A deep learning-based precision injection molded part defect detection method

Abstract

The application discloses a kind of precision injection part defect detection methods based on deep learning, it is related to computer vision technical field, including, to standardization multi-view image sequence is initialized and reconstruction, generate high-fidelity three-dimensional geometric structure, simultaneously identify the area that can cause false detection, generate view angle occlusion mask set;Based on high-fidelity three-dimensional geometric structure, construct enhanced defect training dataset and defect detection model, and update defect detection model weight by back propagation, generate trained defect detection model;Through trained defect detection model, from standardization multi-view image sequence extract two-dimensional texture feature map set and two-dimensional geometric prior feature map set, and generate defect detection set in conjunction with view angle occlusion mask set;From defect detection set, filter out hierarchical defect set, and carry out size quantification, generate structured detection report in conjunction with quantity determination rule base.The application achieves the effect of reducing false detection caused by structure occlusion.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for detecting defects in precision injection molded parts based on deep learning. Background Technology

[0002] Precision injection molded parts are widely used in high-end manufacturing, and their surface quality directly affects product performance. Currently, defect detection in precision injection molded parts primarily relies on machine vision methods. Conventional methods typically involve acquiring images of the workpiece surface using industrial cameras and then combining image filtering, edge enhancement, and texture feature extraction to identify defect areas. With the development of artificial intelligence technology, some solutions have begun to incorporate deep learning network architectures. These models automatically learn defect features through training, thereby improving the level of automation and recognition efficiency, and meeting the basic requirements of production lines for high-precision inspection.

[0003] However, existing conventional methods have limitations in handling complex geometries. On the one hand, traditional models rely heavily on two-dimensional texture information and lack three-dimensional geometric prior fusion, making it difficult to distinguish between surface textures and real defects, and susceptible to false detections due to structural occlusion interference. On the other hand, the detection process lacks an uncertainty quantification mechanism, making it impossible to effectively classify low-confidence results, and thus failing to meet the requirements of precision manufacturing for quantifying defect physical dimensions and assessing reliability. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for detecting defects in precision injection molded parts, which solves the problems of easy misdetection of complex geometries and insufficient reliability assessment of the physical dimensions of defects.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a deep learning-based method for detecting defects in precision injection molded parts, comprising: acquiring and preprocessing multi-view original RGB image sequences to generate standardized multi-view image sequences; initializing and reconstructing the standardized multi-view image sequences to generate high-fidelity three-dimensional geometric structures, while identifying regions that may lead to false detections, and generating a set of view occlusion masks; constructing an enhanced defect training dataset and a defect detection model based on the high-fidelity three-dimensional geometric structures, and updating the weights of the defect detection model through backpropagation to generate a trained defect detection model; extracting a set of two-dimensional texture feature maps and a set of two-dimensional geometric prior feature maps from the standardized multi-view image sequences using the trained defect detection model, and combining them with the set of view occlusion masks to generate a defect detection set; selecting a hierarchical defect set from the defect detection set, quantizing its size, and combining it with a quantity judgment rule base to generate a structured inspection report.

[0007] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the steps of acquiring multi-view original RGB image sequences and performing preprocessing to generate standardized multi-view image sequences are as follows: The precision injection molded parts are scanned by an industrial camera array, and the intrinsic and extrinsic parameter matrices from multiple perspectives are recorded to generate a sequence of original RGB images from multiple perspectives. By using photometric consistency correction, the exposure differences and color temperature deviations between different viewpoints in the original RGB image sequence are eliminated, and a photometrically balanced RGB image sequence is generated. Based on the photometric equalization RGB image sequence, noise suppression is performed through multi-scale Gaussian filtering to smooth background noise and preserve high-frequency gradient information of defect edges, thereby generating a denoised and enhanced image sequence. The corresponding depth confidence map is obtained from the denoised and enhanced image sequence and associated with the two-dimensional texture information in the denoised and enhanced image sequence to generate a standardized multi-view image sequence.

[0008] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the steps of initializing and reconstructing a standardized multi-view image sequence to generate a high-fidelity three-dimensional geometric structure are as follows: Extract extrinsic parameter matrices and sparse point clouds from standardized multi-view image sequences to construct an initial geometric representation of the scene; The initial scene geometric representation is densely reconstructed using a multi-view stereo vision algorithm, and dense point cloud data is generated by combining geometric constraints. The Poisson surface reconstruction algorithm is used to extract isosurfaces from dense point cloud data to form high-density grid data of the surface of precision injection molded parts, and restore the subtle geometric features of the local structure of the precision injection molded parts to construct a high-fidelity three-dimensional geometric structure.

[0009] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the steps of identifying areas that may lead to false detections and generating a set of view occlusion masks are as follows: Extract the normal and curvature information of vertices from high-fidelity 3D geometric structures, and quantify the geometric abrupt change areas on the surface of precision injection molded parts to form a geometric feature enhancement map; Based on the geometric feature-enhanced map, the occlusion mask for each viewpoint is calculated using the visibility ray casting algorithm, and regions in the two-dimensional image that may lead to false detections are identified, generating a set of viewpoint occlusion masks.

[0010] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the steps of constructing an enhanced defect training dataset and a defect detection model based on a high-fidelity three-dimensional geometric structure are as follows: Virtual defect geometry is generated at random locations in high-fidelity 3D geometric grid data to simulate common defect morphologies on the production line and generate 3D geometry containing virtual defects. Based on a 3D geometric structure containing virtual defects, an enhanced defect training dataset is constructed by simulating light source reflection and material texture through rasterization rendering algorithm and combining it with standardized multi-view image sequences. Based on the convolutional neural network architecture, residual convolutional backbone layers are constructed using residual blocks, feature pyramid layers are constructed through bilinear interpolation, and detection heads are built through convolutional layers. A defect detection model is constructed by serially connecting the residual convolutional backbone layer, the feature pyramid layer, and the detection head through a data stream.

