas. At the same time, the traditional method has poor adaptability to different material properties and workpiece geometries, and cannot perform differentiated processing, resulting in uneven electrode wire loss, high energy consumption during the processing process, unstable surface roughness, and other problems, which seriously affect the processing quality and production efficiency. Summary of the invention

[0004] The present invention provides a method and system for intelligent planning of cutting paths for wire cutting machines based on image processing. The present invention realizes the coordinated optimization of cutting paths and electric spark parameters, balances the contradictions among machining accuracy, machining time and energy consumption, and enables the system to achieve optimal process matching on workpieces of different materials and geometric complexity.

[0005] In a first aspect, the present invention provides a method for intelligently planning a cutting path of a wire cutting machine based on image processing, the method comprising: Perform multi-angle image acquisition and preprocessing on the workpiece to be processed to obtain the workpiece feature map; Inputting the workpiece feature map into a perception area recognition network for processing to obtain a target processing area recognition map; Performing feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map; Performing cutting path planning according to the enhanced feature map and the importance weight map to obtain an initial cutting path; The initial cutting path is subjected to multi-objective optimization to obtain a target cutting path and a corresponding electric spark parameter configuration table.

[0006] In a second aspect, the present invention provides an intelligent planning system for cutting paths of wire cutting machines based on image processing, and the intelligent planning system for cutting paths of wire cutting machines based on image processing comprises: An acquisition module is used to acquire and preprocess multi-angle images of the workpiece to be processed to obtain a workpiece feature map; A region recognition module, used for inputting the workpiece feature map into a perception region recognition network for processing to obtain a target processing region recognition map; A feature enhancement module, used to perform feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map; A path planning module, used for performing cutting path planning according to the enhanced feature map and the importance weight map to obtain an initial cutting path; The multi-objective optimization module is used to perform multi-objective optimization on the initial cutting path to obtain a target cutting path and a corresponding electric spark parameter configuration table.

[0007] In the technical solution provided by the present invention, through the multi-angle acquisition and adaptive preprocessing technology of the industrial camera, the geometric feature information of the workpiece can be fully captured, effectively overcoming the information loss problem caused by the traditional single-view acquisition, and providing a high-quality workpiece feature map for subsequent accurate identification. The dual-stream feature extraction architecture of the perception area recognition network is adopted to realize the coordinated extraction of macro-geometric features and micro-texture details, solve the problem of information transmission loss in the traditional method when processing sharp angles and fine features, and accurately identify the target area that needs differentiated processing. Through the supplementation and fusion mechanism of global and local features, the system can pay special attention to small feature areas such as sharp corners and fine holes, accurately depict the complex contour features of the workpiece, and solve the problem of insufficient recognition of small features by traditional methods. Intelligent path planning based on enhanced feature maps and importance weight maps can implement differentiated control point density settings and feed speed adjustments for different geometric feature areas, ensure the stability and efficiency of the cutting process, and significantly improve the quality of complex contour processing. By establishing a multi-objective optimization model that comprehensively considers geometric constraints and process constraints, the coordinated optimization of cutting path and EDM parameters is achieved, and the contradiction between machining accuracy, machining time and energy consumption is balanced, so that the system can achieve the best process matching on workpieces of different materials and geometric complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0009] Figure 1 A schematic diagram of an embodiment of a method for intelligent planning of cutting paths for a wire cutting machine based on image processing in an embodiment of the present invention; Figure 2Schematic diagram of an embodiment of a wire cutting machine cutting path intelligent planning system based on image processing in an embodiment of the present invention. DETAILED DESCRIPTION

[0010] Embodiments of the present invention provide a method and system for intelligent planning of cutting paths for wire cutting machines based on image processing. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0011] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the wire cutting machine cutting path intelligent planning method based on image processing in the embodiment of the present invention includes: Step S101, performing multi-angle image acquisition and preprocessing on the workpiece to be processed to obtain a workpiece feature map; It is understandable that the execution subject of the present invention may be a wire cutting machine cutting path intelligent planning system based on image processing, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0012] Specifically, a high-precision, multi-angle image acquisition system is constructed. The system is based on an industrial-grade high-resolution camera with a resolution of no less than 4096×3072 pixels and is equipped with a ring-shaped LED light source to ensure shadow-free and uniform illumination. It is fixed above the machine tool worktable and maintained at a vertical installation distance of 350mm to obtain workpiece images that meet the geometric accuracy requirements. On this basis, the original image datasets of multiple sides of the workpiece are collected by multi-angle rotation to ensure the image integrity and detail expression ability of complex geometric shapes at different angles, and obtain the original image sequence. An adaptive Gaussian filtering operation is performed on the original image sequence, and a dynamic kernel size adjustment mechanism is used to control the filter kernel size between 3×3 and 7×7 according to the local noise distribution of the image. In this way, random high-frequency noise is effectively suppressed, edge structure information is retained, and a denoised image dataset is generated. Histogram equalization is performed on the denoised image dataset, and the image contrast is improved and the boundary grayscale gradient is enhanced by the piecewise linear mapping method. At the same time, the bicubic interpolation algorithm is applied for geometric correction to eliminate the deformation error introduced by image distortion, and a corrected image dataset with unified perspective and enhanced details is obtained. In order to extract the workpiece body area, the corrected image is segmented, and the improved Otsu threshold method is combined with morphological processing such as closing operation and corrosion and expansion to extract a continuous and complete workpiece contour map. The contour information from different angles is integrated into a single image representation through coordinate registration and image fusion technology. Edge detection processing is performed on the workpiece contour map, and the improved Canny algorithm is used to calculate the edge intensity map in combination with the gradient amplitude and direction information to enhance the gradient response of features such as corners, holes, slopes and slender edges. The edge intensity map is uniformly converted to a standard size and the pixel value is normalized to generate the workpiece feature map.

