A method for intelligent detection and quantitative evaluation of residual powder of a porous structure bone implant instrument based on visual recognition

Abstract

The application relates to a kind of intelligent detection and quantitative evaluation methods of residual powder of porous structure bone implant instrument based on visual identification, relate to nondestructive testing technical field, to solve the problem of low efficiency, limited precision, strong subjectivity of traditional residual powder detection technology, the following steps are innovated in the application: using high-resolution electron microscope (SEM) and dual-mode cooperative acquisition technology of computer tomography (CT), the pictures of surface topography features and internal residual powder state of porous structure are obtained, the images are preprocessed in batches using Matlab, the residual powder is positioned using Otsu method and U-Net model, the particle size distribution characteristics of residual powder are quantified, the geometric characteristics of porous structure, the multi-scale feature fusion is carried out on the preprocessed image, the accurate positioning of residual powder, lightweight preliminary screening, output high confidence residual powder distribution and feature map, finally, the spatial distribution, morphological characteristics and melting state of residual powder are quantified, and quality detection evaluation is output, the application can realize efficient identification and quality detection of residual powder of porous structure bone implant instrument.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing technology, specifically to a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition. Background Technology

[0002] Visual recognition technology, often referred to as computer vision, involves capturing image information of target objects using high-resolution cameras or webcams, and then analyzing and processing these images using image processing algorithms and deep learning techniques. In recent years, visual recognition technology has been widely applied in various fields such as industrial automation, healthcare, and security monitoring.

[0003] Traditional residual powder detection techniques heavily rely on the experience, eyesight, and subjective judgment of the inspectors, resulting in numerous limitations: low efficiency, limited accuracy, and strong subjectivity. Visual recognition technology, on the other hand, enables high-precision detection and localization of residual powder, reducing manual intervention. Furthermore, real-time image analysis allows for rapid analysis of changes in residual powder. It is applicable to porous structures of various shapes and materials and can handle complex background interference. Summary of the Invention

[0004] In order to solve the problems of traditional residual powder detection technology, this invention provides a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] A method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition includes the following steps:

[0007] Step 1: Acquire images of the porous structure using a dual-modal collaborative acquisition technique combining high-resolution electron microscopy (SEM) and computed tomography (CT) to obtain the surface morphology and internal residual powder state of the porous structure. Machine vision algorithms are used to analyze the preview images from the SEM and CT scans in real time, automatically adjusting parameters to ensure image clarity. A two-way image correlation and interactive function is added to significantly improve data visualization.

[0008] Step Two: Batch preprocessing of images is performed using Matlab software. Intelligent cropping is achieved through Gaussian convolution and Souvola binarization, combined with GPU-accelerated nonlocal means denoising and CLAHE enhancement to improve image quality. An innovative approach is adopted: a dual-threshold Otsu algorithm combined with a U-Net model is used for three-region segmentation. Topological features of the porous structure are analyzed through 3D voxel reconstruction and a watershed algorithm, with GPU parallel computing integrated to optimize efficiency. Powder residue identification is performed by fusing feature extraction and random forest classification, ultimately establishing a three-dimensional quantification system to output physical indicators such as porosity and powder residue rate. The entire process innovatively integrates traditional algorithms with deep learning to achieve preliminary identification and evaluation of porous structures and powder residue.

[0009] Step 3: After image preprocessing, a multi-scale feature pyramid (P2-P5) is generated using ResNet, and cross-layer fusion is performed using a Feature Pyramid Network (FPN) to combine deep semantic information with shallow detail features. Next, an Adaptive Region Proposal Network (RPN) is used to dynamically optimize the size and proportion of the anchor boxes, and high-confidence candidate boxes are further selected through boundary regression and non-maximum suppression (NMS). Pixel-level feature alignment is performed using ROIAlign to ensure accurate localization of the candidate boxes. Subsequently, a lightweight classifier combined with FocalLoss filters the background region, preserving the target region. Finally, high-confidence residual powder candidate regions and their feature maps are output to support subsequent detection tasks.