[0011] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the step of updating the defect detection model weights through backpropagation to generate the trained defect detection model includes the following steps: The enhanced defect training dataset is input into the defect detection model. Multi-scale feature maps are extracted from the standardized multi-view image sequence through the residual convolution backbone layer. The multi-scale feature maps are then fused using bilinear interpolation through the feature pyramid layer. Finally, the detection head outputs the training defect category probability in parallel using convolution operations. Based on the training defect category probabilities, the error is calculated using the cross-entropy loss function and the weights are updated through backpropagation to generate the trained defect detection model.

[0012] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the step of extracting a set of two-dimensional texture feature maps and a set of two-dimensional geometric prior feature maps from a standardized multi-view image sequence using a trained defect detection model is as follows: Based on the trained defect detection model, texture semantic features of images are extracted from standardized multi-view image sequences to generate a set of two-dimensional texture feature maps. The geometric feature enhancement map in the high-fidelity 3D geometric structure is projected onto the 2D texture feature map set to generate a 2D geometric prior feature map set.

[0013] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the steps for generating the defect detection set are as follows: A weighted fusion of a two-dimensional texture feature map set and a two-dimensional geometric prior feature map set is performed through a cross-modal attention fusion mechanism to suppress background interference and enhance the response of defect regions, thereby generating a multimodal fused feature map set. Based on the multimodal fusion feature map set and the viewpoint occlusion mask set, a defect detection set is generated by masking the predicted response of the occluded area in the detection head of the trained defect detection model.

[0014] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the step of selecting a hierarchical defect set from the defect detection set is as follows: The defect detection set is subjected to multiple random forward propagations through Monte Carlo random inactivation inference to calculate the variance distribution of the prediction results and generate a set of prediction uncertainties. From the set of prediction uncertainties, high-confidence defects and low-confidence suspicious defects are selected to generate a graded defect set.

[0015] As a preferred embodiment of the deep learning-based precision injection molded part defect detection method of the present invention, the steps of performing dimensional quantization, combining with a quantity judgment rule base, and generating a structured inspection report are as follows: The defect size of the graded defect set is quantified, and combined with the scale information of the high-fidelity three-dimensional geometric structure, the actual physical size and volume of the defect are obtained to form a set of defect physical parameters. Based on the set of physical parameters of defects and the set of graded defects, combined with the quality judgment rule base, pass / fail judgment labels and defect level labels are generated, and a structured inspection report is generated.

[0016] The beneficial effects of this invention are as follows: by constructing a high-fidelity three-dimensional geometric structure and combining it with image filtering, the fine geometric features of the surface are restored, thereby reducing false detections caused by structural occlusion; through a cross-modal attention fusion mechanism, a weighted fusion of two-dimensional texture and three-dimensional geometric prior information is achieved, and combined with deep learning model training, the accuracy of defect detection and the ability to quantify physical dimensions under complex structures are improved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a deep learning-based method for detecting defects in precision injection molded parts.

[0019] Figure 2 A flowchart for high-fidelity 3D geometry construction and viewpoint occlusion mask generation.

[0020] Figure 3A flowchart for generating structured inspection reports.

[0021] Figure 4 This is a graph showing the change in the percentage of structural occlusion area and the average false detection rate.

[0022] Figure 5 This is a graph showing the change in the overall quality score over time. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides a method for detecting defects in precision injection molded parts based on deep learning, including the following steps: S1. Acquire and preprocess original RGB image sequences from multiple perspectives to generate standardized multi-view image sequences.

[0027] By scanning precision injection molded parts with an industrial camera array and recording the intrinsic and extrinsic parameter matrices from multiple perspectives, a multi-view original RGB image sequence is generated.

[0028] Furthermore, the precision injection molded part is fixed at a rotating station, and images are simultaneously acquired by multiple surround industrial camera arrays (e.g., 50 mm focal length, 10 ms exposure time, and 8 cameras). The intrinsic parameter matrix (including focal length, principal point coordinates, and distortion coefficients) and extrinsic parameter matrix (including rotation matrix and translation vector) of each camera are recorded synchronously. The acquired RGB images are then bound to the corresponding matrices by timestamp to generate a multi-view original RGB image sequence with pose information.

[0029] It should be noted that the distortion coefficients are obtained using the Zhang Zhengyou calibration method and include radial distortion parameters and tangential distortion parameters. The exemplary value range is typically -0.5 to 0.5. By using photometric consistency correction, the exposure differences and color temperature deviations between different viewpoints in the original RGB image sequence are eliminated, and a photometrically balanced RGB image sequence is generated.

[0030] Furthermore, based on the original RGB image sequence from multiple perspectives, the average luminance value and chromaticity mean of each perspective image in the HSV color space are obtained. The average luminance value is the pixel mean of the V channel of each perspective image, and the average chromaticity value is the pixel mean of the S and H channels. The center perspective is selected as the reference benchmark, and the ratio of the average luminance value of the other perspectives to the reference benchmark is used as the luminance gain coefficient (e.g., 1.2). The difference between the chromaticity mean of the other perspectives and the reference benchmark is used as the chromaticity offset (e.g., 0.05). Based on the luminance gain coefficient and the chromaticity offset, a linear transformation formula is applied to the RGB channels of each pixel for numerical mapping, so that the luminance statistical distribution and color attribute values ​​of the multi-view images tend to be consistent, generating a luminance-equalized RGB image sequence.

[0031] Based on the photometrically equalized RGB image sequence, noise suppression is achieved through multi-scale Gaussian filtering, which smooths background noise and preserves high-frequency gradient information of defect edges, thereby generating a denoised and enhanced image sequence.

[0032] Furthermore, the input RGB image sequence is photometrically equalized, and Gaussian filtering is applied to the R, G, and B channels according to the corresponding detail preservation scale and noise suppression scale (e.g., the standard deviation of the detail preservation scale is 1.0 and the standard deviation of the noise suppression scale is 2.0). Through Gaussian convolution operations (with convolution kernel sizes of 5×5 and 7×7 respectively), a first-scale smoothed image and a second-scale smoothed image are obtained using a two-dimensional Gaussian function. The first-scale smoothed image and the second-scale smoothed image are weighted and fused to generate a preliminary smoothed image, which preserves the high-frequency gradient information of defect edges and suppresses background noise. The fused smoothed images of the R, G, and B channels are merged to generate a denoising and enhanced image sequence.