[0013] Step S102, inputting the workpiece feature map into the perception area recognition network for processing to obtain a target processing area recognition map; Specifically, the standardized workpiece feature map is input into two parallel substructures of the perception area recognition network, where the residual feature extraction branch adopts a four-level cascaded residual block architecture. Each level of residual block adopts a bottleneck structure, which includes a deep module composed of a convolutional layer, a normalization layer and a ReLU activation function. It has the ability to compress and expand feature channels layer by layer, thereby performing step-by-step spatial structure abstraction and deep feature extraction of the image, and generating a main feature map with high-level semantic information expression capabilities. The main feature map effectively identifies geometric key areas such as sharp corners, inner holes, and slits. At the same time, the workpiece feature map texture preservation branch is composed of three levels of convolutional layers with different receptive field scales. Each level of convolution kernel adopts different sizes of 3×3, 5×5 and 7×7, respectively, and is supplemented by a multi-scale filtering mechanism to retain edge details and local texture changes, thereby generating an auxiliary feature map with multi-scale resolution capabilities. Based on the heterogeneous feature pyramid technology, a multi-level skip connection path is established between the main feature map and the auxiliary feature map. This connection mechanism allows the low-level texture features and high-level semantic features to interact and fuse in a multi-scale structure, avoiding the problems of information isolation and feature faults, and forming an enhanced feature set. A channel-level adaptive weight calculation mechanism is applied to the enhanced feature set. By performing average pooling, full connection mapping and Softmax normalization operations on each feature channel, the weight coefficient of each channel in process judgment is calculated to generate a process optimization feature representation with both regional significance and process sensitivity. The process optimization feature representation is input into the pixel-level classification decoder, which adopts a combined structure of deconvolution upsampling and skip connection to gradually restore the spatial resolution and perform precise boundary positioning and target area screening. The processing area probability of each pixel is evaluated through the Softmax classification layer, and the target processing area identification map is output.

[0014] Step S103, performing feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map; Specifically, with the target processing area identification map as input, N local patch images centered on the key points of the target area are extracted from the image through complex feature sampling operations. Each patch is 32×32 pixels in size and covers complex geometric features such as sharp corners, inner hole edges, and sharp transition areas, forming a complex feature patch set of workpieces with structural recognition value. The patch set is input into the identity-related feature reconstruction module composed of global and local attention supplementary branches. The module extracts local geometric representations layer by layer through a three-level convolutional structure, and introduces an attention mechanism in the feature channel dimension. The response intensity of each feature channel is modeled differentially through the local perception path and the global context path, and a feature importance weight matrix is ​​generated. The matrix assigns different response intensities to the features of different areas in each patch, effectively amplifying the recognition sensitivity of the key processing area. On this basis, the original target processing area recognition map is synchronously input into the enhanced hierarchical feature fusion branch, where the initial feature extraction layer extracts the basic spatial structure features through standard convolution to generate the basic feature map; then enters the hierarchical feature extraction module with three-layer structure progression, each layer uses dense connection strategy to ensure the integrity of feature transmission while maintaining the expansion of the receptive field, forming a multi-scale feature set containing low-level details, middle-level structure and high-level semantics. For the multi-scale feature set, in order to make full use of the expression ability of different feature levels, a differentiated weighted integration mechanism is adopted. The attention weight of each scale feature map is calculated through global average pooling and fully connected layers, and fusion is performed accordingly to obtain a fused feature representation that integrates boundary perception, morphological understanding and processing intention. In the fusion stage, the fused feature representation is combined with the aforementioned feature importance weight matrix across branches, and the channel response value at each position in the fused feature map is multiplicatively coupled with the weight matrix value at the corresponding position, and the information is integrated through weighted accumulation on the cross-scale structure to generate an enhanced feature map. At the same time, the channel attention response generated in the fusion process is independently output as an importance weight map.

[0015] The basic feature map is input into the first hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for feature decomposition. The local spatial variation information is extracted using the standard convolution module, and a preliminary first-layer feature map is constructed. In order to expand the receptive field and enhance the feature expression dimension, the first-layer feature map is subjected to residual feature expansion processing. In this process, 1×1 convolution and 3×3 convolution kernels are used alternately, and a jump connection structure is used to maintain the continuity of the original feature path to generate the first-layer enhanced feature map. The first-layer enhanced feature map is input into the second-layer hierarchical feature extraction layer. In this layer, in addition to performing conventional spatial abstract convolution, a jump connection channel from the first-layer enhanced feature map to the third layer is also constructed synchronously. This channel maintains the resolution of the original texture feature in the form of a full channel. At the same time, the self-attention mechanism is introduced inside the second layer to perform intra-channel attention screening on the second-layer feature map. Redundant or non-critical information is suppressed through inter-channel importance calculation to generate the second-layer attention weighted feature map. The skip path data of the second-layer attention weighted feature map and the first-layer enhanced feature map are input into the third-layer hierarchical feature extraction layer, which focuses on feature complementary fusion processing and uses dense connection structure to perform spatial alignment and dimensional fusion of cross-layer data to generate the third-layer fusion feature map. The first-layer enhanced feature map, the second-layer attention weighted feature map and the third-layer fusion feature map are input into the multi-scale aggregation module, and a multi-scale feature set is formed by feature channel compression, spatial weighted fusion and multi-scale response superposition.

[0016] Step S104, performing cutting path planning according to the enhanced feature map and the importance weight map to obtain an initial cutting path; Specifically, by performing edge extraction operations on the enhanced feature map, the improved Canny edge detection algorithm based on gradient response enhancement is used to obtain the initial edge point set. The main contour segments are retained during the extraction process, and the sharp corners, hole edges and curvature mutation points in the image are fully responded to to ensure the complete recognition of complex areas. Subsequently, the edge points are weighted and adjusted in combination with the importance weight map. The position of each edge point is fine-tuned by superimposing the response coefficient of its corresponding position in the weight map on its position coordinates, and a set of first edge point sets reflecting the geometric feature accuracy and process priority is generated. The first edge point set is prioritized, and a priority function is constructed according to the importance response value, local edge continuity and geometric stability of the corresponding area of ​​the point. The sorting result is constructed as the second edge point set, which is reorganized and arranged according to the local structural continuity, so that the fitting process has the basis of contour semantic consistency. On this basis, the second edge point set is subjected to curve fitting processing, and the discrete edge points are constructed into the first contour curve using the third-order B-spline interpolation algorithm. The curve has C² continuity and ensures the smoothness of movement during the electrode wire cutting process. In order to improve the geometric adaptability of the path to complex structures, the control point density of the first contour curve is subdivided based on the importance weight map. The control point distribution density is increased in the area with higher weight response value, and the low density is maintained in the straight contour area to generate the second contour curve, which can accurately cover key processing structures such as tiny holes, sharp corner transitions and concave areas. Based on the geometric configuration of the second contour curve, the entry point is optimized. By analyzing the material thickness gradient, the local edge curvature distribution and the internal stress field estimation results of the workpiece, a comprehensive scoring function is constructed. The position with the lowest stress concentration, the most stable curvature and the moderate material thickness is selected as the optimal processing starting position on the contour curve to ensure that the heat affected area and path stability in the initial stage of processing are optimal. Based on the curvature function of the second contour curve and the regional distribution information of the importance weight map, an adaptive feed speed scheduling mechanism is established. By reducing the feed speed in high curvature or high weight areas and increasing the cutting speed in straight low weight areas, a feed speed distribution function is constructed. Taking the second contour curve as the path basis, the optimized processing starting point is used as the path starting point, and the cutting speed of each segment on the path is assigned segmentally according to the cutting feed speed function to construct an initial cutting path with continuous structure, adjustable speed and reasonable starting point, which includes the electrode wire center trajectory. The electrode wire compensation trajectory is constructed by adding offset parameters according to the material properties to form standardized path data.

