Jade defect intelligent detection method and system based on machine vision and deep learning

Abstract

The invention relates to the technical field of computer vision, and discloses a jade defect intelligent detection method and system based on machine vision and deep learning, and the method comprises the following steps: S1, based on a high-resolution industrial camera and a laser three-dimensional scanner, adopting a multi-mode synchronous collection strategy, and rotating a jade sample through a precise motion control system, a jade surface high-resolution two-dimensional color image and high-precision three-dimensional point cloud data are respectively obtained, and a jade multi-mode original data set is generated. A high-resolution two-dimensional color image and high-precision three-dimensional point cloud data are integrated through a multi-modal synchronous acquisition strategy, and multi-dimensional feature expression under unified coordinates is constructed, so that the limitation of a single data source is effectively overcome; an image registration algorithm and a feature pyramid network are combined with a point cloud network to perform multi-modal feature fusion, complementarity of color texture and geometric morphology information is enhanced, and image quality is optimized based on adaptive histogram equalization and non-local mean filtering.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method and system for intelligent detection of defects in jade based on machine vision and deep learning. Background Technology

[0002] Computer vision is a core branch of artificial intelligence, dedicated to enabling machines to acquire, process, analyze, and understand information in digital images and videos through algorithms and models, thereby simulating or even surpassing human visual perception capabilities. This field deeply integrates technologies such as image processing, pattern recognition, machine learning, and deep learning. Core tasks include object detection, image segmentation, feature extraction, and 3D reconstruction. Among them, the intelligent detection method and system for jade defects based on machine vision and deep learning refers to an intelligent solution that comprehensively utilizes high-resolution imaging equipment, 3D scanning technology, and deep learning algorithms to automatically identify, locate, classify, and quantify various defects (such as cracks, impurities, and blemishes) in jade materials.

[0003] Existing technologies typically rely on a single imaging modality, such as using only two-dimensional image analysis, lacking support from three-dimensional geometric information, resulting in insufficient detection capabilities for internal defects or surface deformations in jade. Data processing workflows are often fragmented and unintegrated, with image and point cloud data failing to be effectively registered and fused, leading to one-sided and inconsistent feature extraction. Deep learning models have relatively simple structures, often employing single-path processing, failing to fully utilize the complementary advantages of multimodal data, and the lack of attention mechanisms weakens the model's ability to focus on key areas. Defect quantification methods are mostly based on two-dimensional parameters, such as area or perimeter, lacking precise measurements in three-dimensional space, such as depth or volume, resulting in incomplete evaluation results. The report generation process often lacks interpretable elements, such as heatmap visualization, making the decision-making process too opaque and relying on human experience for review and verification. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent detection method and system for jade defects based on machine vision and deep learning. This solves the problems of existing technologies that typically rely on a single imaging mode, such as using only two-dimensional image analysis, lacking support from three-dimensional geometric information, resulting in insufficient detection capabilities for internal defects or surface deformations in jade; and data processing flows that are often fragmented and unintegrated, with image and point cloud data failing to be effectively registered and fused, leading to one-sided and inconsistent feature extraction.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent detection method for jade defects based on machine vision and deep learning, comprising the following steps: S1: Based on a high-resolution industrial camera and a laser 3D scanner, a multimodal synchronous acquisition strategy is adopted, and the jade sample is rotated through a precision motion control system to obtain high-resolution two-dimensional color images and high-precision three-dimensional point cloud data of the jade surface, thereby generating a multimodal raw dataset of jade. S2: Based on the jade multimodal raw dataset generated by S1, the two-dimensional image and the three-dimensional point cloud are aligned at the pixel level, and color and texture features are extracted from the two-dimensional image and geometric morphology features are extracted from the three-dimensional point cloud. A multi-dimensional feature representation under a unified coordinate system is constructed to generate a registration and fusion multimodal feature map. S3: Based on the registration and fusion multimodal feature map generated by S2, the two-dimensional image data is enhanced and denoised, and the three-dimensional point cloud data is smoothed and simplified to generate an optimized multimodal data cube. S4: Based on the optimized multimodal data cube generated by S3, it processes two-dimensional image information to segment surface defects, processes three-dimensional geometric information to detect deformation defects, integrates dual-path information, and generates a binary segmentation mask for jade defects and a preliminary classification confidence matrix. S5: Based on the binary segmentation mask of jade defects generated by S4, the mask is mapped back to the three-dimensional spatial coordinates of the multimodal feature map generated by S2 through registration and fusion. Each defect region is isolated, and the bounding box size, average depth, and point density geometric parameters of the point cloud are calculated. Three-dimensional quantization measurement is performed to generate a set of three-dimensional quantization parameters for defects. S6: Based on the preliminary classification confidence matrix generated in S4 and the three-dimensional quantization parameter set of defects generated in S5, the defect type is finally determined, a heat map of the highlighted defect area is generated, and it is overlaid with the original image to automatically generate a structured inspection report, generating a smart inspection report of jade defects with heat map.

[0006] Preferably, generating the jade multimodal raw dataset based on S1 includes the following steps: S101: Based on a high-precision electronically controlled rotary table and a trigger synchronization controller, the industrial area array color camera is triggered at a specific angle position, and a global shutter exposure mode is used to acquire a sequence of two-dimensional color images of the surface of jade samples from multiple angles, generating a multi-angle two-dimensional image sequence of jade. S102: Based on the synchronous trigger signal of S101, the laser 3D scanner is controlled to synchronously scan the surface of the jade sample at the same angular position. The 3D contour information of the surface is obtained by laser triangulation and a multi-angle 3D point cloud sequence of jade is generated. S103: Based on the jade multi-angle two-dimensional image sequence generated by S101 and the jade multi-angle three-dimensional point cloud sequence generated by S102, a timestamp alignment algorithm is used to associate all data to the same rotation angle coordinate system and store them in the database to generate the jade multimodal original dataset.