[0033] It should be noted that high-frequency gradient information refers to image information that retains the detailed changes in the edges of defects on the surface of precision injection molded parts through multi-scale Gaussian filtering, and is used to emphasize locally rapidly changing brightness or color features.

[0034] The corresponding depth confidence map is obtained from the denoised and enhanced image sequence and associated with the two-dimensional texture information in the denoised and enhanced image sequence to generate a standardized multi-view image sequence.

[0035] Furthermore, adjacent viewpoint image pairs are selected from the denoised and enhanced image sequence, and pixel disparity values ​​are obtained using binocular stereo vision triangulation. Outliers with pixel disparity values ​​greater than the disparity threshold are removed and converted into depth confidence maps. RGB channel data (grayscale values ​​of the red, green, and blue color channels of each pixel) are extracted from the denoised and enhanced image sequence as two-dimensional texture information. The depth values ​​in the depth confidence map (the physical distance from the pixel to the camera optical center, i.e., the origin of the camera coordinate system) are used as three-dimensional geometric information. The two-dimensional texture information and three-dimensional geometric information are bound according to the one-to-one correspondence of pixel coordinates and sorted by timestamp to generate a standardized multi-view image sequence.

[0036] It should be noted that binocular stereo vision triangulation obtains pixel disparity values ​​by matching corresponding pixels in adjacent viewpoint images, and uses triangular geometric relationships to solve for the depth of the pixel's corresponding three-dimensional spatial coordinates by combining intrinsic and extrinsic parameter matrices. The disparity threshold is set based on the camera baseline distance, focal length, and expected detection depth range. The minimum disparity setting is calculated by the focal length, baseline distance, and maximum detection depth. The exemplary value range is usually 0 to 256 (8-bit quantization) or 0 to 1024 (16-bit quantization). Two-dimensional texture information refers to the RGB channel grayscale value sequence extracted from the denoised and enhanced image sequence, which is used to reflect the texture pattern and color characteristics of the surface of the precision injection molded part.

[0037] S2. Initialize and reconstruct the standardized multi-view image sequence to generate a high-fidelity three-dimensional geometric structure, while identifying areas that may cause false detections and generating a set of view occlusion masks.

[0038] Extrinsic parameter matrices and sparse point clouds are extracted from standardized multi-view image sequences to construct an initial geometric representation of the scene.

[0039] Furthermore, based on standardized multi-view image sequences, the structure-of-motion (SOMO) algorithm is used to detect scale-invariant feature transformation (SMT) points. Nearest neighbor matching is used to match feature point pairs between adjacent views, eliminating mismatched pairs. Binocular stereo triangulation is then used to obtain the 3D coordinates of the feature points, forming a sparse point cloud. Based on the sparse point cloud, the rotation and translation parameters of the 3D coordinates of the sparse point cloud and the camera extrinsic parameter matrix are jointly adjusted using a reprojection error minimization objective function to optimize the camera pose and generate an optimized extrinsic parameter matrix. A 3D Cartesian coordinate system is constructed with the optical center of the first camera as the origin, and the sparse point cloud and the optimized extrinsic parameter matrix are transformed to a unified world coordinate system to generate an initial geometric representation of the scene.

[0040] It should be noted that the motion reconstruction algorithm is a computer vision method for recovering the structure of a three-dimensional scene and the camera motion trajectory from a multi-view two-dimensional image sequence. It achieves three-dimensional reconstruction through feature matching, binocular stereo vision triangulation, and optimization to minimize reprojection error.

[0041] The initial scene geometry is densely reconstructed using a multi-view stereo vision algorithm, and then dense point cloud data is generated by combining geometric constraints.

[0042] Furthermore, based on the initial scene geometric representation, a multi-view stereo vision algorithm is used to densely reconstruct the initial scene geometric representation (e.g., setting the number of depth planes to 64 layers). Specifically, the viewpoint with complete field of view coverage, optimal image quality, and located at the center of the camera array is selected as the reference viewpoint image. The homography matrix is ​​obtained using the optimized extrinsic parameter matrix. The reference viewpoint pixels are projected onto the image plane of the adjacent source viewpoints (other camera viewpoints adjacent to the reference viewpoint used to assist depth). Candidate depth values ​​are obtained by iteratively updating the depth plane index. The normalized cross-correlation value between the reference viewpoint and the source viewpoints is obtained as a photometric consistency measure, which is used to quantify the appearance similarity of the same spatial point under different viewpoints to evaluate the matching credibility. When the angle between the normal vectors (the angle between the surface normal vector and the camera's line of sight) is greater than the angle threshold, the candidate depth value is determined to be unreliable and is removed. Reliable depth values ​​are combined to form an initial depth map. The initial depth map is iteratively updated until the reprojection error converges to generate an optimized depth map. The optimized depth maps generated for each viewpoint are filtered and fused to remove depth abrupt change anomalies. The retained depth pixels are back-projected into three-dimensional space to generate dense point cloud data.

[0043] It should be noted that the multi-view stereo vision algorithm is a computer vision method that uses a sequence of two-dimensional images from multiple perspectives, combined with the camera's intrinsic and extrinsic parameter matrices, to solve for pixel-level depth values ​​through photometric consistency matching and geometric constraints, thereby achieving dense 3D reconstruction of a scene. The homography matrix is ​​a 3x3 transformation matrix used to describe the projection mapping relationship of planar points between the reference viewpoint and the source viewpoint. The included angle threshold is usually determined by statistically analyzing the distribution of the angle between the surface normal vector of the defect-free region and the camera's line of sight and selecting a high-confidence quantile (e.g., the 95th percentile) as the criterion. The exemplary value range is usually from 45 degrees to 75 degrees.

[0044] The Poisson surface reconstruction algorithm is used to extract isosurfaces from dense point cloud data to form high-density grid data of the surface of precision injection molded parts, and restore the subtle geometric features of the local structure of the precision injection molded parts to construct a high-fidelity three-dimensional geometric structure.