[0010] Step 4: Construct a multi-scale feature fusion network to enhance small target detection capabilities, and combine a two-stage candidate box optimization strategy with the YOLOv8 framework to achieve efficient localization. Extend the detection-segmentation-fusion multi-task collaborative architecture, resolving task conflicts through gradient orthogonalization and adaptive weighting mechanisms. Finally, output a quality assessment, integrating a dynamic threshold alarm mechanism to achieve efficient identification and quality detection of residual material from porous bone implant devices. Attached Figure Description

[0011] Figure 1 This is a flowchart of a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition, according to the present invention.

[0012] Figure 2 This is a recognition diagram of the intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition according to the present invention.

[0013] Figure 3 This is a schematic diagram of residual powder on the support rod surface of the present invention, which is a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition.

[0014] Figure 4This image shows the detection data results of a visual recognition-based intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices according to the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below: This embodiment is implemented under the premise of the technical solution of the present invention, and a detailed implementation method is given, but the protection scope of the present invention is not limited to the following embodiment. Specific Implementation Example 1: This embodiment relates to a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition. The detection method includes the following steps: Step 1: Acquire images of the porous structure using a dual-modal collaborative acquisition technique combining high-resolution electron microscopy (SEM) and computed tomography (CT) to obtain the surface morphology and internal residual powder state of the porous structure. Machine vision algorithms are used to analyze the preview images from the SEM and CT scans in real time, automatically adjusting parameters to ensure image clarity. A two-way image correlation and interactive function is added to significantly improve data visualization. Step Two: Batch preprocessing of images is performed using MATLAB software. Intelligent cropping is achieved through Gaussian convolution and Souvola binarization, combined with GPU-accelerated nonlocal means denoising and CLAHE enhancement to improve image quality. An innovative approach is adopted: a dual-threshold Otsu algorithm combined with a U-Net model is used for three-region segmentation. Topological features of the porous structure are analyzed through 3D voxel reconstruction and a watershed algorithm, with GPU parallel computing integrated to optimize efficiency. Powder residue identification is performed by fusing feature extraction and random forest classification, ultimately establishing a three-dimensional quantification system to output physical indicators such as porosity and powder residue rate. The entire process innovatively integrates traditional algorithms with deep learning to achieve preliminary identification and evaluation of porous structures and powder residue. Step 3: After image preprocessing, a multi-scale feature pyramid (P2-P5) is generated using ResNet, and cross-layer fusion is performed using a Feature Pyramid Network (FPN) to combine deep semantic information with shallow detail features. Next, an Adaptive Region Proposal Network (RPN) is used to dynamically optimize the size and proportion of the anchor boxes, and high-confidence candidate boxes are further selected through boundary regression and non-maximum suppression (NMS). Pixel-level feature alignment is performed using ROIAlign to ensure accurate localization of the candidate boxes. Subsequently, a lightweight classifier combined with Focal Loss filters background regions, retaining high-potential target regions. Finally, high-confidence residual powder candidate regions and their feature maps are output, providing accurate support for subsequent detection tasks. Step 4: Construct a multi-scale feature fusion network to enhance small target detection capabilities, and combine a two-stage candidate box optimization strategy with the YOLOv8 framework to achieve efficient localization. Extend the detection-segmentation-fusion multi-task collaborative architecture, resolving task conflicts through gradient orthogonalization and adaptive weighting mechanisms. Finally, output a quality assessment, integrating a dynamic threshold alarm mechanism to achieve efficient identification and quality detection of residual material from porous bone implant devices. Specific Implementation Example 2: This embodiment relates to a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition. The method for obtaining the surface morphology features and internal residual powder status of the porous structure further includes the following steps: The porous sample is fixed on the sample stage using a special fixture with positioning grooves to ensure that the sample pose is consistent when scanning with high-resolution electron microscope and computed tomography equipment. Gold nanosphere markers are implanted on the sample surface to provide a reliable cross-scale spatial reference for multimodal imaging. The characterization equipment was activated, and high-resolution images of the porous structure surface were acquired using a high-resolution electron microscope (SEM) to assess its surface morphology quality. Simultaneously, computed tomography (CT) was used to penetrate the part with X-rays, acquiring projection data of the porous structure. Reconstruction algorithms were then used to generate three-dimensional volume data to analyze internal defects. SEM focuses on identifying high-resolution details on the porous structure surface, while CT focuses on identifying the internal three-dimensional structure. The combination of both comprehensively covers the dual "surface-interior" characteristics of the porous structure. Furthermore, a spatial mapping matrix was established using gold sphere registration, and cross-modal data fusion was achieved through registration and the ICP algorithm. Machine vision algorithms are used for real-time image processing to simultaneously optimize image quality. SEM images employ multi-scale edge enhancement techniques, fusing Sobel and Gabor filters to extract micron-level