[0007] Preferably, the generation, registration, and fusion of multimodal feature maps based on S2 includes the following steps: S201: Based on the single set of data in the original multimodal jade dataset generated by S1, the scale-invariant feature transformation algorithm is used to extract key points and feature descriptors from the two-dimensional image and its corresponding three-dimensional point cloud projection image, respectively, to generate initial feature point pairs for the two-dimensional and three-dimensional modes. S202: Based on the initial feature point pairs generated in S201, the random sampling consensus algorithm is used to remove erroneous matching point pairs, and the optimal homography transformation matrix is ​​calculated to complete the accurate registration of two-dimensional image pixels and three-dimensional point cloud coordinates, generating a fine registration parameter matrix. S203: Based on the fine registration parameter matrix generated by S202, perform perspective transformation on the original two-dimensional image to make it completely aligned with the three-dimensional point cloud data in space, and generate the registered two-dimensional image. S204: Based on the registered 2D image and its corresponding 3D point cloud generated by S203, a dual-branch convolutional neural network is used to extract deep features respectively. The extracted color texture features and geometric morphology features are then fused in the channel dimension through a feature splicing layer to generate a registered and fused multimodal feature map.

[0008] Preferably, generating the optimized multimodal data cube based on S3 includes the following steps: S301: Based on the registered two-dimensional image generated by S2, the contrast-limited adaptive histogram equalization algorithm is used to enhance the detail information in the low-contrast area of ​​the image, highlight the potential defect features, and generate a contrast-enhanced two-dimensional image. S302: Based on the contrast-enhanced two-dimensional image generated by S301, a non-local mean filtering algorithm is used for image denoising. While smoothing the background, the edge information of the defects is preserved to the maximum extent, and a denoised two-dimensional image is generated. S303: Based on the 3D point cloud data generated by S1, the voxel grid downsampling algorithm is used to reduce the point cloud density and reduce the amount of data while maintaining the overall shape of the point cloud, thus generating a downsampled 3D point cloud. S304: Based on the downsampled 3D point cloud generated by S303, the moving least squares method is used to resample and smooth the point cloud surface to eliminate noise points generated during the scanning process and generate a smoothed 3D point cloud. S305: Based on the denoised 2D image generated by S302 and the smoothed 3D point cloud generated by S304, the two are recombined into a data unit according to the registration relationship of S2 to generate an optimized multimodal data cube.

[0009] Preferably, the generation of a binary segmentation mask for jade defects and a preliminary classification confidence matrix based on S4 includes the following steps: S401: Based on the optimized multimodal data cube generated by S3, it is input into a fully convolutional network with an encoder-decoder structure; the encoder part is constructed using depthwise separable convolutions to extract high-level semantic features hierarchically from the fused data and generate high-level multimodal semantic feature maps; S402: Based on the high-level multimodal semantic feature map generated by S401, transposed convolution is used in the decoder part for upsampling to gradually restore the spatial resolution of the feature map. Features from the medium-resolution layer of the encoder are fused through skip connections to accurately locate the defect boundary and generate a high-resolution defect segmentation score map. S403: Based on the high-resolution defect segmentation score map generated by S402, the Argmax function or thresholding is used to assign a category label to each pixel, generate a binary segmentation result, and generate a binary segmentation mask for jade defects. S404: At the end of the encoder of S401, a global average pooling layer and a softmax classifier are connected in parallel. Based on the extracted global features, the types of defects in the image are initially predicted, and an initial classification confidence matrix is ​​generated.

[0010] Preferably, the generation of a three-dimensional quantization parameter set for defects based on S5 includes the following steps: S501: Based on the jade defect binary segmentation mask generated by S4, the connected component labeling algorithm is used to identify and distinguish the independent defect regions in the mask, assign a unique identifier to each defect, and generate an independent defect region identifier map. S502: Based on the independent defect region identification map generated by S501, use it as an index to extract the 3D point cloud subset corresponding to each defect region from the optimized multimodal data cube generated by S3, and generate an independent defect 3D point cloud cluster. S503: For each independent defect 3D point cloud cluster generated by S502, the principal component analysis algorithm is used to calculate its direction and size in 3D space, and its minimum outer envelope cuboid is constructed to generate the defect 3D size and direction vector. S504: Based on the independent defect 3D point cloud clusters generated by S502, calculate the average depth and point density of the point cloud clusters, and generate defect depth and density parameters; S505: Integrates the defect 3D size and direction vector generated by S503 with the defect depth and density parameters generated by S504 to form a structured quantization record for each defect, generating a set of defect 3D quantization parameters.

[0011] Preferably, the generation of a jade defect intelligent detection report with an interpretable heatmap based on S6 includes the following steps: S601: Based on the preliminary classification confidence matrix generated in S4 and the three-dimensional quantification parameter set of defects generated in S5, a rule-based expert system is used to make a decision on the final type of each defect and generate the final defect classification label. S602: Based on the S4 neural network model, it uses the gradient weighted class activation mapping algorithm to calculate the image region that contributes the most to the final classification result during the network decision-making process and generate a heat map, generating an interpretable heat map of the defect region. S603: Based on the interpretability heatmap of the defect area generated by S602, it is superimposed with the denoised two-dimensional image generated by S3 to intuitively display the focus area of ​​the model and generate a visualized detection result map. S604: Integrates the final defect classification labels generated by S601, the three-dimensional quantitative parameter set of defects generated by S5, and the visual detection result map generated by S603. According to the predefined template, it automatically generates a structured report containing defect type, quantity, size, location, and visual evidence, and generates a jade defect intelligent detection report with an interpretable heat map.

[0012] The intelligent jade defect detection system based on machine vision and deep learning includes the following modules; The data acquisition and preprocessing module, based on a high-resolution industrial camera and a laser 3D scanner, simultaneously acquires two-dimensional color image sequences and three-dimensional point cloud sequences of jade samples, optimizes the images and point clouds, and generates an optimized multimodal data cube. The multimodal fusion analysis module, based on the optimized multimodal data cube generated by the data acquisition and preprocessing module, processes two-dimensional image information, processes three-dimensional geometric information, integrates information, outputs pixel-level segmentation results and preliminary classification confidence of defects, and generates defect segmentation masks and preliminary confidence scores. The defect precise quantification module, based on the defect segmentation mask generated by the multimodal fusion analysis module, maps the mask back to the three-dimensional space of the optimized multimodal data cube, isolates the defects, calculates the three-dimensional spatial size and orientation of each isolated defect point cloud cluster, calculates the average depth, volume, and point density quantification parameters, and generates a set of three-dimensional quantification parameters for defects. The decision and report generation module, based on the preliminary confidence level generated by the multimodal fusion analysis module and the three-dimensional quantitative parameter set of defects generated by the defect precision quantification module, performs final arbitration on the defect type, generates a heat map of the highlighted defect area, merges it with the original image, automatically generates a structured report, and generates a jade defect intelligent detection report with an interpretable heat map.