[0045] Furthermore, dense point cloud data is input, and an octree structure (e.g., depth 10) is constructed using the Poisson surface reconstruction algorithm. The point cloud normal vectors are obtained and uniformly oriented outward. The Poisson equation of the gradient field of the indicator function is solved, and the isosurface is extracted to form high-density mesh data of the precision injection molded part surface. For local structures of the precision injection molded part, a local refinement scale factor is set. Based on the local refinement scale factor, by reducing the size of the leaf nodes of the octree leaf and increasing the density of the mesh vertices, subtle geometric features (e.g., small edges, chamfers, and threads) are restored and isolated noise meshes are removed. The initial triangular mesh data is converted into watertight mesh data to construct a high-fidelity three-dimensional geometric structure.

[0046] It should be noted that the Poisson surface reconstruction algorithm is a 3D reconstruction method that converts dense point cloud data with normal vectors into a continuous closed triangular mesh surface by solving the Poisson equation of the gradient field of the indicator function. The local structure of the precision injection molded part refers to the tiny edges, chamfers and threaded structures on the surface of the precision injection molded part, which usually have complex geometric shapes with a diameter of less than 1 mm. The local refinement scale factor is set based on the ratio of the local structural feature size (such as diameter, width and height) of the precision injection molded part to the global average point spacing, and the value range is usually from 1.5 to 3.0.

[0047] The normal and curvature information of vertices are extracted from high-fidelity 3D geometry, and the geometric abrupt change regions on the surface of precision injection molded parts are quantified to form a geometric feature enhancement map.

[0048] Furthermore, each grid vertex in the high-fidelity 3D geometric structure is traversed, and the cross product operation is performed on the edge vector formed by the coordinate difference of the grid vertex to obtain the normal information (the X, Y, and Z component values ​​of the unit normal vector). The principal curvature value is calculated through the covariance matrix of the neighborhood points to obtain the curvature information (Gaussian curvature and mean curvature). Regions where the curvature gradient (the rate of change of the mean curvature value between adjacent vertices on the grid surface) exceeds the curvature change rate threshold are marked as geometric abrupt change regions. The XYZ components of the unit normal vector, Gaussian curvature, mean curvature, and the markings of geometric abrupt change regions are fused into five-channel feature data according to the vertex index and mapped to the two-dimensional UV parameter space to generate a geometric feature enhancement map.

[0049] It should be noted that the curvature change rate threshold is set based on the curvature distribution statistics of the normal surface geometry of the precision injection molded part, and is determined by a fixed quantile (e.g., the 95th percentile value) of the curvature gradient in the defect-free area. An exemplary value range is typically 0.3 to 1.0.

[0050] Based on the geometric feature-enhanced map, the occlusion mask for each viewpoint is calculated using the visibility ray casting algorithm, and regions in the two-dimensional image that may lead to false detections are identified, generating a set of viewpoint occlusion masks.

[0051] Furthermore, based on the geometric feature enhancement map and high-fidelity 3D geometry, a visibility ray projection algorithm is used to emit rays from the camera optical center to the grid vertices for each viewpoint, and the distance between the intersection points of the rays and the triangular facets in the grid is obtained. The minimum value is selected as the visibility depth. Surface areas with intersection distances greater than the visibility depth are projected onto the image plane and marked as occluded areas to generate a binary occlusion mask. Geometric abrupt change region markers in the geometric feature enhancement map are read, and the overlapping parts of the binary occlusion mask and the geometric abrupt change regions are marked as areas that may cause false detections. All viewpoints are traversed to generate a viewpoint occlusion mask set.

[0052] It should be noted that the visibility ray casting algorithm is a method of generating an occlusion mask by emitting rays from the camera's optical center to the 3D mesh and calculating the distance between the intersection points, and then determining the surface visibility based on the minimum depth value.

[0053] like Figure 4 The graphs showing the relationship between the proportion of structural occlusion area and the mean false detection rate are presented. The grouping variable of the curves is the octree depth of the Poisson surface reconstruction algorithm, and the control variable is whether or not a viewpoint occlusion mask set is used. The Poisson surface reconstruction algorithm uses an octree structure to convert dense point cloud data into high-density grid data to construct a high-fidelity 3D geometric structure. Changes in octree depth will affect the degree of restoration of subtle geometric features. The viewpoint occlusion mask set identifies areas in the 2D image that may lead to false detections based on geometric feature enhancement maps and visibility ray casting algorithms. Under the same structural occlusion area proportion, the mean false detection rate using the viewpoint occlusion mask set is lower than the mean false detection rate without the viewpoint occlusion mask set. The difference increases with the increase of the structural occlusion area proportion, reflecting the effect of reducing false detections caused by structural occlusion.

[0054] S3. Based on a high-fidelity 3D geometric structure, an enhanced defect training dataset and a defect detection model are constructed. The weights of the defect detection model are updated through backpropagation to generate the trained defect detection model.

[0055] Virtual defect geometry is generated at random locations in high-fidelity 3D geometric grid data to simulate common defect morphologies on the production line, thus generating a 3D geometric structure containing virtual defects.

[0056] Furthermore, a high-fidelity 3D geometric structure is input, and a random vertex on the mesh surface is selected as the defect center (e.g., the defect influence radius is 2 mm). Based on the Gaussian distribution standard deviation and neighborhood distance, the displacement vector is calculated using the Gaussian displacement formula, combined with the unit normal vector and the maximum displacement. This displacement vector is then added to the original mesh surface vertex to obtain the new coordinates of the mesh vertex after generating the virtual defect. The vertex is offset along the normal vector to form a depression or convexity, generating a virtual defect geometric structure. The preprocessed RGB channel data in the standardized multi-view image sequence is read and mapped to the deformation area (the set of mesh vertices within the defect influence radius with the randomly selected defect center vertex as the origin) through UV coordinates. The local color brightness is adjusted to simulate stains or color differences, generating a 3D geometric structure containing virtual defects.