surface textures, and dynamic contrast stretching optimizes local grayscale distribution to avoid overexposure or underexposure. CT projection uses a flat-panel response correction matrix to dynamically update and suppress ring artifacts, combined with a GPU-accelerated real-time CLAHE algorithm to enhance the visibility of internal structures in low-contrast areas. To improve image sharpness, SEM images assess detail integrity through edge density and dynamic range scoring, while CT projection measures structural visibility based on local entropy and optimizes radiation safety through a dose efficiency index. Furthermore, dynamic parameter adjustments are possible. Based on a PID closed-loop control model, the SEM electron beam focusing current and scanning speed can be adjusted in real-time, and the X-ray energy and exposure time of CT are dynamically optimized through a linear model to ensure image sharpness. Through real-time processing and dynamic parameter closed-loop control, simultaneous acquisition and optimization of surface and internal structural information are achieved. To establish a two-way image-based interactive function, a unified spatial mapping relationship must first be established through coordinate registration. Feature point matching is then used to precisely align the two-dimensional pixel coordinate system of the SEM image with the spatial coordinate system of the three-dimensional model. In the interactive interface, when a specific area is selected in the SEM image, the system uses a coordinate transformation algorithm to map the selected area's coordinates to three-dimensional space. Voxel marking technology is used to highlight the corresponding residual powder area, and spatial indexing accelerates localization. Conversely, when a residual powder accumulation area is clicked in the three-dimensional model, the system reverse-calculates the projected coordinates of that location in the SEM image, combines image pyramid technology to quickly locate the corresponding area, and overlays a heatmap of morphological parameters. Throughout the process, GPU acceleration is added to ensure smooth visualization under large-scale data. Specific Implementation Example 3: This embodiment relates to a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition. The method further includes the following steps during batch image preprocessing using Matlab software: First, a batch preprocessing module based on Matlab is constructed. The dir function is used to traverse image files in a specified folder. The imread function is used to read the image within the loop and then perform preprocessing operations: first, Gaussian kernel is used for spatial convolution to achieve smooth denoising, then the Sauvola algorithm is applied to calculate the local dynamic threshold, then the effective region is extracted through binarization, and finally, a non-black pixel boundary detection algorithm based on row and column projection is designed to automatically calculate the effective region to complete intelligent cropping. In the multi-stage enhancement process, firstly, adaptive median filtering is applied to the image to suppress impulse noise. This method dynamically adjusts the size of the filtering window to adapt to different noise environments. The specific process is as follows: Dynamic adjustment of the filtering window: The size of the filtering window is automatically adjusted based on the noise characteristics of local areas of the image. For areas with impulse noise, the filtering window is enlarged to include more neighboring pixels for noise smoothing. For clearer areas, the window is reduced to retain more detail. Impulse noise removal: By replacing the center pixel with the median of its neighboring pixels, impulse noise in the image is effectively removed while preserving the image's edge structure. Next, the processed image data is migrated to the GPU to improve the efficiency of denoising calculations. `gpuArray` is used to migrate the image data from the CPU to the GPU, thus fully utilizing the parallel computing advantages of the GPU. Then, SSIM-weighted nonlocal means denoising is performed, as follows: Migrate data to GPU: Use the gpuArray function to migrate image data to the GPU, avoiding data transfer bottlenecks between the CPU and GPU. Calculate SSIM similarity: The GPU is used to compute the SSIM value of each pixel in the image and its neighboring pixels in parallel. SSIM is used to represent the similarity between two regions. Regions with higher SSIM values ​​are assigned larger weighting coefficients to enhance the denoising effect. Nonlocal mean denoising: Utilizing calculated SSIM weights, each pixel is weighted and averaged with its neighboring pixels to smooth noise while preserving image details, especially at edges. It effectively removes Gaussian noise and maintains image texture and edge details. If an image suffers from insufficient contrast after denoising, the CLAHE algorithm is used to enhance local contrast. Local equalization in CLAHE enhances local details by performing histogram equalization within a small neighborhood, while avoiding noise amplification caused by over-enhancement. To prevent over-enhancement, CLAHE sets a contrast limit to ensure that significant artifacts and noise are not introduced during local enhancement. Finally, to further enhance image details, the Laplacian operator is used for edge sharpening. The Laplacian operator calculates the second derivative of an image and detects edges through convolution operations. This operator effectively highlights image details and enhances contrast in edge regions. The Laplacian-processed image is then weighted back to the original image, thereby improving edge sharpness while maintaining smoothness in other areas of the image. In the image segmentation stage, the innovative dual-threshold Otsu algorithm is adopted. First, by maximizing the inter-class variance of the three classes (pore region, skeleton region, and residual powder region), the optimal threshold combination T1 and T2 are obtained. The core formula is as follows: Gray-level histogram statistics: Let the total number of pixels in the image be N, the gray level range be [0, L-1], and the number of pixels at gray level i be n. i Its probability is The probabilities and average gray levels of the three classes: Porous region: gray level 0≤i≤T1, probability Average gray level Skeleton: Gray level T1≤i≤T2, probability Average gray level Residual powder area: grayscale