[0013] This invention provides a method and system for intelligent detection of defects in jade based on machine vision and deep learning. It has the following beneficial effects: 1. This invention integrates high-resolution two-dimensional color images and high-precision three-dimensional point cloud data through a multimodal synchronous acquisition strategy, constructing a multi-dimensional feature representation under a unified coordinate system, effectively overcoming the limitations of a single data source. It employs an image registration algorithm combined with a feature pyramid network and a point cloud network for multimodal feature fusion, enhancing the complementarity of color texture and geometric morphology information. Image quality is optimized based on adaptive histogram equalization and non-local mean filtering, while voxel grid downsampling and moving least squares are used to simplify the point cloud, significantly improving data purity and processing efficiency. A multi-branch U-shaped network with an attention mechanism enables parallel processing and fusion of two-dimensional and three-dimensional information. An attention-guided module strengthens the focus on defect areas, achieving pixel-level accurate segmentation and preliminary classification. The segmentation results are mapped back to three-dimensional space, and connected component analysis and point cloud clustering algorithms are used to calculate bounding box size, average depth, and point density parameters, completing the three-dimensional quantitative measurement of defects. Finally, a multi-feature decision fusion algorithm combined with gradient-weighted class activation mapping generates a heatmap, outputting a structured inspection report. The overall process significantly improves detection accuracy, enhances automation, improves interpretability, and reduces labor costs.

[0014] 2. This invention integrates a high-resolution industrial camera and a laser 3D scanner through a multi-sensor synchronous triggering control algorithm to achieve synchronous acquisition of 2D color image sequences and 3D point cloud sequences, ensuring data spatiotemporal consistency and providing a solid foundation for subsequent analysis. It employs contrast-limited adaptive histogram equalization and non-local mean filtering algorithms to optimize image quality, enhance details in low-contrast areas, and suppress noise. Simultaneously, it utilizes voxel grid downsampling and moving least squares to simplify and smooth point cloud data, improving data purity and processing efficiency. A dual-branch deep learning network based on an attention mechanism processes 2D images and 3D geometric information in parallel, extracting color and texture features through a convolutional neural network. The cloud network extracts morphological features and integrates multimodal information with an attention-guided feature fusion layer to enhance the focusing ability of defect areas, achieving pixel-level accurate segmentation and preliminary classification. The segmentation results are mapped back to three-dimensional space, and a connected component analysis algorithm is used to isolate defect areas. Principal component analysis is combined with the algorithm to calculate three-dimensional dimensions, orientation, average depth, volume, and point density parameters, completing the accurate three-dimensional quantitative measurement of defects. Finally, a multi-feature decision fusion algorithm is used to arbitrate the defect type, and a gradient-weighted class activation mapping algorithm is applied to generate a heatmap. After being fused with the original image, a structured report is output. The overall process significantly improves detection accuracy, automation, and interpretability, while reducing false detection rate and manual dependence. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the main steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a detailed schematic diagram of S6 of the present invention; Figure 8 This is a system block diagram of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example: Please see the appendix Figure 1 - Appendix Figure 8 This invention provides an intelligent detection method for jade defects based on machine vision and deep learning, including the following steps; S1: Based on a high-resolution industrial camera and a laser 3D scanner, a multimodal synchronous acquisition strategy is adopted, and the jade sample is rotated through a precision motion control system to obtain high-resolution two-dimensional color images and high-precision three-dimensional point cloud data of the jade surface, thereby generating a multimodal raw dataset of jade. S2: Based on the jade multimodal raw dataset generated by S1, the two-dimensional image and the three-dimensional point cloud are aligned at the pixel level, and color and texture features are extracted from the two-dimensional image and geometric morphology features are extracted from the three-dimensional point cloud. A multi-dimensional feature representation under a unified coordinate system is constructed to generate a registration and fusion multimodal feature map. S3: Based on the registration and fusion multimodal feature map generated by S2, the two-dimensional image data is enhanced and denoised, and the three-dimensional point cloud data is smoothed and simplified to generate an optimized multimodal data cube. S4: Based on the optimized multimodal data cube generated by S3, it processes two-dimensional image information to segment surface defects, processes three-dimensional geometric information to detect deformation defects, integrates dual-path information, and generates a binary segmentation mask for jade defects and a preliminary classification confidence matrix. S5: Based on the binary segmentation mask of jade defects generated by S4, the mask is mapped back to the three-dimensional spatial coordinates of the multimodal feature map generated by S2 through registration and fusion. Each defect region is isolated, and the bounding box size, average depth, and point density geometric parameters of the point cloud are calculated. Three-dimensional quantization measurement is performed to generate a set of three-dimensional quantization parameters for defects. S6: Based on the preliminary classification confidence matrix generated in S4 and the three-dimensional quantization parameter set of defects generated in S5, the defect type is finally determined, a heat map of the highlighted defect area is generated, and it is overlaid with the original image to automatically generate a structured inspection report, generating a smart inspection report of jade defects with heat map.

[0018] The steps for generating the jade multimodal raw dataset based on S1 are as follows: S101: Based on a high-precision electronically controlled rotary table and a trigger synchronization controller, the industrial area array color camera is triggered at a specific angle position, and a global shutter exposure mode is used to acquire a sequence of two-dimensional color images of the surface of jade samples from multiple angles, generating a multi-angle two-dimensional image sequence of jade. Based on a high-precision electronically controlled rotary table and a trigger synchronization controller, control commands are generated according to the preset acquisition angle interval to trigger an industrial area array color camera to acquire images. With the help of a ring LED light source, the jade sample is fixed at the center of the rotary table and continuously rotates to acquire multiple two-dimensional color images. The images are numbered and stored in angular order to generate a multi-angle two-dimensional image sequence of the jade.