[0017] Step S105, performing multi-objective optimization on the initial cutting path to obtain a target cutting path and a corresponding electric spark parameter configuration table.

[0018] Specifically, combined with the actual processing requirements, the processing accuracy requirements, workpiece contour tolerance, electrode wire tension limit and maximum cutting speed are set as hard constraints of the optimization model, and a joint objective function is constructed for the three aspects of accuracy, time and energy consumption to form a multi-objective optimization model with a reasonable structure. The enhanced feature map is discretized, and the image space is divided into several grid units using a fixed-size two-dimensional grid division method. Differentiated weight coefficients are assigned to each grid unit according to the grayscale response intensity of the corresponding area in the importance weight map, forming a feature importance matrix with spatial regional significance annotation capabilities. This matrix is ​​used to judge the influence level of each segment in the path on the processing quality and control strategy. The initial cutting path is input into the optimization model and combined with the feature importance matrix to perform the particle swarm optimization process. In this process, each particle encoding contains the coordinate vector of the path control point and the local cutting parameter vector, such as feed speed, electrode wire compensation, etc. The optimization algorithm introduces an adaptive inertia weight strategy to dynamically adjust the particle search range and convergence speed. Its weight decays nonlinearly with the increase of the number of iterations, effectively avoiding the local optimal trap and obtaining an iteratively updated path optimization solution. In order to ensure the physical continuity of the path and the motion stability during the machining process, a curvature penalty factor is added as a regularization term in the optimization scheme to suppress the sudden change of curvature between adjacent path segments. By calculating the local curvature change at the control point and constraining the accumulation of its square difference, the target cutting path with high curvature consistency and smoothness is obtained. In order to ensure the use of the most appropriate EDM parameters in different workpiece areas, a material-parameter mapping function is constructed based on the workpiece material characteristic parameter matrix and the feature importance matrix. The material parameter matrix includes indicators such as conductivity, thermal conductivity, and hardness, and the feature importance matrix provides the regional machining complexity weight. The two are established through interpolation fitting or neural mapping models. The optimal voltage and current combination parameters are dynamically calculated at each position point on the target cutting path to generate an EDM parameter configuration table. The configuration table has spatial position resolution and can achieve fine matching of discharge power in different machining areas in the control system.

[0019] In the embodiment of the present invention, through the multi-angle acquisition and adaptive preprocessing technology of the industrial camera, the geometric feature information of the workpiece can be fully captured, effectively overcoming the information loss problem caused by the traditional single-view acquisition, and providing a high-quality workpiece feature map for subsequent accurate identification. The dual-stream feature extraction architecture of the perception area recognition network is adopted to realize the coordinated extraction of macroscopic geometric features and microscopic texture details, solve the problem of information transmission loss in the traditional method when processing sharp angles and fine features, and accurately identify the target area that needs differential processing. Through the supplementation and fusion mechanism of global and local features, the system can pay special attention to small feature areas such as sharp corners and fine holes, accurately depict the complex contour features of the workpiece, and solve the problem of insufficient recognition of small features by traditional methods. Intelligent path planning based on enhanced feature maps and importance weight maps can implement differentiated control point density settings and feed speed adjustments for different geometric feature areas, ensure stable and efficient cutting process, and significantly improve the quality of complex contour processing. By establishing a multi-objective optimization model that comprehensively considers geometric constraints and process constraints, the coordinated optimization of cutting paths and EDM parameters is realized, and the contradiction between processing accuracy, processing time and energy consumption is balanced, so that the system can achieve the best process matching on workpieces of different materials and geometric complexity.

[0020] In a specific embodiment, the process of executing step S101 may specifically include the following steps: Use an industrial camera to collect high-resolution images of the workpiece at multiple angles to obtain a data set of original images of the workpiece, and perform adaptive Gaussian filtering on the original image data set to obtain a denoised image data set; Performing histogram equalization and geometric correction processing on the denoised image data set to obtain a corrected image data set, and segmenting and fusing the corrected image data set to obtain a workpiece contour map; Edge detection is performed on the workpiece contour map to obtain an edge intensity map, which is uniformly converted into a standard size and subjected to pixel value normalization to obtain a workpiece feature map.