[0057] The expression for the Gaussian displacement formula is: ; in, To generate new coordinates for the mesh vertices after generating virtual defects; The defect center is randomly selected in a high-fidelity three-dimensional geometric structure; Neighborhood distance refers to the spatial distance between a neighboring vertex and the center of the defect. The standard deviation of the Gaussian distribution determines the range of influence of virtual defects and the smoothness of their edges; The direction of the unit normal vector on the surface of the mesh vertex; This represents the maximum vertex offset at the defect center.

[0058] Based on a 3D geometric structure containing virtual defects, an enhanced defect training dataset is constructed by simulating light source reflection and material texture through a rasterization rendering algorithm and combining it with a standardized multi-view image sequence.

[0059] Furthermore, based on the 3D geometry containing virtual defects, light source parameters (e.g., position coordinates, intensity value 1.0, and color temperature 5500K) and material parameters (e.g., diffuse reflection coefficient 0.8, specular reflection coefficient 0.2, and roughness 0.3) are set. Using a rasterization rendering algorithm, the 3D vertex coordinates are converted to 2D screen coordinates according to the intrinsic and extrinsic parameter matrices. Occluded fragments are removed, leaving visible surfaces. Simultaneously, pixel shading operations are performed based on the light source and material parameters to obtain the illumination reflection value of each pixel, simulating light source reflection and material texture (referring to the albedo and roughness maps mapped onto the 3D mesh surface, used to characterize the color distribution and micro-geometric roughness of the object's surface), generating a rendered image sequence. This rendered image sequence is then fused with a standardized multi-view image sequence, aligned by view index and timestamp. The true RGB color data in the standardized multi-view image sequence is retained, and the illumination reflection changes and color deviations of the virtual defects are superimposed to construct an enhanced defect training dataset.

[0060] It should be noted that the rasterization rendering algorithm is a real-time rendering method that maps the vertex and face data of a 3D geometric model to a 2D image plane through projection transformation, and performs visibility determination and pixel color shading. The enhanced defect training dataset randomly generates virtual defects on a high-fidelity 3D geometric structure and simulates lighting and material changes, enabling the model to learn richer and more realistic defect morphologies. Compared with conventional training data, it can significantly improve the accuracy and robustness of defect recognition under complex structures.

[0061] Based on the convolutional neural network architecture, residual convolutional backbone layers are constructed using residual blocks, feature pyramid layers are constructed through bilinear interpolation, and detection heads are built through convolutional layers.

[0062] Furthermore, based on the convolutional neural network architecture, a residual convolutional backbone layer containing multiple stages is constructed (for example, each stage contains 3 residual blocks, with output channels of 64, 128, 256, and 512 respectively). Using bilinear interpolation, the deep feature layers output by the residual convolutional backbone layer are upsampled to generate shallow feature layer sizes. Different levels of feature layers are fused through element-wise addition to construct a feature pyramid layer. A detection head containing multiple convolutional operations is connected to the output of the feature pyramid layer, for example, the convolutional kernel size is set to 3×3, and the number of output channels is set to the number of defect categories.

[0063] A defect detection model is constructed by serially connecting the residual convolutional backbone layer, the feature pyramid layer, and the detection head through a data stream.

[0064] Furthermore, the output feature map of the residual convolution backbone layer is aligned and connected with the input port of the feature pyramid layer according to the number of channels (e.g., 512 channels). The multi-scale fusion output of the feature pyramid layer is matched and connected with the convolution input of the detection head according to the spatial resolution (e.g., 1 / 16 of the original image size), forming a complete forward propagation path from the original image input to the defect category probability output, thus constructing a defect detection model.

[0065] The enhanced defect training dataset is input into the defect detection model. Multi-scale feature maps are extracted from the standardized multi-view image sequence through the residual convolution backbone layer. The multi-scale feature maps are then fused using bilinear interpolation through the feature pyramid layer. Finally, the detection head uses convolution operations to output the training defect category probability in parallel.

[0066] Furthermore, the enhanced defect training dataset is input into the defect detection model (e.g., input image size is 512×512 pixels), and multi-scale feature maps (e.g., feature maps with 64, 128, 256 and 512 channels respectively) are extracted through residual convolution backbone layers. The multi-scale feature maps are input into the feature pyramid layer, and bilinear interpolation is used to upsample the deep features of the multi-scale feature maps to the shallow feature size. Multi-scale information is then fused through element-wise addition to generate a fused feature map. The fused feature map is input into the detection head, and convolution operations are performed to map it to the defect category dimension (e.g., 3 layers of convolution kernels with a size of 3×3). The training defect category probability output is then output, and the number of output channels is equal to the number of defect categories (e.g., dents, bumps, cracks, bubbles, scratches, stains, color differences, missing materials and flash).

[0067] Based on the training defect category probabilities, the error is calculated using the cross-entropy loss function and the weights are updated through backpropagation to generate the trained defect detection model.

[0068] Furthermore, the training defect category probabilities and the real defect labels (defect category annotation information corresponding to each training sample in the enhanced defect training dataset) are input into the cross-entropy loss function (for example, setting the initial learning rate to 0.001 and the weight decay coefficient to 0.0005) to obtain the classification error value. Based on the classification error value, the gradient values ​​of the residual convolution backbone layer, feature pyramid layer and convolution kernel in the detection head are obtained through the backpropagation algorithm, and the weight parameters of the convolution kernel are updated using the stochastic gradient descent update formula. The training is iterated until the loss value reaches below the convergence threshold, and the final weight parameters are saved into the defect detection model architecture to generate the trained defect detection model.

[0069] It should be noted that the convergence threshold is set based on the size of the training set, the number of defect categories, and the model complexity. It is set as a fixed percentage (e.g., 5%) of the loss value in the validation set (an independent subset of samples divided from the enhanced defect training dataset) during the early stages of training, and the value range is usually from 0.001 to 0.1.

[0070] The stochastic gradient descent update formula is: ; in, To update the weight parameters of the convolutional kernel; To update the weight parameters of the previous convolutional kernel; The initial learning rate is used to control the step size of weight parameter updates; The gradient value of the convolution kernel; This is the weight decay coefficient, set based on the need to prevent overfitting and the magnitude of the parameters. An exemplary value range is typically 1 × 10⁻⁶. -5 Up to 1×10 -3 .