[0019] S102: Based on the synchronous trigger signal of S101, the laser 3D scanner is controlled to synchronously scan the surface of the jade sample at the same angular position. The 3D contour information of the surface is obtained by laser triangulation and a multi-angle 3D point cloud sequence of jade is generated. The above content describes the acquisition of three-dimensional contour information using laser triangulation, according to the formula... ; Calculate the depth value Z.

[0020] In the formula, f represents the lens focal length, L represents the baseline distance, and Δx represents the pixel offset.

[0021] The lens focal length f was obtained through scanner optical calibration and was set to 25 mm. The baseline distance L was obtained through scanner mechanical structure measurement and was set to 200 mm. The pixel offset Δx was calculated after capturing the laser line deformation image using a CMOS sensor. For example, if Δx is 50 pixels at a certain point, the Z value is calculated using the formula Z = (25 × 200) / 50 = 100 mm. This result indicates that the depth value Z is 100 mm, representing the vertical distance of a point on the jade surface relative to the scanner. This depth is used to construct 3D point cloud coordinates and further generate a multi-angle 3D point cloud sequence of the jade. S103: Based on the jade multi-angle two-dimensional image sequence generated by S101 and the jade multi-angle three-dimensional point cloud sequence generated by S102, a timestamp alignment algorithm is used to associate all data to the same rotation angle coordinate system and store them in the database to generate the jade multimodal original dataset.

[0022] Based on the jade multi-angle two-dimensional image sequence generated by S101 and the jade multi-angle three-dimensional point cloud sequence generated by S102, the timestamps of the images and point clouds are extracted, the timestamp difference is calculated, the images and point cloud data collected at the same angle are paired, all paired data are stored in a relational database, an index table is built with the angle value as the primary key, and the jade multimodal original dataset is generated.

[0023] The generation, registration, and fusion of multimodal feature maps based on S2 includes the following steps: S201: Based on the single set of data in the original multimodal jade dataset generated by S1, the scale-invariant feature transformation algorithm is used to extract key points and feature descriptors from the two-dimensional image and its corresponding three-dimensional point cloud projection image, respectively, to generate initial feature point pairs for the two-dimensional and three-dimensional modes. Based on a single set of data from the original multimodal jade dataset generated by S1, the two-dimensional image is blurred, a pyramid is constructed, extreme points are detected, neighboring pixel values ​​are compared, key points are located, gradient magnitude and direction are calculated, a direction histogram is calculated, the main direction is determined, the neighborhood of the key point is rotated, sub-regions are divided, gradient histograms are calculated, and feature descriptors are generated. The same process is performed on the three-dimensional point cloud projection image to extract key points, and the distance ratio method is used for matching to generate initial feature point pairs.

[0024] S202: Based on the initial feature point pairs generated in S201, the random sampling consensus algorithm is used to remove erroneous matching point pairs, and the optimal homography transformation matrix is ​​calculated to complete the accurate registration of two-dimensional image pixels and three-dimensional point cloud coordinates, generating a fine registration parameter matrix. The above content calculates the optimal homography transformation matrix according to the formula. ; Calculate the homography matrix H.

[0025] In the formula, X represents the source point coordinate matrix, Y represents the target point coordinate matrix, and H represents the homography transformation matrix.

[0026] The source point coordinate matrix X is obtained from the initial feature point pairs. For example, if four pairs of matching points are selected, the source point coordinates p1 are (100, 200, 1), p2 are (150, 250, 1), p3 are (200, 300, 1), and p4 are (250, 350, 1). The target point coordinates Y are obtained from the corresponding points, q1 are (110, 210, 1), q2 are (160, 260, 1), q3 are (210, 310, 1), and q4 are (260, 360, 1). The X matrix is ​​constructed as a 4×3 matrix with row vectors representing the homogeneous coordinates of p1, p2, p3, and p4. The Y matrix is ​​also a 4×3 matrix with row vectors representing the homogeneous coordinates of q1, q2, q3, and q4. The X^TX matrix is ​​then calculated to obtain a 3×3 matrix, for example, X^TX = [[sum(x_i^2), sum(x_iy_i), ...]. [sum(x_i)], [sum(x_i y_i), sum(y_i^2), sum(y_i)], [sum(x_i), sum(y_i),n]], n=4. Substitute into the numerical calculation X^TX to find the inverse matrix (X^TX){-1}, then calculate XT Y to obtain a 3×3 matrix. Finally, H= (X^TX)^{-1} X^TY. Substitute into the numerical calculation H matrix elements. The results show that the H matrix maps the source point to the target point, which is used for accurate registration and generates a fine registration parameter matrix.

[0027] S203: Based on the fine registration parameter matrix generated by S202, perform perspective transformation on the original two-dimensional image to make it completely aligned with the three-dimensional point cloud data in space, and generate the registered two-dimensional image. Based on the fine registration parameter matrix generated by S202, the coordinates of each pixel in the original two-dimensional image are transformed, the transformed coordinates are calculated, interpolation is performed, the new pixel value is calculated using the weighted average of the surrounding pixels, the boundary region is processed, the size of the transformed image is determined, and a registered image aligned with the three-dimensional point cloud space is generated.

[0028] S204: Based on the registered 2D image and its corresponding 3D point cloud generated by S203, a dual-branch convolutional neural network is used to extract deep features respectively. The extracted color texture features and geometric morphology features are then fused in the channel dimension through a feature splicing layer to generate a registered and fused multimodal feature map.

[0029] Based on the registered 2D image and its corresponding 3D point cloud generated by S203, the 2D image is input into the convolutional neural network branch, and the 3D point cloud is input into the point cloud network branch. Features are extracted, feature vectors are output, the feature vectors are concatenated, and normalization processing is performed to generate a multimodal feature map of registration fusion.

[0030] The generation of the optimized multimodal data cube based on S3 includes the following steps: S301: Based on the registered two-dimensional image generated by S2, the contrast-limited adaptive histogram equalization algorithm is used to enhance the detail information in the low-contrast area of ​​the image, highlight the potential defect features, and generate a contrast-enhanced two-dimensional image. Based on the registered 2D image generated by S2, local regions are divided, gray-level histograms are calculated, truncation is performed, pixel counts are redistributed, cumulative distribution function is calculated, gray-level transformation is performed, and a contrast-enhanced 2D image is generated.