[0021] Specifically, a high-resolution industrial camera system is used to perform multi-angle image acquisition operations on the workpiece to be processed. The system consists of a main camera with a pixel size of no less than 4096×3072, an autofocus lens, a high color rendering LED ring light source, and a precision rotation angle control platform. During the acquisition process, the camera is fixed on the upper reference axis of the wire cutting equipment and maintains the optical axis direction perpendicular to the workpiece surface. At the same time, the workpiece is rotated at multiple angles by the rotating platform in an equiangular stepping manner, so that the camera captures a full-view image at each angle. Under a complete rotation angle sequence, the original image data set covering the side profiles and detailed structures of the workpiece is obtained. The original image data set of the workpiece is subjected to preliminary denoising. An adaptive Gaussian filtering mechanism is introduced to achieve local image smoothing on the premise of retaining edge features. For each frame of the original image, a local variance perception strategy is used to dynamically adjust the filter kernel size, so that a larger 7×7 convolution kernel is used in the noise-strong area, and a smaller 3×3 or 5×5 kernel is used for filtering in the texture-clear area, forming a uniform and effective denoised image data set. During the filtering process, the kernel function adopts a zero-mean two-dimensional Gaussian distribution model, and its standard deviation σ is automatically adjusted according to the local brightness change to ensure the edge integrity of the structural area. The denoised image data set is input into the image enhancement module in turn, and histogram equalization is performed on it. This process takes the uniformity of grayscale distribution as the optimization goal. By redistributing the pixel grayscale values, the contrast of the image is improved in low-light or strongly reflected areas, thereby improving the overall grayscale hierarchy and texture clarity of the image. After the histogram equalization is completed, the geometric correction mechanism is introduced, and the image is dedistorted using the calibration data based on the camera's intrinsic and extrinsic matrix. In particular, for the radial distortion and tangential distortion problems of industrial cameras, the reverse mapping relationship between pixel coordinates and world coordinates is established, and the sub-pixel reconstruction of the original pixel position is performed in combination with the bicubic interpolation algorithm to generate a geometrically accurate corrected image data set. After the geometric correction is completed, in order to extract the main contour of the workpiece structure, the image data set is subjected to regional segmentation and perspective fusion processing. The regional segmentation adopts the improved Otsu threshold algorithm combined with morphological operations, which can effectively separate the workpiece body from the background and eliminate artifacts such as adhesion, holes, and fractures. For the structural fusion between multi-angle images, the contour lines extracted at different angles are aligned and superimposed according to the pose and reconstructed by using the multi-image fusion strategy based on affine transformation and feature point registration to form a continuous, complete workpiece contour map with three-dimensional depth perception potential. The contour map is subjected to edge detection operation. In the edge detection stage, the improved Canny algorithm based on image gradient field analysis is adopted, the intensity gradient value at the edge point is considered, and the direction information, response connectivity and noise-background response ratio are integrated to extract the edge intensity map through non-maximum suppression and dual threshold tracking mechanism. The edge intensity map contains the complete workpiece geometric contour, inner and outer boundaries, sharp corner position and curvature change area, and is the key carrier for constructing the workpiece feature space representation.Perform image normalization on the edge intensity map. The original image is uniformly scaled to 512×512 pixels through size normalization to adapt it to the fixed input size network model and avoid the misalignment of the network layer input and output size due to size inconsistency; the pixel value normalization operation linearly maps the grayscale value from the original [0,255] to [0,1], making the subsequent convolution operation more stable and accelerating the network convergence. The edge map after size normalization and pixel normalization constitutes the standardized workpiece feature map.

[0022] In a specific embodiment, the process of executing step S102 may specifically include the following steps: The workpiece feature map is input into a perception region recognition network, which includes a residual feature extraction branch and a texture preservation branch; The workpiece feature map is progressively extracted with deep features through the four-stage cascade residual block in the residual feature extraction branch to obtain the main feature map; The workpiece feature map is input into the texture preservation branch, and then cascaded and processed by three levels of convolution layers with different scales and multi-scale filtering to obtain an auxiliary feature map. Based on the heterogeneous feature pyramid technology, a multi-level skip connection path between the main feature map and the auxiliary feature map is established to obtain an enhanced feature set; A channel adaptive weight calculation mechanism preset by the wire cutting process is performed on the enhanced feature set to obtain a process optimization feature representation; The process optimization feature representation is input into the pixel-level classification decoder for boundary positioning and target area screening to obtain the target processing area identification map.

[0023] Specifically, a dual-branch perceptual region recognition network with geometric perception, texture preservation and region discrimination capabilities is constructed. The network takes the standardized workpiece feature map as input. Its goal is to automatically identify the key areas suitable for wire cutting processing from the image and provide accurate and coherent regional boundary descriptions. The network consists of two parallel functional sub-networks, one of which is the residual feature extraction branch, which is used to extract the deep semantic structure in the image, and the other is the texture preservation branch, which is used to preserve local high-frequency details and edge texture information. In the residual feature extraction branch, the input image is subjected to layer-by-layer feature extraction through four-level cascaded residual blocks. Each residual block is composed of "convolution layer + batch normalization + ReLU activation + shortcut connection". The spatial features at different scales are extracted by alternating 1×1 and 3×3 convolution kernels, and the residual structure is used to prevent gradient disappearance and feature degradation, ensuring information penetration. As the network deepens layer by layer, this branch effectively captures the large-scale structural relationship, geometric configuration changes and semantic aggregation effect of the target area in the image, forming a main feature map with high expressive ability. At the same time, the original workpiece feature map is input into the texture preservation branch, which is composed of three convolution layers of different scales, using 3×3, 5×5 and 7×7 convolution kernels respectively to obtain texture details under different receptive fields, and supplemented by a multi-scale filtering processing module. The boundary details and texture discontinuities are enhanced through local response normalization and local direction enhancement mechanisms, so that it can enhance the response to process-sensitive areas such as sharp corners, holes, grooves, etc. while maintaining spatial resolution, and output an auxiliary feature map that is extremely sensitive to edge contours. In order to achieve feature synergy and semantic complementarity between the two branches, a fusion mechanism based on a heterogeneous feature pyramid structure is adopted to construct a multi-level jump connection path between the main feature map and the auxiliary feature map. The edge information in the shallow auxiliary feature map is directly transferred to the middle and high-level semantic map of the main branch through jump connections, and the feature channels are upsampled and aligned, and then cascaded splicing operations are performed to form an enhanced feature set. Based on the enhanced feature set, a channel adaptive weight calculation mechanism is introduced. By performing global average pooling on each channel feature map and sending the pooling result to a two-layer fully connected neural network and a Softmax normalization layer, the relative importance coefficient of each channel in the feature expression is calculated. The channel attention mechanism dynamically re-weights the features according to the different geometric structures and texture complexities of the workpieces, improves the response of key channels and suppresses redundant information, thus forming a process optimization feature representation.The process optimization feature representation is then input into the pixel-level classification decoder built on the decoding path. The decoder is mainly composed of a symmetrical upsampling module, a jump connection module, and a pixel-by-pixel Softmax classifier. Its workflow is as follows: the deep process optimization features are upsampled back to the original image size step by step, and the local feature maps from the shallow layer in the encoding path are fused at each level. The spatial positioning accuracy is restored through the fusion structure, and the boundary semantic consistency is enhanced; then a channel compression layer is constructed using a 1×1 convolution to output the classification probability of each pixel under the "processing / non-processing" label, and the Softmax function is used for normalization, so that the probability value corresponding to each pixel is distributed between [0,1], indicating the confidence level of its belonging to the target processing area. The entire pixel classification output is mapped to a target processing area identification map of the same size as the original image, in which each pixel value represents its credibility in the area determined to be processable.