[0071] S4. Using the trained defect detection model, extract a set of two-dimensional texture feature maps and a set of two-dimensional geometric prior feature maps from the standardized multi-view image sequence, and combine them with the viewpoint occlusion mask set to generate a defect detection set.

[0072] Based on the trained defect detection model, texture semantic features of images are extracted from standardized multi-view image sequences to generate a set of two-dimensional texture feature maps.

[0073] Furthermore, the standardized multi-view image sequence is input into the trained defect detection model, and multi-stage convolution operations (e.g., fourth-order, with output channels of 64, 128, 256 and 512 respectively) are performed through the residual convolution backbone layer to output multi-scale feature maps; texture semantic features of the defect region are extracted from the multi-scale feature maps, and the multi-scale feature maps are sorted by view index and timestamp to generate a two-dimensional texture feature map set.

[0074] It should be noted that texture semantic features refer to the feature maps output by the intermediate layers of a convolutional neural network, which contain high-level visual information such as edges, corners, and color patterns.

[0075] The geometric feature enhancement map in the high-fidelity 3D geometric structure is projected onto the 2D texture feature map set to generate a 2D geometric prior feature map set.

[0076] Furthermore, the five-channel feature data (including the XYZ components of the unit normal vector, Gaussian curvature, and mean curvature) in the geometric feature enhancement map of the high-fidelity 3D geometric structure are mapped to the 3D mesh vertices according to UV coordinates. The 3D vertex coordinates are converted into 2D image pixel coordinates using the camera projection matrix. The 2D image pixel coordinates are used as the target position index for filling the five-channel feature data to fill the corresponding pixel positions in the 2D texture feature map set, maintaining a one-to-one correspondence between pixel coordinates (e.g., spatial resolution of 1 / 16 of the input image size). Uncovered pixels are filled by bilinear interpolation. The coordinates and feature values ​​of the four filled pixels around the uncovered pixel are selected, and a linear combination of the five-channel feature values ​​is calculated by weighting by the inverse distance. This value is then assigned to the uncovered pixel to complete the feature filling, generating a 2D geometric prior feature map set with the same spatial resolution as the 2D texture feature map set.

[0077] It should be noted that the camera projection matrix refers to the mathematical transformation matrix obtained by combining the camera intrinsic parameter matrix and the extrinsic parameter matrix, which is used to convert the three-dimensional vertex coordinates into two-dimensional image pixel coordinates.

[0078] By employing a cross-modal attention fusion mechanism, a weighted fusion of a set of two-dimensional texture feature maps and a set of two-dimensional geometric prior feature maps is performed to suppress background interference and enhance the response of defective regions, thereby generating a set of multimodal fused feature maps.

[0079] Furthermore, based on the two-dimensional texture feature map set and the two-dimensional geometric prior feature map set, a cross-modal attention fusion mechanism (e.g., setting the convolution kernel size to 3×3) is used to convert the two-dimensional texture feature map set into a query vector and the two-dimensional geometric prior feature map set into a key vector and a value vector. The dot product similarity between the query vector and the key vector is obtained to obtain the attention weight matrix. The attention weight matrix is ​​normalized using the Softmax function. The normalized attention weight matrix is ​​combined with the value vector to obtain the weighted features. The weighted features are added element-wise to the two-dimensional texture feature map set to suppress background interference and enhance the response of defective regions, generating a multimodal fusion feature map set.

[0080] It should be noted that the defect area refers to the pixel area on the surface of the precision injection molded part that has common defects in the production line (including dents, bulges, cracks, bubbles, scratches, stains, color differences, missing material and flash). Based on the multimodal fusion feature map set and the viewpoint occlusion mask set, a defect detection set is generated by masking the predicted response of the occluded area in the detection head of the trained defect detection model.

[0081] Furthermore, the multimodal fusion feature map set and the viewpoint occlusion mask set are input into the detection head of the trained defect detection model. The detection head performs convolution operations to output an initial defect category probability map. The binary occlusion mask in the viewpoint occlusion mask set is aligned with the initial defect category probability map according to the pixel position. For example, the mask value of the occluded area is set to 0 and the mask value of the visible area is set to 1. The initial defect category probability map is multiplied element-wise with the binary occlusion mask to mask the predicted response of the occluded area. The defect category probability, bounding box coordinates (position and size parameters of the defect area in the image), and defect category probability confidence of the remaining visible area are extracted to generate a defect detection set.

[0082] It should be noted that the predicted response of the masked area refers to setting the probability value of the corresponding masked area in the initial defect category probability map to zero.

[0083] S5. Select a graded defect set from the defect detection set, quantify its size, and generate a structured inspection report by combining it with the quantity judgment rule base.

[0084] The defect detection set is subjected to multiple random forward propagations through Monte Carlo random inactivation inference to calculate the variance distribution of the prediction results and generate a set of prediction uncertainties.

[0085] Furthermore, the defect detection set is input into the trained defect detection model, and the random deactivation layer is kept active. A random forward propagation operation (e.g., 50 times) is performed through Monte Carlo random deactivation inference. In each forward propagation, neurons are randomly discarded (e.g., 10% of neurons are discarded), and the defect category probability and bounding box coordinates of each propagation output are obtained. Based on the defect category probability and bounding box coordinates, the variance value of the defect category probability and the variance value of the bounding box coordinates at each pixel position are obtained to form a variance distribution. The variance distribution is associated with the corresponding defect detection set to generate a set of prediction uncertainties containing the probability fluctuation range and the position fluctuation range, which serves as the input data for the hierarchical defect set.

[0086] It should be noted that the random deactivation layer refers to the network layer in the defect detection model after training, which is used to randomly discard neurons with probability during the propagation process to prevent overfitting during the training phase and to maintain activation during the Monte Carlo random deactivation inference phase to generate randomness.

[0087] From the set of prediction uncertainties, high-confidence defects and low-confidence suspicious defects are selected to generate a graded defect set.