[0031] S302: Based on the contrast-enhanced two-dimensional image generated by S301, a non-local mean filtering algorithm is used for image denoising. While smoothing the background, the edge information of the defects is preserved to the maximum extent, and a denoised two-dimensional image is generated. Based on the contrast-enhanced 2D image generated by S301, a search window and a similarity window are set, pixel similarity weights are calculated, denoised pixel values ​​are calculated, and a denoised 2D image is generated.

[0032] S303: Based on the 3D point cloud data generated by S1, the voxel grid downsampling algorithm is used to reduce the point cloud density and reduce the amount of data while maintaining the overall shape of the point cloud, thus generating a downsampled 3D point cloud. Based on the 3D point cloud data generated by S1, a voxel mesh is created, the point cloud space is divided, the centroid coordinates of points within each voxel are calculated, and the centroids are used to replace all points within the voxel to generate the downsampled 3D point cloud.

[0033] S304: Based on the downsampled 3D point cloud generated by S303, the moving least squares method is used to resample and smooth the point cloud surface to eliminate noise points generated during the scanning process and generate a smoothed 3D point cloud. In the above content, the moving least squares method is used to resample and smooth the point cloud surface, according to the formula... ; Calculate the coordinates of the smoothed point cloud.

[0034] In the formula, Represents the weighting function. Represents the original point coordinates. This represents the local fitting coefficient.

[0035] The weighting function uses a Gaussian function. Where h is the bandwidth parameter, set to twice the average spacing of the point cloud. For example, if the average spacing of the point cloud is 0.2mm, then h=0.4mm. For each point x, a set of points is selected within its neighborhood radius r=2h, and the local fitting coefficient a(x) is calculated by solving the least squares problem min sum(w_i(x)(p_i^T a(x) - f_i)^2), where f_i is a certain attribute value of point p_i, such as the z-coordinate. Substituting specific values, let the coordinates of point x be (10,10,10), and its neighborhood contain points p1(10.1,10.2,10.3) and p2(9.9,9.8,9.7), with weights w1=exp(-0.14 / 0.16)=0.42, w2=exp(-0.26 / 0.16)=0.20. Solving for a(x) minimizes the weighted sum of squares, we get a(x) = [a0, a1, a2]^T. We then calculate f(x) = p^T a(x). This result shows that f(x) is the smoothed coordinate value, which can be used to generate a smoothed 3D point cloud.

[0036] S305: Based on the denoised 2D image generated by S302 and the smoothed 3D point cloud generated by S304, the two are recombined into a data unit according to the registration relationship of S2 to generate an optimized multimodal data cube.

[0037] Based on the denoised 2D image generated by S302 and the smoothed 3D point cloud generated by S304, the 3D point coordinates corresponding to the pixels of the 2D image are calculated according to the registration relationship. The pixel values ​​are associated with the 3D coordinates to create a multimodal data cube and generate an optimized multimodal data cube.

[0038] The steps for generating a binary segmentation mask for jade defects and a preliminary classification confidence matrix based on S4 are as follows: S401: Based on the optimized multimodal data cube generated by S3, it is input into a fully convolutional network with an encoder-decoder structure; the encoder part is constructed using depthwise separable convolutions to extract high-level semantic features hierarchically from the fused data and generate high-level multimodal semantic feature maps; Based on the optimized multimodal data cube generated by S3, the input is a fully convolutional network with an encoder-decoder structure. The network is processed using depthwise separable convolutional layers, and pooling operations are performed to reduce the spatial size, extract texture and geometric features, and generate a high-level multimodal semantic feature map.

[0039] S402: Based on the high-level multimodal semantic feature map generated by S401, transposed convolution is used in the decoder part for upsampling to gradually restore the spatial resolution of the feature map. Features from the medium-resolution layer of the encoder are fused through skip connections to accurately locate the defect boundary and generate a high-resolution defect segmentation score map. Based on the high-level multimodal semantic feature map generated by S401, transposed convolution is used for upsampling to gradually reduce the number of output channels. Encoder features are fused through skip connections to gradually restore the spatial resolution to the original size, outputting the defect segmentation score for each pixel and generating a high-resolution defect segmentation score map.

[0040] S403: Based on the high-resolution defect segmentation score map generated by S402, the Argmax function or thresholding is used to assign a category label to each pixel, generate a binary segmentation result, and generate a binary segmentation mask for jade defects. Based on the high-resolution defect segmentation score map generated by S402, the score values ​​of each channel at each pixel position are compared, and the channel index with the highest score is selected as the category label to generate a binary segmentation result, thus generating a binary segmentation mask for jade defects.

[0041] S404: At the end of the encoder of S401, a global average pooling layer and a softmax classifier are connected in parallel. Based on the extracted global features, the types of defects in the image are initially predicted, and an initial classification confidence matrix is ​​generated.

[0042] The above content involves a preliminary prediction of the types of defects present in the image based on the extracted global features, according to the formula... Calculate the initial classification confidence level.

[0043] In the formula, Represents the score for the i-th category. Represents the number of defect types. This represents the confidence level of the i-th category.

[0044] Global features are extracted at the encoder end, with a feature map size of 64×64×512. After passing through a global average pooling layer, the average value of each channel is calculated, outputting a 512-dimensional feature vector. This vector is input to a fully connected layer, where the weight matrix has a size of 512×K, where K=3 represents three defect types, such as cracks, impurities, and normal. The fully connected layer outputs a score vector z = [z1, z2, z3]. The score is calculated through forward propagation of the network. For example, for z = [2.0, 1.0, 0.1], the exponential function is calculated: exp(2.0)=7.389, exp(1.0)=2.718, exp(0.1)=1.105. The summation is sum_exp = 7.389 + 2.718 + 1.105 = 11.212. The confidence score σ(z1) is calculated as 7.389 / 11.212 = 0.659, and σ(z2) is calculated as σ(z3). = 2.718 / 11.212 = 0.242, σ(z3) = 1.105 / 11.212 = 0.099, which indicates that the confidence vector is [0.659, 0.242, 0.099], representing the probability of each defect type, used to generate the preliminary classification confidence matrix.