[0024] In a specific embodiment, the process of executing step S103 may specifically include the following steps: Extract N local patch images centered on the target point from the target processing area recognition map to form a complex feature patch set of the workpiece; The complex feature patch set of the artifact is input into the identity-related feature reshaping module of the global and local attention supplementary branches for feature weight analysis to obtain the feature importance weight matrix; The target processing area identification map is input into the initial feature extraction layer of the enhanced hierarchical feature fusion branch for feature extraction to obtain a basic feature map; The basic feature map is input into the three-layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for processing to obtain a multi-scale feature set; The multi-scale feature sets are differentiated and weighted integrated to obtain a fused feature representation, and the feature importance weight matrix and the fused feature representation are combined across branches to obtain an enhanced feature map and an importance weight map.

[0025] Specifically, representative geometric key points are extracted from the target processing area recognition map as the center of local patch extraction. The target processing area recognition map represents the probability of each pixel belonging to the processing area in the form of confidence. By setting the confidence threshold, a set of pixels with high confidence and complex structure is screened out, and N representative target points are extracted in the structurally significant area using criteria such as edge gradient, curvature extremum, and connected domain center. With each target point as the center, a fixed-size window is used to crop the corresponding image patch from the original image to form a set of complex feature patches of the workpiece covering typical complex structures (such as junctions, closed holes, sharp corners, etc.). The set of complex feature patches of the workpiece is input into the identity-related feature reconstruction module, which is composed of global and local attention supplementary branches. The local path extracts fine-grained texture features of each patch through stacked convolutional layers, while the global path extracts the structural expression features and geometric context information of the patch through spatial compression, self-attention mechanism, and channel attention mechanism. In the local branch, the three layers of convolution kernels are staggered with 3×3, 5×5 and 3×3, respectively, and the Leaky ReLU activation function is combined to enhance the responsiveness to edge and texture changes; in the global branch, each patch is globally averaged and pooled, compressed into a channel vector, and then sent to two fully connected layers to extract the nonlinear mapping features between channels, and the channel weight coefficients are generated using the Sigmoid function, and then the channel coefficients are back-projected to the original space to form a weight map. The local path emphasizes boundary sensitivity, and the global path emphasizes structural consistency. After the two are fused, a unified feature importance weight matrix is ​​generated through feature stacking and channel fusion. At the same time, the target processing area recognition map is input into the initial feature extraction layer of the enhanced hierarchical feature fusion branch. The layer structure is a standard convolution layer. The first round of encoding modeling is performed on the entire image through multi-channel convolution kernels to extract the basic feature map containing boundary distribution, grayscale gradient and regional shape. This map retains the spatial structural resolution of the image. The basic feature map is input into a three-layer hierarchical feature extraction module, which uses a dense connection structure to construct a feature progression path. The intermediate features output by each layer are not only input into the next layer, but also sent to higher layers through jump connections, thereby maintaining the multi-layer transmission capability of local details. The first layer uses a smaller convolution kernel size to capture micro-structure features, the second layer focuses on shape area modeling, and the third layer focuses on contextual structure and large-scale semantic fusion. This progressive relationship constitutes a multi-scale feature set, and each scale feature map in the set corresponds to different structural scale information. Differential weighted integration processing is performed on the multi-scale feature set. Channel pooling and channel attention coefficient calculation are performed on the output feature map of each layer, and then they are uniformly mapped to a common dimension through a soft fusion mechanism. Differential fusion coefficients are generated according to the response significance and spatial activity of different channels, and fused feature representations are generated through a weighted accumulation mechanism. The fused feature representation and the feature importance weight matrix are combined across branches.Channel-by-channel multiplication is performed on the channel features of each spatial position in the fusion feature map, and multiplied with the channel weights of the corresponding weight matrix to form a weighted feature response, and then normalization is performed in the spatial dimension to construct a feature response map with consistent response strength to generate an enhanced feature map. The channel weight coefficient map output in this process is retained as an importance weight map, which records the comprehensive expression of participation, significance and geometric complexity of each area of ​​the workpiece in the feature generation process.

[0026] In a specific embodiment, the execution step inputs the basic feature map into the three-layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for processing, and the process of obtaining the multi-scale feature set may specifically include the following steps: Input the basic feature map into the first hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for feature decomposition to obtain a first-layer feature map; Apply residual feature expansion processing to the first-layer feature map, extract multi-scale texture information through alternating combinations of 1×1 and 3×3 convolution kernels, and obtain the first-layer enhanced feature map; The first layer enhanced feature map is input into the second layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch, and a skip connection channel from the first layer enhanced feature map to the third layer is constructed to obtain the second layer feature map and its skip path; Perform self-attention feature filtering on the second-layer feature map to obtain the second-layer attention weighted feature map; The skip path data of the second-layer attention weighted feature map and the first-layer enhanced feature map are input into the third-layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for feature complementary fusion to obtain the third-layer fusion feature map; Multi-scale feature aggregation is performed on the first-layer enhanced feature map, the second-layer attention weighted feature map, and the third-layer fusion feature map to obtain a multi-scale feature set.

[0027] Specifically, the basic feature map is input into the first-layer hierarchical feature extraction module of the enhanced hierarchical feature fusion branch. This module serves as the low-level entrance of the entire fusion structure and is used to perform primary structure decoupling and feature decomposition operations. At this stage, the standard convolution layer is used to extract shallow geometric texture changes, boundary contour trends and initial spatial relationships to form a first-layer feature map containing coarse-grained semantic contours. Residual feature expansion processing is applied to the first-layer feature map, and a local residual unit chain is constructed by alternately stacking 1×1 and 3×3 convolution kernels. The 1×1 convolution kernel is used for compression and mapping in the channel dimension in this structure, thereby reducing the parameter scale and improving the nonlinear transformation capability; the 3×3 convolution kernel captures the spatial structural relationship and enhances the perception of small-scale features such as pore edges and intersection angle tips. While the two convolution kernels are alternately combined, the standard residual connection is introduced to perform element-by-element addition fusion of the input feature map and the convolution stack output, so as to retain the low-level feature expression path and accumulate the deep response, and generate the first-layer enhanced feature map with stronger spatial resolution and texture discrimination. The first-layer enhanced feature map is input into the second-layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch. At the same time, a skip connection path is built inside the system, which directly reaches the subsequent third-layer fusion module from the first-layer enhanced feature map to compensate for the low-level structural information in the higher semantic layer to avoid semantic loss. The second-layer hierarchical extraction layer itself adopts a deeper receptive field and a more complex channel combination strategy to enhance the recognition ability of mesoscale structures, which is suitable for identifying transition areas of regional structures, such as surface gradients and boundary inward folds. In order to make the extraction process of this layer feature selective, the output second-layer feature map is screened by applying a self-attention mechanism. This mechanism constructs a channel description vector through global average pooling, and then sends it to a set of nonlinear transformation networks to generate channel response weights. These weights are applied to the original feature map to form the second-layer attention weighted feature map. The second-layer attention weighted feature map and the first-layer enhanced feature map are input into the third-layer hierarchical feature extraction layer through a skip channel. The third layer, as the highest layer of the entire fusion branch, focuses on the modeling of semantic configuration and regional layout. Its internal structure introduces a feature complementation mechanism, that is, channel alignment, size standardization and feature splicing operations are performed on the two input feature maps respectively. On this basis, the shallow local features are integrated with the deep global expression through the convolutional fusion structure to improve the synergy of features at different scales and structural levels, and obtain the third-layer fused feature map. Multi-scale feature aggregation operations are performed on the feature maps of the above three key levels. This operation performs channel compression processing on the first-layer enhanced feature map, the second-layer attention weighted feature map and the third-layer fused feature map. After unifying the dimensions, they are fused according to the weighted strategy in the spatial position. The fusion coefficient is set according to the source level of the feature map and its local response strength. At the same time, the regional weighting parameter is introduced in combination with the spatial attention mechanism to realize the redistribution of spatial dimension information and form a multi-scale feature set.