[0088] Furthermore, based on the prediction uncertainty set, predictions with defect category probability variance values ​​below the probability variance threshold and bounding box coordinate variance values ​​below the location variance threshold are marked as high-confidence defects, while predictions with defect category probability variance values ​​and bounding box coordinate variance values ​​exceeding either threshold are marked as low-confidence suspicious defects. The marking results are sorted according to confidence level and the defect category probability and bounding box coordinates are bound together to generate a graded defect set.

[0089] It should be noted that the probability variance threshold is set based on the fixed percentile (e.g., 95th percentile) of the distribution statistics of the probability variance values ​​of defect categories in the set of prediction uncertainty, and the exemplary value range is usually 0.005 to 0.02; the location variance threshold is determined by the upper quantile (e.g., 95th percentile) of the statistical range of location fluctuations of normal detection samples, and the exemplary value range is usually 0.3 pixels to 1.0 pixels.

[0090] The defect size of the graded defect set is quantified, and combined with the scale information of the high-fidelity three-dimensional geometric structure, the actual physical size and volume of the defect are obtained to form a set of defect physical parameters.

[0091] Furthermore, the input is a set of graded defects. The bounding box coordinates and pixel count of each graded defect item are read. Combined with the scale information of the high-fidelity 3D geometry (e.g., 0.05 mm per pixel), the pixel size is converted into physical size values ​​(defect length, width, and height values). The point cloud density and 3D coordinate distribution of the pixels in the defect area are statistically analyzed, and the volume value of the space occupied by the defect is solved by the voxel integration method. The defect category, physical size value, and volume value of the space occupied by the defect are bound by the defect index to form a set of defect physical parameters.

[0092] It should be noted that the scale information refers to the conversion relationship between pixel units and physical length units in a high-fidelity 3D geometric structure, which is derived from the focal length parameter in the camera intrinsic parameter matrix and the triangulation geometry in the dense point cloud reconstruction process; the voxel integration method is a method of calculating the volume of space occupied by the defect by multiplying the number of occupied voxels by the volume of a single voxel by the integral.

[0093] Based on the set of physical parameters of defects and the set of graded defects, combined with the quality judgment rule base, pass / fail judgment labels and defect level labels are generated, and a structured inspection report is generated.

[0094] Furthermore, based on the set of physical parameters of defects and the set of graded defects, the weight coefficients (e.g., 0.4, 0.3, and 0.3) and the boundary values ​​of the scoring intervals in the quality judgment rule base are read; the physical size value, volume value, and probability confidence of defect category are weighted and fused according to the weight coefficients to calculate the comprehensive quality score; the pass / fail judgment label and defect level label are mapped according to the range in which the comprehensive quality score falls, for example, a score higher than 0.75 is mapped to a pass and a minor level, a score between 0.5 and 0.75 is mapped to a pass and a moderate level, and a score lower than, for example, 0.5 is mapped to a fail and a severe level; a structured inspection report is generated by binding the defect category, physical parameter value, comprehensive quality score, and judgment label according to the defect index.

[0095] It should be noted that the quality judgment rule base is constructed by statistically analyzing the correlation between the physical parameter distribution of historical qualified samples and defect levels, and setting weight coefficients and scoring mapping rules. The scoring interval boundary values ​​are set based on the statistical distribution of the comprehensive quality score in defect-free samples and typical defect samples. They are set by drawing a histogram of the scoring distribution and selecting the quantile point with the best category separation as the boundary value. The exemplary value range is usually 0.3 to 0.9. The exemplary setting of the scoring interval boundary value 1 is 0.5 and the scoring interval boundary value 2 is 0.75 as two dividing points.

[0096] like Figure 5The graph showing the overall quality score changes over time is presented in the main graph, which shows the changes in the overall quality score of the control path and the overall quality score of the experimental path over time. The graph is also superimposed with the boundary value 1 and the boundary value 2 of the scoring interval to show the interval mapping relationship between the overall quality score and the pass / fail label and the defect level label. The zoomed-in window captures a local interval of the time sequence to observe the detailed differences between the two overall quality score curves near the boundary.

[0097] The experimental path adopts the complete technical solution proposed in this invention, which specifically includes: constructing a high-fidelity three-dimensional geometric structure and generating a set of view occlusion masks; weightedly fusing the set of two-dimensional texture feature maps and the set of two-dimensional geometric prior feature maps through a cross-modal attention fusion mechanism; extracting multimodal features using a trained defect detection model; and combining the set of view occlusion masks to mask the predicted response of the occluded area; and finally generating a hierarchical defect set and quantizing its size through Monte Carlo random inactivation inference.

[0098] The comparison path uses conventional deep learning methods, specifically: it only uses two-dimensional image sequences for defect detection, does not construct three-dimensional geometric structures, does not generate a set of viewpoint occlusion masks, does not adopt cross-modal attention fusion mechanisms, and only relies on single texture features for defect recognition and classification.

[0099] Feature peak 1, feature peak 2, and feature peak 3 are local peak points of the comprehensive quality score of the experimental path within the magnified window. The peak point indicates that the set of physical parameters of the defect corresponding to the defect detection set and the probability confidence of the defect category are weighted by the quality judgment rule base to obtain a higher comprehensive quality score at that moment. The distance of the peak point relative to the boundary value 2 of the score interval reflects the stability of the qualified judgment label. The control path does not use the viewpoint occlusion mask set and cross-modal attention fusion mechanism, and the comprehensive quality score is more likely to drop and fluctuate near the boundary, reflecting the impact of occlusion false detection and interference from complex structural background on the judgment stability.