[0045] The generation of a 3D quantization parameter set for defects based on S5 includes the following steps: S501: Based on the jade defect binary segmentation mask generated by S4, the connected component labeling algorithm is used to identify and distinguish the independent defect regions in the mask, assign a unique identifier to each defect, and generate an independent defect region identifier map. Based on the binary segmentation mask for jade defects generated by S4, each pixel of the image is traversed, the label values ​​of neighboring pixels are checked, new labels are assigned, equivalent label pairs are recorded, equivalent labels are processed, labels are reassigned, and an independent defect region identification map is generated.

[0046] S502: Based on the independent defect region identification map generated by S501, use it as an index to extract the 3D point cloud subset corresponding to each defect region from the optimized multimodal data cube generated by S3, and generate an independent defect 3D point cloud cluster. Based on the independent defect region identification map generated by S501, according to the correspondence between two-dimensional pixel coordinates and three-dimensional point clouds, for each pixel in the defect region, the corresponding three-dimensional point coordinates are found, the three-dimensional points are collected to form a subset of the point cloud, stored as an independent file, and an independent defect three-dimensional point cloud cluster is generated.

[0047] S503: For each independent defect 3D point cloud cluster generated by S502, the principal component analysis algorithm is used to calculate its direction and size in 3D space, and its minimum outer envelope cuboid is constructed to generate the defect 3D size and direction vector. In the above content, the principal component analysis algorithm is used to calculate its direction in three-dimensional space, according to the formula. ; Calculate the covariance matrix.

[0048] In the formula, This represents the coordinate vector of the i-th point in the point cloud. Represents the centroid coordinates of the point cloud. This represents the number of points in the point cloud.

[0049] The centroid coordinates μ of the point cloud are obtained by calculating the average of the coordinates of all points. For example, if the point cloud contains 100 points, μ = (sum(x) / 100, sum(y) / 100, sum(z) / 100). Assuming μ = (10.5, 20.3, 15.2), the difference vector between each point and the centroid is calculated. For example, if the coordinates of point p1 are (10.0, 20.0, 15.0), the difference vector is (-0.5, -0.3, -0.2). The outer product matrix of the difference vectors is calculated, with a dimension of 3×3. The outer product matrix of all points is summed and divided by the number of points n to obtain the covariance matrix C. The eigenvalues ​​and eigenvectors of C are calculated. The eigenvector corresponding to the largest eigenvalue is the principal direction vector of the point cloud. This result shows that the principal direction vector describes the main extension direction of the defect in three-dimensional space and is used to construct the minimum outer envelope cuboid and generate the three-dimensional size and direction vector of the defect.

[0050] S504: Based on the independent defect 3D point cloud clusters generated by S502, calculate the average depth and point density of the point cloud clusters, and generate defect depth and density parameters; Based on the independent defect 3D point cloud cluster generated by S502, the average value of the Z coordinates of all points is calculated, the bounding box volume of the point cloud is calculated, the point density is calculated, and the defect depth and density parameters are generated.

[0051] S505: Integrates the defect 3D size and direction vector generated by S503 with the defect depth and density parameters generated by S504 to form a structured quantization record for each defect, generating a set of defect 3D quantization parameters.

[0052] Integrate the defect 3D size and orientation vector generated by S503 with the defect depth and density parameters generated by S504, create a structured record for each defect, and store the record as a JSON format or database record to generate a defect 3D quantization parameter set.

[0053] Generating a smart jade defect detection report with an interpretable heatmap based on S6 includes the following steps: S601: Based on the preliminary classification confidence matrix generated in S4 and the three-dimensional quantification parameter set of defects generated in S5, a rule-based expert system is used to make a decision on the final type of each defect and generate the final defect classification label. Based on the preliminary classification confidence matrix generated in S4 and the defect 3D quantization parameter set generated in S5, the confidence matrix contains the category probability of each defect, such as crack 0.65, impurity 0.25, and normal 0.10. The 3D parameters include depth density in the dimension direction. An expert rule system is set: Rule 1: a crack is judged as a crack when the confidence is greater than 0.6 and the length is greater than 2 mm; Rule 2: an impurity is judged as a crack when the confidence is greater than 0.5 and the density is less than 100 points / cm. 3 When all confidence levels are below 0.5, a second judgment is made using three-dimensional parameters. The conditions of each rule are compared, and the category corresponding to the rule that meets the conditions is selected to generate the final defect classification label.

[0054] S602: Based on the S4 neural network model, it uses the gradient weighted class activation mapping algorithm to calculate the image region that contributes the most to the final classification result during the network decision-making process and generate a heat map, generating an interpretable heat map of the defect region. The above content employs a gradient-weighted class activation mapping algorithm to calculate the image region that contributes most to the final classification result during the network decision-making process, according to the formula... ; Calculate the category activation graph.

[0055] In the formula, This represents the importance weight of the k-th feature map to category c. Represents the k-th feature map. The activation graph represents category c.

[0056] First, calculate the gradient of the network output with respect to category c, ∂yc / ∂Ak, where yc is the score of category c, and Ak is the k-th feature map of the last convolutional layer. The feature map size is 64×64×512. For each feature map k, calculate the global average pooling to obtain the weight α_k^c = (1 / Z) * ∑_i∑_j ∂yc / ∂A_{ij}k, where Z=64×64=4096 is the number of pixels in the feature map. For example, for category c, 512 weight values ​​are calculated. Multiply the weights by the corresponding feature maps and sum them, ∑_k α_k^c* A^k. Use the ReLU function to retain positive activation values ​​and generate an activation map. This activation map is upsampled to the original image size of 2048×2048 through bilinear interpolation to generate an interpretable heatmap of the defect area.

[0057] S603: Based on the interpretability heatmap of the defect area generated by S602, it is superimposed with the denoised two-dimensional image generated by S3 to intuitively display the focus area of ​​the model and generate a visualized detection result map. The interpretable heatmap of the defect area generated by S602 is 2048×2048 pixels with values ​​ranging from 0 to 1. It is overlaid with the denoised 2D image generated by S3. The 2D image is in RGB format. The heatmap is converted to Jet color mapping, with red representing high-activation areas and blue representing low-activation areas. The overlay weights are set to 0.5 for the heatmap and 0.5 for the original image. The pixel value after overlay is calculated as: original image pixel value × 0.5 + heatmap pixel value × 0.5 × 255. A visualized detection result image is then generated.

