Quality detection method for labor protection shoes based on artificial intelligence visual detection

Abstract

The application relates to the technical field of safety shoes detection, and discloses a safety shoe quality detection method based on artificial intelligence visual detection. The method first collects multi-angle image data of safety shoes, generates pretreatment image data sets through noise filtering and size standardization processing; then extracts crack depth, suture line offset and material wear area and other surface defect features from the data sets by using a convolutional neural network to obtain a visual feature vector set; next, the safety shoes are divided into multiple quality grade intervals corresponding to preset quality threshold ranges according to the set, a preliminary quality classification result is generated; then, the result and defect position information are fused, a defect dense area is identified by using a spatial clustering algorithm, and a severity score of the defect dense area is calculated to generate a quality evaluation hotspot map; finally, based on the hotspot map and the preliminary classification result, a decision tree model is used to integrate quality decisions, and a final quality detection report containing a qualified mark or a rework suggestion is output. The method improves the intelligence and accuracy of safety shoe quality detection.

Description

Technical Field

[0001] This invention relates to the field of safety shoe inspection technology, specifically a safety shoe quality inspection method based on artificial intelligence visual inspection. Background Technology

[0002] With the rapid development of the manufacturing industry, the production scale of safety shoes has continued to expand, and the market has placed higher demands on the efficiency and accuracy of their quality inspection.

[0003] Currently, traditional quality inspection of work shoes mainly relies on manual visual inspection. Inspectors judge whether the product is qualified by observing the surface condition, stitching, and material integrity of the work shoes. This method not only requires a large amount of manpower, but the test results are also easily affected by factors such as the inspectors' subjective experience and fatigue, making it difficult to standardize the testing standards and posing a risk of missed or false positives.

[0004] Manual inspection is inefficient and cannot meet the rapid inspection needs of large-scale production lines, often becoming a bottleneck restricting production progress. In addition, manual inspection often fails to accurately identify minor defects such as tiny cracks or slight seam deviations, and these defects may gradually expand during long-term use, affecting the protective performance of safety shoes and posing safety hazards to users.

[0005] With the development of automation technology, some companies have begun to introduce machine vision inspection equipment. However, most existing equipment can only perform simple dimensional measurements and preliminary identification of appearance defects. It is difficult to accurately quantify and analyze the specific characteristics of defects, such as crack depth and the size of material wear areas. Moreover, these devices usually lack the ability to comprehensively evaluate defects and cannot scientifically classify the quality grade of safety shoes based on information such as the location and density of defects, resulting in limited practicality of the inspection results. Summary of the Invention

[0006] The purpose of this invention is to provide a method for quality inspection of safety shoes based on artificial intelligence visual inspection, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for quality inspection of safety shoes based on artificial intelligence visual inspection, the method comprising:

[0008] Collect multi-angle image data of safety shoes, perform noise filtering and size standardization on the multi-angle image data to eliminate lighting interference and scale differences, and generate a preprocessed image dataset;

[0009] Based on a preprocessed image dataset, a convolutional neural network is used to extract surface defect features of work shoes. The surface defect features include crack depth, stitch line offset, and material wear area, generating a set of visual feature vectors.

[0010] Based on the set of visual feature vectors, safety shoes are divided into multiple quality level intervals, each of which corresponds to a preset quality threshold range, generating preliminary quality classification results.

[0011] By integrating the preliminary quality classification results with the defect location information in the visual feature vector set, a spatial clustering algorithm is used to identify defect-dense regions, and the severity score of each defect-dense region is calculated to generate a quality assessment heatmap.

[0012] Based on the quality assessment heatmap and preliminary quality classification results, a decision tree model is used to integrate quality decisions and output a final quality inspection report for safety shoes. The final quality inspection report includes a pass mark or rework recommendation.

[0013] Preferably, the process of acquiring multi-angle image data of the safety shoes, performing noise filtering and size standardization on the multi-angle image data to eliminate lighting interference and scale differences, and generating a preprocessed image dataset includes: acquiring top, side and bottom view images of the safety shoes through an industrial camera to form an original image data stream;

[0014] The Gaussian filtering algorithm is applied to smooth the original image data stream, remove random noise points, and adjust the image brightness to a uniform reference value to generate a smooth image set.