T2≤i≤L-1, probability Average gray level Between-class variance maximize: in, This represents the overall average gray level of the image. Based on the mathematical principles of the dual-threshold Otsu algorithm, the pixel distribution of each gray level in the preprocessed image is statistically analyzed to generate a normalized histogram p. i The threshold combination is iteratively optimized using dynamic programming, and the accumulated probability and gray-level mean are updated through a recursive formula: First, pre-calculate the cumulative amount, cumulative probability S(k), and cumulative gray mean U(k): Dynamic programming iteration: For each candidate T1, recursively calculate the optimal value of T2. Using a recursive formula, the time complexity is reduced from O(L) 2 ) decreased to O(L). After determining the values ​​of T1 and T2, a third-order quantization model is established to divide the pore region, skeleton region, and residual powder candidate region: gray values ​​below T1 are classified as pore regions, gray values ​​in the T1-T2 range are classified as skeleton regions, and gray values ​​above T2 are classified as residual powder candidate regions. At the same time, the candidate region ROI is input into the pre-trained U-Net model, and the traditional segmentation results and deep learning predictions are combined with pixel-by-pixel logical AND operations to generate high-confidence candidate regions, thereby reducing the false detection rate. For the identification of porous structures, firstly, topological features are reconstructed based on a 3D voxel mesh. The number and volume distribution of 3D connected components are quantified using the `bwconncomp` function in Matlab, and the traditional Euler number is extended by combining the Betti number. The complex structure of the pores is analyzed using a computational topology library. Secondly, to achieve accurate 3D partitioning of porous structures, an optimized 3D watershed algorithm is adopted. Initial markers are generated through distance transformation, and a region merging strategy is combined to suppress over-segmentation, achieving accurate 3D partitioning of porous structures. Finally, for curvature calculation, the centerline of the 3D skeleton is extracted and a local triangular mesh is constructed. The radius of curvature is dynamically estimated based on the geometric relationship between adjacent nodes, and a surface normal vector change rate algorithm is introduced to enhance the accuracy of spatial curvature representation. In addition, GPU acceleration and parallel block processing technologies can be integrated. Matlab's `gpuArray` is used to migrate computationally intensive tasks such as 3D convolution and watershed to the graphics card hardware, and `parfor` multi-threaded processing is combined to improve the efficiency of large-scale volume data processing. In the residual powder intelligent identification stage, circular structuring elements of different sizes are used to perform opening operations to eliminate tiny noise particles while retaining residual powder structures of different scales. Subsequently, multi-dimensional features are extracted based on connected component analysis. Morphological features are used to calculate the pixel area, aspect ratio of the circumscribed rectangle, and eccentricity of the minimum circumscribed ellipse for each connected region. Gray-scale features are extracted from the pixel mean and variance within the statistical region. Spatial distribution features are extracted by calculating the nearest neighbor distance between adjacent residual powder candidate regions. Texture features are extracted using the Gray-Level Co-occurrence Matrix (GLCM). The importance of each feature is calculated using a pre-trained random forest classifier, and the importance of the features is further verified by the Boruta algorithm to ensure that the selected features are statistically significant. In the final quantitative analysis stage, an innovative three-dimensional measurement system was established to calculate the pixel proportions of the pore area, skeleton area, and residual powder area, and multiply them by the pixel calibration coefficient to convert them into actual areas. Simultaneously, the mean and standard deviation of grayscale values ​​for each area were statistically analyzed to generate a quantitative report. Furthermore, the following characteristics of residual powder were quantified and analyzed: Sphericity: By analyzing the geometry of residual powder, the degree to which it approximates a sphere is quantified, and its morphological characteristics are evaluated. Surface roughness: Analyze the unevenness of the residual powder surface and quantify its surface roughness to evaluate the physical properties of the residual powder. Particle size distribution: By measuring the size of the residual powder and generating a particle size distribution histogram, the size and distribution characteristics of the particles are quantified. Bonding strength between residual powder and substrate: The interface characteristics between residual powder and substrate are analyzed, and the bonding strength is evaluated by texture, contrast and edge features. Simultaneously, the geometric characteristics of the porous structure are analyzed, and porosity, pore size distribution, and pore connectivity are calculated, while the connectivity of the pore structure is monitored. Finally, through multi-dimensional feature quantification and structural analysis, the preliminary identification and assessment of the porous structure and residual powder are completed, generating a structure-physical property correlation report. Specific Implementation Example 4: This embodiment relates to a method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition. The analysis of the preprocessed image further includes the following steps: The image preprocessed in Matlab is input into ResNet to generate a multi-scale feature pyramid (P2-P5). The deep features (P5) capture global semantic information such as residual powder texture, while the shallow features (P2) preserve high-resolution details. The fusion of deep and shallow features is achieved through the natural cross-layer connections of FPN, avoiding the gradient vanishing problem of traditional CNNs. Target localization using RPN begins with anchor box design. Adaptive anchor boxes dynamically adjust their size and proportions to accommodate different target sizes and