[0058] S604: Integrates the final defect classification labels generated by S601, the three-dimensional quantitative parameter set of defects generated by S5, and the visual detection result map generated by S603. According to the predefined template, it automatically generates a structured report containing defect type, quantity, size, location, and visual evidence, and generates a jade defect intelligent detection report with an interpretable heat map.

[0059] The system integrates the final defect classification labels generated by S601, the 3D quantization parameter set of defects generated by S5, and the visualized detection result map generated by S603. The classification labels include cracks, the parameter set includes size, direction, depth, and density, and the visualization map is an overlay image. According to the predefined report template, which includes a title, detection time, number of defects, defect list, parameter table, and image region, the data is filled into the corresponding positions in the template. The defect list is arranged by serial number, and each defect displays its type, size, and location coordinates. The parameter table displays the 3D quantization values, and the image region is embedded in the visualization map to generate a structured report in PDF format. It also generates an intelligent detection report for jade defects with an interpretable heat map.

[0060] The intelligent jade defect detection system based on machine vision and deep learning includes the following modules; The data acquisition and preprocessing module, based on a high-resolution industrial camera and a laser 3D scanner, simultaneously acquires two-dimensional color image sequences and three-dimensional point cloud sequences of jade samples, optimizes the images and point clouds, and generates an optimized multimodal data cube. The multimodal fusion analysis module, based on the optimized multimodal data cube generated by the data acquisition and preprocessing module, processes two-dimensional image information, processes three-dimensional geometric information, integrates information, outputs pixel-level segmentation results and preliminary classification confidence of defects, and generates defect segmentation masks and preliminary confidence scores. The defect precise quantification module, based on the defect segmentation mask generated by the multimodal fusion analysis module, maps the mask back to the three-dimensional space of the optimized multimodal data cube, isolates the defects, calculates the three-dimensional spatial size and orientation of each isolated defect point cloud cluster, calculates the average depth, volume, and point density quantification parameters, and generates a set of three-dimensional quantification parameters for defects. The decision and report generation module, based on the preliminary confidence level generated by the multimodal fusion analysis module and the three-dimensional quantitative parameter set of defects generated by the defect precision quantification module, performs final arbitration on the defect type, generates a heat map of the highlighted defect area, merges it with the original image, automatically generates a structured report, and generates a jade defect intelligent detection report with an interpretable heat map.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent detection of jade defects based on machine vision and deep learning, characterized in that, Includes the following steps: S1: Based on a high-resolution industrial camera and a laser 3D scanner, a multimodal synchronous acquisition strategy is adopted, and the jade sample is rotated through a precision motion control system to obtain high-resolution two-dimensional color images and high-precision three-dimensional point cloud data of the jade surface, thereby generating a multimodal raw dataset of jade. S2: Based on the jade multimodal raw dataset generated by S1, the two-dimensional image and the three-dimensional point cloud are aligned at the pixel level, and color and texture features are extracted from the two-dimensional image and geometric morphology features are extracted from the three-dimensional point cloud. A multi-dimensional feature representation under a unified coordinate system is constructed to generate a registration and fusion multimodal feature map. S3: Based on the registration and fusion multimodal feature map generated by S2, the two-dimensional image data is enhanced and denoised, and the three-dimensional point cloud data is smoothed and simplified to generate an optimized multimodal data cube. S4: Based on the optimized multimodal data cube generated by S3, it processes two-dimensional image information to segment surface defects, processes three-dimensional geometric information to detect deformation defects, integrates dual-path information, and generates a binary segmentation mask for jade defects and a preliminary classification confidence matrix. S5: Based on the binary segmentation mask of jade defects generated by S4, the mask is mapped back to the three-dimensional spatial coordinates of the multimodal feature map generated by S2 through registration and fusion. Each defect region is isolated, and the bounding box size, average depth, and point density geometric parameters of the point cloud are calculated. Three-dimensional quantization measurement is performed to generate a set of three-dimensional quantization parameters for defects. S6: Based on the preliminary classification confidence matrix generated in S4 and the three-dimensional quantization parameter set of defects generated in S5, the defect type is finally determined, a heat map of the highlighted defect area is generated, and it is overlaid with the original image to automatically generate a structured inspection report, generating a smart inspection report of jade defects with heat map.

2. The intelligent detection method for jade defects based on machine vision and deep learning according to claim 1, characterized in that, The steps for generating the jade multimodal raw dataset based on S1 are as follows: S101: Based on a high-precision electronically controlled rotary table and a trigger synchronization controller, the industrial area array color camera is triggered at a specific angle position, and a global shutter exposure mode is used to acquire a sequence of two-dimensional color images of the surface of jade samples from multiple angles, generating a multi-angle two-dimensional image sequence of jade. S102: Based on the synchronous trigger signal of S101, the laser 3D scanner is controlled to synchronously scan the surface of the jade sample at the same angular position. The 3D contour information of the surface is obtained by laser triangulation and a multi-angle 3D point cloud sequence of jade is generated. S103: Based on the jade multi-angle two-dimensional image sequence generated by S101 and the jade multi-angle three-dimensional point cloud sequence generated by S102, a timestamp alignment algorithm is used to associate all data to the same rotation angle coordinate system and store them in the database to generate the jade multimodal original dataset.

3. The intelligent detection method for jade defects based on machine vision and deep learning according to claim 1, characterized in that, The generation, registration, and fusion of multimodal feature maps based on S2 includes the following steps: S201: Based on the single set of data in the original multimodal jade dataset generated by S1, the scale-invariant feature transformation algorithm is used to extract key points and feature descriptors from the two-dimensional image and its corresponding three-dimensional point cloud projection image, respectively, to generate initial feature point pairs for the two-dimensional and three-dimensional modes. S202: Based on the initial feature point pairs generated in S201, the random sampling consensus algorithm is used to remove erroneous matching point pairs, and the optimal homography transformation matrix is ​​calculated to complete the accurate registration of two-dimensional image pixels and three-dimensional point cloud coordinates, generating a fine registration parameter matrix. S203: Based on the fine registration parameter matrix generated by S202, perform perspective transformation on the original two-dimensional image to make it completely aligned with the three-dimensional point cloud data in space, and generate the registered two-dimensional image. S204: Based on the registered 2D image and its corresponding 3D point cloud generated by S203, a dual-branch convolutional neural network is used to extract deep features respectively. The extracted color texture features and geometric morphology features are then fused in the channel dimension through a feature splicing layer to generate a registered and fused multimodal feature map.