[0015] Each image in the smoothed image set is scaled to a fixed pixel size to ensure that all images have a consistent resolution. Then, histogram equalization is used to enhance image contrast, generating a standardized image set.

[0016] A standardized image set is integrated, and the data is organized according to time sequence and viewpoint type to construct a preprocessed image dataset, which is directly used in subsequent feature extraction steps.

[0017] Preferably, the step of extracting surface defect features of work shoes based on a preprocessed image dataset using a convolutional neural network, wherein the surface defect features include crack depth, suture line offset, and material wear area, and generating a set of visual feature vectors, includes: loading the preprocessed image dataset into a pre-trained convolutional neural network model, extracting deep features of the image through multiple convolutional layers and pooling layers, and generating an original feature map;

[0018] Key regions are segmented from the original feature map. The key regions include the toe, upper and sole. Texture gradient and color distribution parameters are calculated for each key region to generate a subset of region features.

[0019] Merge the regional feature subsets of all key regions, apply principal component analysis algorithm to reduce dimensionality, retain the main feature components, and form a compressed feature vector;

[0020] The compressed feature vectors are associated with the positional coordinates in the original feature maps to construct a set of visual feature vectors, which is then passed as input to the quality level classification module.

[0021] Preferably, the step of dividing the work shoes into multiple quality grade intervals based on the visual feature vector set, with each quality grade interval corresponding to a preset quality threshold range, and generating a preliminary quality classification result includes: parsing the feature values ​​in the visual feature vector set and matching them with the preset quality threshold range, wherein the quality threshold range includes acceptable crack length and maximum wear area;

[0022] A support vector machine classifier is used to divide the feature values ​​into multiple classes, outputting the quality level of each safety shoe sample and generating the original classification label;

[0023] By aggregating the defect count statistics from the original classification labels and visual feature vector sets, the sample distribution density of each quality level interval is calculated to generate classification density distribution data;

[0024] Based on the classification density distribution data, outliers in the original classification labels are corrected, and the final classification result is determined through a majority voting mechanism to form a preliminary quality classification result. This preliminary quality classification result is used to fuse defect location information.

[0025] Preferably, the process of fusing the preliminary quality classification results with the defect location information in the visual feature vector set, identifying densely populated defect areas using a spatial clustering algorithm, calculating the severity score for each densely populated defect area, and generating a quality assessment heatmap includes: extracting the defect location coordinates from the visual feature vector set, mapping them to the coordinate system of the work shoe's 3D model, generating a spatial location dataset; combining the quality level labels from the preliminary quality classification results, performing density peak clustering on the spatial location dataset to identify high-density defect clusters, generating a list of densely populated defect areas; calculating the average defect size and frequency for each densely populated defect area, and applying a weighted formula to obtain a severity score, wherein the severity score is based on defect type weight and location importance;

[0026] By integrating the severity scores and spatial coordinates of all defect-dense regions, a heatmap overlay image is generated, and a quality assessment heatmap is constructed, which serves as the input source for the decision tree model.

[0027] Preferably, the quality decision is integrated using a decision tree model based on the quality assessment heatmap and preliminary quality classification results, and the final quality inspection report for the safety shoes is output. The final quality inspection report includes a pass mark or rework suggestion, including: inputting the severity score in the quality assessment heatmap and the quality level in the preliminary quality classification results into the decision tree model, and generating a decision path through a recursive segmentation algorithm;

[0028] The key nodes in the decision-making path are evaluated, including the defect repairability and safety risk coefficients. The overall quality score is calculated, and decision scoring data is generated.

[0029] Compare the decision score data with the preset pass threshold. If the overall quality score is lower than the pass threshold, it is marked as a rework suggestion; otherwise, it is marked as pass.

[0030] The decision-making results are compiled into a structured report, which includes details of the defect location and recommended treatment measures, and the final quality inspection report of the safety shoes is output.