shapes, improving detection accuracy and efficiency. Next, target localization is performed. For each anchor box, binary classification and bounding box regression are conducted, and non-maximum suppression is used to filter candidate boxes, retaining those with high confidence. Finally, candidate boxes are mapped to corresponding feature maps, and bilinear interpolation is used to achieve ROI alignment, eliminating quantization errors inherent in traditional ROIPooling and ensuring accurate feature extraction. A lightweight initial screening is performed by adding a lightweight fully connected layer after the RPN. Based on the candidate region features extracted by ROI Align, binary classification is performed. Focal Loss is used to alleviate class imbalance, retaining only high-confidence candidate regions and filtering out more than 90% of background interference, thereby reducing the computational load in subsequent stages. The module outputs a multi-scale feature pyramid feature map (P2-P5) and a high-confidence residual powder distribution. Specific Implementation Example 5: This embodiment relates to a visual recognition-based intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices. The method quantifies the spatial distribution, morphological characteristics, and melting state of the residual powder, and the final quality evaluation output further includes the following steps: The multi-scale feature module output from Example 4 replaces its default feature extraction path. A dynamic upsampling mechanism is introduced into the P2-P5 feature pyramid, and adaptive fusion of features at adjacent scales is achieved through learnable parameters. Combined with a spatial attention-guided cross-scale feature selection module, the ability to express the details of residual powder on small targets is significantly improved. The detection architecture adopts a two-stage candidate box optimization strategy: the first stage loads the morphological preliminary screening candidate boxes from Example 3 as priors; the second stage uses a lightweight CNN to verify confidence and implement a dynamic anchor box elimination strategy, combined with an online pruning mechanism to reduce computational redundancy. The efficient detection framework of YOLOv8 is retained to ensure real-time processing speed. The multi-task extended design constructs a three-branch architecture of detection, segmentation, and melting evaluation: The detection branch utilizes YOLOv8's bounding box regression and classification head to output the precise location, size, and quantity of residual powder. The segmentation branch designs a lightweight mask head, integrating an edge-sensitive convolutional module and a multi-scale porous spatial pyramid (ASPP), generating pixel-level masks based on P2 features and calculating morphological parameters such as area and eccentricity. The melting evaluation branch integrates a 3-layer MLP at the end of the detection head to classify the residual powder state into free, semi-molten, and fully molten states. Multi-dimensional feature fusion and evaluation construct a four-level analysis system: Spatial distribution analysis is based on the detection frame density distribution, using adaptive bandwidth kernel density estimation (KDE) to generate a hierarchical heatmap, and combining it with Voronoi diagrams to divide the core area, transition area, and edge area, quantifying the density proportion of each region. Morphological parameter statistical analysis utilizes the pixel-level mask output by the segmentation branch to calculate the morphological characteristics of each residual powder, establishing an abnormal morphology screening rule library. Melting state correlation analysis aligns the output probability of the melting evaluation branch with the spatiotemporal process parameters to analyze the spatial distribution patterns of residual powder in different states. The quality inspection and evaluation system constructs a three-layer scoring network, comprehensively considering the proportion of liquid residual powder, core area density, and morphological parameters to generate a score from 0 to 100, ultimately outputting a quality inspection evaluation. Combined with dynamic threshold alarms, it achieves real-time quality inspection, realizing efficient identification and quality inspection of residual powder in porous bone implant devices. While the present invention has been described in detail through specific embodiments, those skilled in the art should understand that the above examples are for illustrative purposes only and not intended to limit the scope of the invention. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A method for intelligent detection and quantitative evaluation of residual powder in porous bone implant devices based on visual recognition, characterized in that: Includes the following steps: Step 1: Acquire images of the porous structure. Use a dual-modal collaborative acquisition technique of high-resolution electron microscopy (SEM) and computed tomography (CT) to acquire images of the surface morphology features and internal residual powder state of the porous structure. Step 2: Batch image preprocessing is performed using Matlab to intelligently crop effective regions and suppress noise. Multi-stage enhancement improves contrast, and the Otsu method combined with the U-Net model achieves efficient and accurate localization of residual powder. The sphericity, surface roughness, and particle size distribution characteristics of residual powder are quantified, and the bonding strength between residual powder and the matrix interface, as well as the geometric characteristics of the porous structure, are monitored to achieve preliminary identification and evaluation of the porous structure and residual powder. Step 3: Analyze the preprocessed image to capture the material and uniformity of residual powder, and retain high-resolution images of the edges of tiny residual powders to preserve deep features. Lightweight initial screening filters out background interference, ultimately outputting a high-confidence residual powder distribution and feature map, providing accurate support for subsequent quality testing of bone implant devices. Step 4: Quantify the spatial distribution, morphological characteristics and melting state of residual powder, and finally output the quality inspection evaluation to achieve efficient identification and quality inspection of residual powder of porous bone implant devices.

2. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition as described in claim 1, characterized in that: The method for obtaining photographs of the surface morphology and internal residual powder state of porous structures is as follows: To acquire images of porous structures, a dual-modal collaborative acquisition technique combining high-resolution electron microscopy (SEM) and computed tomography (CT) is employed to obtain the surface morphology features and internal residual powder state of the porous structures. The combination of the two techniques can comprehensively cover the dual "surface-interior" characteristics of the porous structures. By using machine vision algorithms to analyze preview images from high-resolution electron microscopes and computed tomography scans in real time, the parameters of characterization equipment can be intelligently calibrated to ensure image clarity and simultaneously optimize imaging quality. To better establish the connection between surface morphology and internal structure, a two-way image correlation and interaction function has been added. Selecting a specific area in the SEM image can automatically locate and highlight the corresponding internal residual powder distribution in the 3D model. Clicking on a residual powder accumulation area in the 3D model will display the corresponding surface SEM image and morphology parameters, significantly improving data visualization.

3. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition as described in claim 1, characterized in that: The specific steps for batch image preprocessing using Matlab software are as follows: Batch preprocessing of images is performed by a pre-written program in Matlab that iterates through each image file, reading and preprocessing each image. The program can identify valid regions of the image, perform intelligent cropping, and automatically remove invalid background. Furthermore, Gaussian smoothing preprocessing is added during image reading to suppress high-frequency noise interference with subsequent threshold calculations. In the multi-stage image enhancement stage, an improved adaptive median filtering combined with a nonlocal means denoising algorithm is employed to avoid misidentifying small noise points as residual powder, thus maintaining the integrity of pore edges while suppressing noise. The nonlocal means denoising stage introduces structural similarity (SSIM) weights to prevent low-similarity blocks from participating in the calculation, reducing edge blurring. Furthermore, adaptive histogram equalization and sharpening are used to enhance microstructure contrast, addressing the problem of pore identification in low-contrast images. In the stage of image segmentation, a multi-threshold dynamic segmentation algorithm is adopted to narrow the detection range. Based on the Otsu method, the main threshold and high-region threshold are iteratively calculated, and a third-order quantization model is established to divide the pore region, skeleton region, and residual powder candidate region, thereby improving the robustness of residual powder candidate region localization. A pre-trained deep learning model U-Net is added to supplement the traditional segmentation. It predicts each candidate region. The residual powder candidate region of the traditional multi-threshold segmentation and the prediction result of U-Net are logically ANDed, and only the overlapping area of ​​the two is retained as a high-confidence residual powder candidate, thereby reducing the false detection rate. To identify the geometric features of the porous structure, the pore size distribution is first analyzed to generate a pore size distribution histogram. Next, the radius of curvature is calculated to obtain the bending characteristics of the framework. Finally, the number of connected components is calculated. Intelligent identification of residual powder involves filtering out noise through morphological opening operations, extracting multi-dimensional features such as morphology, grayscale, spatial distribution, and texture, and using a random forest classifier combined with the Boruta algorithm to screen highly significant features for accurate identification of residual powder. Finally, in the multi-dimensional quantitative analysis stage, the pixel proportions of pores, framework, and residual powder are converted into actual physical areas, and the grayscale mean and standard deviation are statistically analyzed to generate a quantitative report. The sphericity, surface roughness, and particle size distribution of the residual powder are quantified, and the bonding strength between the residual powder and the matrix is ​​monitored. Simultaneously, the geometric characteristics of the porous structure are analyzed to achieve preliminary identification and assessment of the porous structure and residual powder, and a correlation analysis between structure and physical properties is completed.

4. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition as described in claim 1, characterized in that: The specific steps for analyzing the preprocessed image are as follows: First, the preprocessed image is input into ResNet to extract multi-level features (P2-P5). The deep features (P5) capture global semantic information such as residual powder texture, while the shallow features (P2) preserve high-resolution details. The fusion of deep and shallow features is achieved through the natural cross-level connections of FPN. Based on dynamically adjusting anchor boxes to adapt to target diversity, after generating candidate boxes, high-confidence regions are selected through bounding box regression and binary classification. ROI Align is used to achieve pixel-level alignment feature extraction, avoiding traditional quantization errors. Next, a lightweight classification network is deployed on the candidate regions, and Focal Loss is used to solve class imbalance, quickly filter background interference, and retain high-potential target regions to reduce computational costs. Finally, the multi-scale feature pyramid feature maps (P2-P5) and the initial candidate regions are output.

5. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition according to claim 1, characterized in that: The specific steps for quantifying the spatial distribution, morphological characteristics, and melting state of residual powder, and ultimately outputting quality testing and evaluation, are as follows: Based on dynamic feature fusion to enhance small target detection, a two-stage candidate box optimization strategy is adopted to balance accuracy and real-time performance within the YOLOv8 framework. Multi-task collaboration is used to construct a three-branch system of detection, segmentation, and melting evaluation: detection outputs position and quantity, segmentation generates pixel-level morphological parameters, and melting classifies residual powder states. Furthermore, a gradient coordination mechanism is used to resolve multi-task conflicts. By integrating spatial distribution, morphological statistics, and melting state multidimensional features, the system outputs quality inspection and evaluation, and combines dynamic threshold alarms to achieve real-time quality inspection.

6. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition according to claim 4, characterized in that: It includes a data storage module for storing historical detection data for subsequent analysis and optimization of the detection algorithm.

7. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition according to claim 4, characterized in that: The method also includes a remote monitoring and diagnostic module, allowing users to remotely access and control the system via a network to view the detection progress, results, and detailed information on residual powder in real time. Simultaneously, this module supports remote expert consultation; when the system encounters complex residual powder that is difficult to diagnose, it can automatically or manually request remote assistance and diagnosis from external experts.

8. The intelligent detection and quantitative evaluation method for residual powder in porous bone implant devices based on visual recognition according to claim 5, characterized in that: The method also includes an adaptive learning mechanism that can automatically update and optimize its internal parameters and model structure based on newly detected residual powder samples, so as to continuously improve the accuracy and efficiency of residual powder identification.

Images