[0028] In a specific embodiment, the process of executing step S104 may specifically include the following steps: Perform edge extraction on the enhanced feature map to obtain multiple initial edge points, and perform weighted adjustment on the multiple initial edge points in combination with the importance weight map to obtain a first edge point set; Prioritizing the first edge point set to obtain a second edge point set, and performing curve fitting processing on the second edge point set to obtain a first contour curve; Performing control point density subdivision processing on the first contour curve based on the importance weight map to obtain a second contour curve; According to the second contour curve, the entry point is optimized by comprehensively considering the material thickness gradient, edge curvature distribution and internal stress distribution of the workpiece to obtain the best processing starting position; The feed rate is planned according to the geometric characteristics of the second contour curve and the regional weight distribution of the importance weight map, and a cutting feed rate distribution function is constructed; An initial cutting path is generated based on the second contour curve, the optimal machining start position and the cutting feed speed distribution function.

[0029] Specifically, the contour boundary of the workpiece is extracted from the enhanced feature map. The improved edge detection algorithm is used to make full use of the grayscale gradient information in the image to identify the area with drastic edge changes. The extracted initial edge points are distributed on the outer boundary, inner hole boundary and local geometric mutation area of ​​the workpiece to form an edge point set. The edge points are combined with the importance weight map. By reading the response value of each edge point in the weight map, the spatial position and processing weight of the edge point are fine-tuned, so that the edge points in the high-weight area are offset to the structural core and obtain a higher priority, while the edge points in the low-weight area are slightly weakened to obtain the first edge point set. The first edge point set is sorted. Based on the spatial distribution of edge points in the image, combined with multiple factors such as the process weight, curvature characteristics of each point and the connection coherence with the surrounding points, it is ensured that the sorting result forms a logically closed and contour-reasonable point sequence in the subsequent curve fitting. On this basis, the second edge point set is fitted into a first contour curve with geometric continuity through curve fitting technology. Based on the regional response information of the importance weight map, the distribution density of the curve control points is readjusted. In high-weight areas, such as sharp corners, slits, micropores and other complex structural parts, more dense control points are added to improve fitting accuracy. In gentle areas, the number of control points is appropriately simplified to reduce calculation redundancy, and a second contour curve with stronger structural adaptability and higher processing guidance is obtained. According to the second contour curve, the entry point optimization is performed by comprehensively considering the material thickness change gradient, edge curvature distribution and internal stress distribution of the workpiece. The starting position avoids the thickness mutation, high curvature corner and stress concentration area, and gives priority to the area with stable structure, gentle curvature and small stress gradient to ensure balanced heat distribution and good electrode wire stability during the cutting process, so as to select the best processing starting point. According to the geometric characteristics of the second contour curve and the response intensity of the corresponding area in the importance weight map, a set of cutting feed speed scheduling strategies is formulated. In areas with complex geometric structures and high weights, the feed speed is appropriately reduced to improve processing accuracy and avoid problems such as ablation, wire drawing and edge burrs; while in areas with simple structures, smooth boundaries and low weights, the feed speed is increased to improve processing efficiency. The speed scheduling mechanism is dynamically adjusted based on each position point of the path, so that the entire machining process can meet both structural accuracy and optimize the overall time consumption. Based on the second contour curve, the initial cutting path is constructed in combination with the determined starting position and the corresponding feed speed scheduling information. The path includes the center motion trajectory of the electrode wire, and the trajectory compensation is appropriately performed according to the machining requirements of different areas, and different feed speeds are preset. Through the above steps, the initial path obtained has continuity in structure, adaptability in process, and flexibility in speed control.

[0030] In a specific embodiment, the process of executing step S105 may specifically include the following steps: A multi-objective optimization model is established by taking machining accuracy requirements, workpiece contour tolerance, electrode wire tension limit and maximum cutting speed as constraints; Discretize the image features of the enhanced feature map to obtain multiple grid units, and assign different weight values ​​to each grid unit according to the importance weight map to obtain the feature importance matrix; Particle swarm optimization is performed on the initial cutting path according to the feature importance matrix and the multi-objective optimization model. The particle encoding includes the coordinates of the path control points and the cutting parameters, and an adaptive inertia weight strategy is introduced to obtain a path optimization iterative solution. Adding a curvature penalty factor to the path optimization iteration scheme limits the curvature change of adjacent path segments to obtain the target cutting path; The material-parameter mapping function is constructed based on the workpiece material characteristic parameter matrix and the feature importance matrix. The optimal voltage and current parameter combinations at different positions of the target cutting path are dynamically calculated through the material-parameter mapping function to obtain the EDM parameter configuration table.