4. The intelligent jade defect detection method based on machine vision and deep learning according to claim 1, characterized in that, The generation of the optimized multimodal data cube based on S3 includes the following steps: S301: Based on the registered two-dimensional image generated by S2, the contrast-limited adaptive histogram equalization algorithm is used to enhance the detail information in the low-contrast area of ​​the image, highlight the potential defect features, and generate a contrast-enhanced two-dimensional image. S302: Based on the contrast-enhanced two-dimensional image generated by S301, a non-local mean filtering algorithm is used for image denoising. While smoothing the background, the edge information of the defects is preserved to the maximum extent, and a denoised two-dimensional image is generated. S303: Based on the 3D point cloud data generated by S1, the voxel grid downsampling algorithm is used to reduce the point cloud density and reduce the amount of data while maintaining the overall shape of the point cloud, thus generating a downsampled 3D point cloud. S304: Based on the downsampled 3D point cloud generated by S303, the moving least squares method is used to resample and smooth the point cloud surface to eliminate noise points generated during the scanning process and generate a smoothed 3D point cloud. S305: Based on the denoised 2D image generated by S302 and the smoothed 3D point cloud generated by S304, the two are recombined into a data unit according to the registration relationship of S2 to generate an optimized multimodal data cube.

5. The intelligent detection method for jade defects based on machine vision and deep learning according to claim 1, characterized in that, The steps for generating a binary segmentation mask for jade defects and a preliminary classification confidence matrix based on S4 are as follows: S401: Based on the optimized multimodal data cube generated by S3, it is input into a fully convolutional network with an encoder-decoder structure; the encoder part is constructed using depthwise separable convolutions to extract high-level semantic features hierarchically from the fused data and generate high-level multimodal semantic feature maps; S402: Based on the high-level multimodal semantic feature map generated by S401, transposed convolution is used in the decoder part for upsampling to gradually restore the spatial resolution of the feature map. Features from the medium-resolution layer of the encoder are fused through skip connections to accurately locate the defect boundary and generate a high-resolution defect segmentation score map. S403: Based on the high-resolution defect segmentation score map generated by S402, the Argmax function or thresholding is used to assign a category label to each pixel, generate a binary segmentation result, and generate a binary segmentation mask for jade defects. S404: At the end of the encoder of S401, a global average pooling layer and a softmax classifier are connected in parallel. Based on the extracted global features, the types of defects in the image are initially predicted, and an initial classification confidence matrix is ​​generated.

6. The intelligent detection method for jade defects based on machine vision and deep learning according to claim 1, characterized in that, The generation of a 3D quantization parameter set for defects based on S5 includes the following steps: S501: Based on the jade defect binary segmentation mask generated by S4, the connected component labeling algorithm is used to identify and distinguish the independent defect regions in the mask, assign a unique identifier to each defect, and generate an independent defect region identifier map. S502: Based on the independent defect region identification map generated by S501, use it as an index to extract the 3D point cloud subset corresponding to each defect region from the optimized multimodal data cube generated by S3, and generate an independent defect 3D point cloud cluster. S503: For each independent defect 3D point cloud cluster generated by S502, the principal component analysis algorithm is used to calculate its direction and size in 3D space, and its minimum outer envelope cuboid is constructed to generate the defect 3D size and direction vector. S504: Based on the independent defect 3D point cloud clusters generated by S502, calculate the average depth and point density of the point cloud clusters, and generate defect depth and density parameters; S505: Integrates the defect 3D size and direction vector generated by S503 with the defect depth and density parameters generated by S504 to form a structured quantization record for each defect, generating a set of defect 3D quantization parameters.

7. The intelligent detection method for jade defects based on machine vision and deep learning according to claim 1, characterized in that, Generating a smart jade defect detection report with an interpretable heatmap based on S6 includes the following steps: S601: Based on the preliminary classification confidence matrix generated in S4 and the three-dimensional quantification parameter set of defects generated in S5, a rule-based expert system is used to make a decision on the final type of each defect and generate the final defect classification label. S602: Based on the S4 neural network model, it uses the gradient weighted class activation mapping algorithm to calculate the image region that contributes the most to the final classification result during the network decision-making process and generate a heat map, generating an interpretable heat map of the defect region. S603: Based on the interpretability heatmap of the defect area generated by S602, it is superimposed with the denoised two-dimensional image generated by S3 to intuitively display the focus area of ​​the model and generate a visualized detection result map. S604: Integrates the final defect classification labels generated by S601, the three-dimensional quantitative parameter set of defects generated by S5, and the visual detection result map generated by S603. According to the predefined template, it automatically generates a structured report containing defect type, quantity, size, location, and visual evidence, and generates a jade defect intelligent detection report with an interpretable heat map.

8. A jade defect intelligent detection system based on machine vision and deep learning, characterized in that, Includes the following modules; The data acquisition and preprocessing module, based on a high-resolution industrial camera and a laser 3D scanner, simultaneously acquires two-dimensional color image sequences and three-dimensional point cloud sequences of jade samples, optimizes the images and point clouds, and generates an optimized multimodal data cube. The multimodal fusion analysis module, based on the optimized multimodal data cube generated by the data acquisition and preprocessing module, processes two-dimensional image information, processes three-dimensional geometric information, integrates information, outputs pixel-level segmentation results and preliminary classification confidence of defects, and generates defect segmentation masks and preliminary confidence scores. The defect precise quantification module, based on the defect segmentation mask generated by the multimodal fusion analysis module, maps the mask back to the three-dimensional space of the optimized multimodal data cube, isolates the defects, calculates the three-dimensional spatial size and orientation of each isolated defect point cloud cluster, calculates the average depth, volume, and point density quantification parameters, and generates a set of three-dimensional quantification parameters for defects. The decision and report generation module, based on the preliminary confidence level generated by the multimodal fusion analysis module and the three-dimensional quantitative parameter set of defects generated by the defect precision quantification module, performs final arbitration on the defect type, generates a heat map of the highlighted defect area, merges it with the original image, automatically generates a structured report, and generates a jade defect intelligent detection report with an interpretable heat map.

Images