[0100] In summary, this invention achieves the restoration of subtle geometric features on the surface by constructing a high-fidelity three-dimensional geometric structure and combining it with image filtering, thereby reducing false detections caused by structural occlusion. Furthermore, by employing a cross-modal attention fusion mechanism, it achieves weighted fusion of two-dimensional texture and three-dimensional geometric prior information, and by combining it with deep learning model training, it improves the accuracy of defect detection and the ability to quantify physical dimensions under complex structures.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting defects in precision injection molded parts based on deep learning, characterized in that, include: Acquire raw RGB image sequences from multiple perspectives and preprocess them to generate standardized multi-view image sequences; The standardized multi-view image sequence is initialized and reconstructed to generate a high-fidelity three-dimensional geometric structure. At the same time, the region that may cause false detection is identified, and a set of view occlusion masks is generated. Based on a high-fidelity 3D geometric structure, an enhanced defect training dataset and a defect detection model are constructed. The weights of the defect detection model are updated through backpropagation to generate a trained defect detection model. The trained defect detection model extracts a set of two-dimensional texture feature maps and a set of two-dimensional geometric prior feature maps from a standardized multi-view image sequence, and combines them with a set of view occlusion masks to generate a defect detection set. A graded defect set is selected from the defect detection set, and its size is quantified. Combined with the quantity judgment rule base, a structured inspection report is generated.

2. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 1, characterized in that, The steps for acquiring and preprocessing multi-view raw RGB image sequences to generate standardized multi-view image sequences are as follows: The precision injection molded parts are scanned by an industrial camera array, and the intrinsic and extrinsic parameter matrices from multiple perspectives are recorded to generate a sequence of original RGB images from multiple perspectives. By using photometric consistency correction, the exposure differences and color temperature deviations between different viewpoints in the original RGB image sequence are eliminated, and a photometrically balanced RGB image sequence is generated. Based on the photometric equalization RGB image sequence, noise suppression is performed through multi-scale Gaussian filtering to smooth background noise and preserve high-frequency gradient information of defect edges, thereby generating a denoised and enhanced image sequence. The corresponding depth confidence map is obtained from the denoised and enhanced image sequence and associated with the two-dimensional texture information in the denoised and enhanced image sequence to generate a standardized multi-view image sequence.

3. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 2, characterized in that, The steps for initializing and reconstructing standardized multi-view image sequences to generate high-fidelity three-dimensional geometric structures are as follows: Extract extrinsic parameter matrices and sparse point clouds from standardized multi-view image sequences to construct an initial geometric representation of the scene; The initial scene geometric representation is densely reconstructed using a multi-view stereo vision algorithm, and dense point cloud data is generated by combining geometric constraints. The Poisson surface reconstruction algorithm is used to extract isosurfaces from dense point cloud data to form high-density grid data of the surface of precision injection molded parts, and restore the subtle geometric features of the local structure of the precision injection molded parts to construct a high-fidelity three-dimensional geometric structure.

4. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 3, characterized in that, The steps for identifying areas that may lead to false detections and generating a set of view occlusion masks are as follows: Extract the normal and curvature information of vertices from high-fidelity 3D geometric structures, and quantify the geometric abrupt change areas on the surface of precision injection molded parts to form a geometric feature enhancement map; Based on the geometric feature-enhanced map, the occlusion mask for each viewpoint is calculated using the visibility ray casting algorithm, and regions in the two-dimensional image that may lead to false detections are identified, generating a set of viewpoint occlusion masks.

5. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 3, characterized in that, The steps for constructing an enhanced defect training dataset and defect detection model based on a high-fidelity 3D geometric structure are as follows: Virtual defect geometry is generated at random locations in high-fidelity 3D geometric grid data to simulate common defect morphologies on the production line and generate 3D geometry containing virtual defects. Based on a 3D geometric structure containing virtual defects, an enhanced defect training dataset is constructed by simulating light source reflection and material texture through rasterization rendering algorithm and combining it with standardized multi-view image sequences. Based on the convolutional neural network architecture, residual convolutional backbone layers are constructed using residual blocks, feature pyramid layers are constructed through bilinear interpolation, and detection heads are built through convolutional layers. A defect detection model is constructed by serially connecting the residual convolutional backbone layer, the feature pyramid layer, and the detection head through a data stream.

6. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 5, characterized in that, The steps for updating the defect detection model weights through backpropagation to generate the trained defect detection model are as follows: The enhanced defect training dataset is input into the defect detection model. Multi-scale feature maps are extracted from the standardized multi-view image sequence through the residual convolution backbone layer. The multi-scale feature maps are then fused using bilinear interpolation through the feature pyramid layer. Finally, the detection head outputs the training defect category probability in parallel using convolution operations. Based on the training defect category probabilities, the error is calculated using the cross-entropy loss function and the weights are updated through backpropagation to generate the trained defect detection model.

7. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 1 or 4, characterized in that, The steps for extracting a set of two-dimensional texture feature maps and a set of two-dimensional geometric prior feature maps from a standardized multi-view image sequence using the trained defect detection model are as follows: Based on the trained defect detection model, texture semantic features of images are extracted from standardized multi-view image sequences to generate a set of two-dimensional texture feature maps. The geometric feature enhancement map in the high-fidelity 3D geometric structure is projected onto the 2D texture feature map set to generate a 2D geometric prior feature map set.

8. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 7, characterized in that, The steps for generating the defect detection set are as follows: A weighted fusion of a two-dimensional texture feature map set and a two-dimensional geometric prior feature map set is performed through a cross-modal attention fusion mechanism to suppress background interference and enhance the response of defect regions, thereby generating a multimodal fused feature map set. Based on the multimodal fusion feature map set and the viewpoint occlusion mask set, a defect detection set is generated by masking the predicted response of the occluded area in the detection head of the trained defect detection model.

9. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 1 or 8, characterized in that, The steps for selecting the graded defect set from the defect detection set are as follows: The defect detection set is subjected to multiple random forward propagations through Monte Carlo random inactivation inference to calculate the variance distribution of the prediction results and generate a set of prediction uncertainties. From the set of prediction uncertainties, high-confidence defects and low-confidence suspicious defects are selected to generate a graded defect set.

10. The method for detecting defects in precision injection molded parts based on deep learning as described in claim 9, characterized in that, The steps for quantifying dimensions, combining them with a quantity determination rule base, and generating a structured inspection report are as follows: The defect size of the graded defect set is quantified, and combined with the scale information of the high-fidelity three-dimensional geometric structure, the actual physical size and volume of the defect are obtained to form a set of defect physical parameters. Based on the set of physical parameters of defects and the set of graded defects, combined with the quality judgment rule base, pass / fail judgment labels and defect level labels are generated, and a structured inspection report is generated.

Images