[0031] Specifically, the constraint system is clarified according to the actual engineering needs, in which the machining accuracy requirements determine the upper limit of the density of the path control points and the lower limit of the fitting accuracy, the workpiece contour tolerance determines the maximum range of deviation from the original contour, the electrode wire tension limit constrains the path curvature change, the minimum bending radius and the angle turning strength, and the maximum cutting speed limits the upper limit of the machining advancement rate of each section along the path. The above factors not only put forward specific restrictions on the spatial geometric structure, but also form linkage constraints on trajectory smoothness, speed stability and electrical parameter rationality at the level of process dynamic execution. Based on these parameters, a multi-objective optimization model with the core goals of minimizing path smoothness, minimizing processing time, minimizing energy consumption and maximizing process stability is established. Each optimization goal has a corresponding weight distribution in the solution process, allowing the system to adjust the optimization focus according to the priority of different processing stages. The image feature discretization processing is performed on the enhanced feature map. The entire enhanced feature map is divided into uniform two-dimensional grid units to form a regular grid structure. In each grid unit, the texture complexity, gradient response value and structural mutation characteristics of the image area are extracted, and the processing weight value is applied to each unit in combination with the importance weight map, so that the high-complexity area or the key processing area has a higher degree of attention and parameter adjustment sensitivity during the optimization process. After the above processing, a feature importance matrix based on the image coordinates and weighted by the process priority is obtained. The initial cutting path is optimized and solved with the feature importance matrix and the established multi-objective optimization model as input. In this stage, the particle swarm optimization algorithm is introduced as the core search tool, and the coordinates of the path control points and the process parameter values ​​are jointly encoded into particle vectors to form an optimization unit that can be iteratively adjusted in the search space. In each round of iteration, each particle updates its position and speed according to its historical optimal state and global optimal state, and at the same time evaluates its comprehensive fitness under multi-objective indicators such as processing accuracy, energy consumption, and speed smoothness in the path space. In order to improve the convergence efficiency and avoid falling into the local optimum, an adaptive inertia weight mechanism is introduced, that is, as the number of iterations increases, the inertia factor of the particle gradually decays, so that the early search has the ability of breadth exploration, and the later convergence has the ability of local refinement, thereby improving the globality and stability of the overall optimization results. During the execution of particle swarm optimization, in order to prevent the path from geometric mutation or physical unprocessable state, an additional curvature penalty term is added to the fitness function to limit the curvature fluctuation value between adjacent control points of the path. The path sequence corresponding to the particle is geometrically analyzed to identify the angle change and curvature deviation between continuous segments. When the curvature mutation of a certain segment exceeds the preset threshold, its fitness is automatically penalized, forcing the optimization algorithm to complete the search update on the basis of maintaining the smoothness of the path. The introduction of the curvature penalty term ensures that the target cutting path obtained not only meets the structural accuracy requirements, but also has a good wire tension control foundation and path continuity advantages at the physical execution level.After the target path is determined, the optimal electrical parameter combination is configured for each position point on the path according to the properties of the processed material. At this time, a material-parameter mapping model is constructed. The model uses the conductivity, thermal conductivity, hardness, melting point, specific heat capacity and other parameters of the workpiece material as input variables. At the same time, combined with the response value of the target path location in the feature importance matrix, the two types of information are combined to generate a mapping function. This function is implemented through empirical formulas, interpolation methods or lightweight neural networks, and its output is the optimal discharge control parameters such as voltage, current or pulse width in the current area. The goal of the entire mapping mechanism is to allocate low-energy and high-precision discharge strategies in high-importance areas, while allowing high-energy and high-speed cutting strategies in low-complexity areas, so as to maximize the overall cutting efficiency while ensuring boundary accuracy and surface quality. After completing the above path space optimization and parameter space configuration, the target cutting path and EDM parameter configuration table are formed. The configuration table is indexed by the points along the path, and each point contains key information such as spatial position, target speed, voltage and current parameters, and discharge timing settings.

[0032] The above describes the wire cutting machine cutting path intelligent planning method based on image processing in the embodiment of the present invention. The following describes the wire cutting machine cutting path intelligent planning system based on image processing in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an intelligent planning system for cutting paths of wire cutting machines based on image processing includes: The acquisition module 201 is used to acquire and preprocess multi-angle images of the workpiece to be processed to obtain a workpiece feature map; The region recognition module 202 is used to input the workpiece feature map into the perception region recognition network for processing to obtain a target processing region recognition map; A feature enhancement module 203 is used to perform feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map; A path planning module 204 is used to plan a cutting path according to the enhanced feature map and the importance weight map to obtain an initial cutting path; The multi-objective optimization module 205 is used to perform multi-objective optimization on the initial cutting path to obtain a target cutting path and a corresponding electric spark parameter configuration table.

[0033] Through the synergy of the above components, the multi-angle acquisition and adaptive preprocessing technology of industrial cameras can fully capture the geometric feature information of the workpiece, effectively overcome the information loss problem caused by traditional single-view acquisition, and provide high-quality workpiece feature maps for subsequent accurate identification. The dual-stream feature extraction architecture of the perception area recognition network is adopted to realize the coordinated extraction of macro-geometric features and micro-texture details, solve the problem of information transmission loss in traditional methods when dealing with sharp angles and small features, and accurately identify the target areas that need differentiated processing. Through the supplementation and fusion mechanism of global and local features, the system can pay special attention to small feature areas such as sharp corners and fine holes, accurately depict the complex contour features of the workpiece, and solve the problem of insufficient recognition of small features by traditional methods. Intelligent path planning based on enhanced feature maps and importance weight maps can implement differentiated control point density settings and feed speed adjustments for different geometric feature areas, ensure stable and efficient cutting process, and significantly improve the quality of complex contour processing. By establishing a multi-objective optimization model that comprehensively considers geometric constraints and process constraints, the coordinated optimization of cutting path and EDM parameters is achieved, and the contradiction between machining accuracy, machining time and energy consumption is balanced, so that the system can achieve the best process matching on workpieces of different materials and geometric complexity.

[0034] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0035] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or partly contributed to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an image processing-based wire cutting machine cutting path intelligent planning device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0036] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent planning of cutting paths of wire cutting machines based on image processing, characterized in that: include: Perform multi-angle image acquisition and preprocessing on the workpiece to be processed to obtain the workpiece feature map; Inputting the workpiece feature map into a perception area recognition network for processing to obtain a target processing area recognition map; Performing feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map; Performing cutting path planning according to the enhanced feature map and the importance weight map to obtain an initial cutting path; The initial cutting path is subjected to multi-objective optimization to obtain a target cutting path and a corresponding electric spark parameter configuration table.

2. The method for intelligent planning of cutting paths of wire cutting machines based on image processing according to claim 1, characterized in that: The multi-angle image acquisition and preprocessing of the workpiece to be processed to obtain the workpiece feature map includes: Using an industrial camera to collect multi-angle high-resolution images of the workpiece to be processed to obtain a workpiece original image data set, and performing adaptive Gaussian filtering on the workpiece original image data set to obtain a noise reduction image data set; Performing histogram equalization and geometric correction processing on the denoised image data set to obtain a corrected image data set, and segmenting and fusing the corrected image data set to obtain a workpiece contour map; Edge detection processing is performed on the workpiece contour map to obtain an edge intensity map, and the edge intensity map is uniformly converted into a standard size and pixel value normalization processing is performed to obtain a workpiece feature map.

3. The method for intelligent planning of cutting paths of wire cutting machines based on image processing according to claim 1, characterized in that: The step of inputting the workpiece feature map into a perception area recognition network for processing to obtain a target processing area recognition map comprises: Inputting the workpiece feature map into a perception region recognition network, wherein the perception region recognition network includes a residual feature extraction branch and a texture preservation branch; Performing progressive depth feature extraction on the workpiece feature map through a four-stage cascade residual block in a residual feature extraction branch to obtain a main feature map; Input the workpiece feature map into the texture preservation branch, perform cascade and multi-scale filtering on the three-level convolution layers with different scales, and obtain an auxiliary feature map; Based on the heterogeneous feature pyramid technology, a multi-level skip connection path between the main feature map and the auxiliary feature map is established to obtain an enhanced feature set; Executing a channel adaptive weight calculation mechanism preset by the wire cutting process on the enhanced feature set to obtain a process optimization feature representation; The process optimization feature representation is input into a pixel-level classification decoder for boundary positioning and target area screening to obtain a target processing area identification map.

4. The method for intelligent planning of cutting paths of wire cutting machines based on image processing according to claim 1, characterized in that: The step of performing feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map includes: Extracting N local patch images centered on the target point from the target processing area identification map to form a workpiece complex feature patch set; Inputting the complex feature patch set of the workpiece into the identity-related feature reshaping module of the global and local attention supplementary branches to perform feature weight analysis to obtain a feature importance weight matrix; Inputting the target processing area identification map into the initial feature extraction layer of the enhanced hierarchical feature fusion branch to perform feature extraction to obtain a basic feature map; Inputting the basic feature map into the three-layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for processing to obtain a multi-scale feature set; The multi-scale feature set is differentially weighted and integrated to obtain a fused feature representation, and the feature importance weight matrix and the fused feature representation are cross-branch combined to obtain an enhanced feature map and an importance weight map.

5. The method for intelligent planning of cutting paths of wire cutting machines based on image processing according to claim 4 is characterized in that: The basic feature map is input into the three-layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for processing to obtain a multi-scale feature set, including: Inputting the basic feature map into the first hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for feature decomposition to obtain a first-layer feature map; Apply residual feature expansion processing to the first layer feature map, extract multi-scale texture information by alternating 1×1 and 3×3 convolution kernels, and obtain the first layer enhanced feature map; Inputting the first layer enhanced feature map into the second layer hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch, and constructing a skip connection channel from the first layer enhanced feature map to the third layer to obtain a second layer feature map and its skip path; Performing self-attention feature screening on the second-layer feature map to obtain a second-layer attention weighted feature map; Input the skip path data of the second layer of attention weighted feature map and the first layer of enhanced feature map into the third layer of hierarchical feature extraction layer of the enhanced hierarchical feature fusion branch for feature complementary fusion to obtain a third layer of fused feature map; Multi-scale feature aggregation is performed on the first layer enhanced feature map, the second layer attention weighted feature map, and the third layer fusion feature map to obtain a multi-scale feature set.

6. The method for intelligent planning of cutting paths of wire cutting machines based on image processing according to claim 1, characterized in that: The step of performing cutting path planning according to the enhanced feature map and the importance weight map to obtain an initial cutting path includes: Performing edge extraction on the enhanced feature map to obtain a plurality of initial edge points, and performing weighted adjustment on the plurality of initial edge points in combination with the importance weight map to obtain a first edge point set; Prioritizing the first edge point set to obtain a second edge point set, and performing curve fitting processing on the second edge point set to obtain a first contour curve; Performing control point density subdivision processing on the first contour curve based on the importance weight map to obtain a second contour curve; According to the second contour curve, the entry point is optimized by comprehensively considering the material thickness gradient, the edge curvature distribution and the internal stress distribution of the workpiece to obtain the best processing starting position; Perform feed rate planning according to the geometric characteristics of the second contour curve and the regional weight distribution of the importance weight map, and construct a cutting feed rate distribution function; An initial cutting path is generated based on the second contour curve, the optimal machining start position and the cutting feed speed distribution function.

7. The method for intelligent planning of cutting paths of wire cutting machines based on image processing according to claim 1, characterized in that: The multi-objective optimization of the initial cutting path to obtain a target cutting path and a corresponding electric spark parameter configuration table includes: A multi-objective optimization model is established by taking machining accuracy requirements, workpiece contour tolerance, electrode wire tension limit and maximum cutting speed as constraints; Discretizing the image features of the enhanced feature map to obtain a plurality of grid units, and assigning different weight values ​​to each grid unit according to the importance weight map to obtain a feature importance matrix; Performing particle swarm optimization on the initial cutting path according to the feature importance matrix and the multi-objective optimization model, the particle encoding includes the coordinates of the path control points and the cutting parameters, and introducing an adaptive inertia weight strategy to obtain a path optimization iterative solution; Adding a curvature penalty factor to the path optimization iteration scheme to limit the curvature change of adjacent path segments to obtain a target cutting path; A material-parameter mapping function is constructed based on the workpiece material characteristic parameter matrix and the characteristic importance matrix, and the optimal voltage and current parameter combination of the target cutting path at different positions is dynamically calculated through the material-parameter mapping function to obtain an EDM parameter configuration table.

8. An intelligent planning system for cutting paths of wire cutting machines based on image processing, characterized in that: Used to implement the wire cutting machine cutting path intelligent planning method based on image processing according to any one of claims 1 to 7, the wire cutting machine cutting path intelligent planning system based on image processing comprises: An acquisition module is used to acquire and preprocess multi-angle images of the workpiece to be processed to obtain a workpiece feature map; A region recognition module, used for inputting the workpiece feature map into a perception region recognition network for processing to obtain a target processing region recognition map; A feature enhancement module, used to perform feature enhancement on the target processing area identification map to obtain an enhanced feature map and an importance weight map; A path planning module, used for performing cutting path planning according to the enhanced feature map and the importance weight map to obtain an initial cutting path; The multi-objective optimization module is used to perform multi-objective optimization on the initial cutting path to obtain a target cutting path and a corresponding electric spark parameter configuration table.

Images