[0031] Preferably, the step of acquiring top, side, and bottom view images of the work shoes using an industrial camera to form a raw image data stream includes: setting the industrial camera to capture images of the work shoes at a fixed angle to ensure coverage of all key viewpoints and generating a multi-view image sequence;

[0032] Multi-view image sequences are timestamped, arranged in the order of capture, and a time-series image library is established.

[0033] Image metadata, including light intensity and shooting distance, is extracted from a time-series image library for subsequent noise filtering.

[0034] Preferably, the step of loading the preprocessed image dataset into the pre-trained convolutional neural network model and extracting deep features of the image through multiple convolutional layers and pooling layers to generate the original feature map includes: initializing the weights of the convolutional neural network model and optimizing the feature extraction layer using transfer learning techniques;

[0035] The preprocessed image dataset is input into the model in batches, forward propagation calculation is performed, and feature activation maps are output.

[0036] Aggregate feature activation maps to generate multi-scale feature representations, which serve as the basis for the original feature mapping.

[0037] Preferably, the feature values ​​in the set of parsed visual feature vectors are matched with a preset quality threshold range, the quality threshold range including acceptable crack length and maximum wear area, including: defining a quality threshold range database to store upper limit values ​​for different defect types;

[0038] Query the feature values ​​in the visual feature vector set and compare them with the quality threshold range database in real time;

[0039] The output comparison results are used as training data for the support vector machine classifier.

[0040] Preferably, the step of extracting the defect location coordinates from the visual feature vector set and mapping them to the coordinate system of the work shoe 3D model to generate a spatial location dataset includes: constructing a standard 3D model of the work shoe, wherein the standard 3D model includes a coordinate reference system;

[0041] Transform the defect location coordinates to the 3D model coordinate system to ensure spatial consistency;

[0042] Verify the accuracy of coordinate transformation and generate a calibrated spatial location dataset.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] By acquiring multi-angle image data of safety shoes and performing noise filtering and size standardization, lighting interference and scale differences can be effectively eliminated, providing a high-quality image data foundation for subsequent defect detection. Multi-angle image acquisition can comprehensively capture information from all parts of the safety shoes, avoiding defects missed due to limited viewing angles and making the inspection more comprehensive.

[0045] Based on the preprocessed image dataset, a convolutional neural network is used to extract surface defect features, including crack depth, suture line offset, and material wear areas, which can accurately identify and quantify these defect features. Convolutional neural networks have powerful feature learning capabilities, enabling them to extract key defect information from complex images. Compared to traditional manual inspection and simple machine vision inspection, they are more effective at identifying subtle defects and can discover potential problems that are difficult to detect.

[0046] Based on a set of visual feature vectors, safety shoes are divided into multiple quality grade intervals, each corresponding to a preset quality threshold range, making quality classification more standardized and objective. This classification method eliminates the subjective arbitrariness of manual inspection, ensures consistency in test results across different batches and by different inspectors, and provides manufacturers with clear quality judgment standards.

[0047] By integrating preliminary quality classification results with defect location information, a spatial clustering algorithm is used to identify areas with high defect density and calculate the severity score for each area, generating a quality assessment heatmap. This heatmap visually displays the key areas and severity of defects in work shoes. This helps manufacturers quickly pinpoint areas with concentrated quality problems, analyze the causes of defects, and thus make targeted improvements to production processes to reduce the recurrence of similar defects.

[0048] Based on the quality assessment heatmap and preliminary quality classification results, a decision tree model is used to integrate quality decisions and output the final quality inspection report, including a pass / fail label or rework recommendations. The decision tree model comprehensively considers multiple factors, making logically clear reasoning and judgments, resulting in a more scientific and reasonable final inspection report. For qualified products, its quality level can be quickly determined; for unqualified products, specific rework recommendations can be provided to guide enterprises in effective rectification, reduce resource waste, and improve overall product quality. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based visual inspection method for quality inspection of safety shoes as described in this invention.

[0050] Figure 2 A flowchart for image data preprocessing;

[0051] Figure 3 Here is a flowchart of the image preprocessing process;

[0052] Figure 4 A flowchart for surface defect feature extraction;

[0053] Figure 5 Figures for feature extraction and feature dimensionality reduction are shown respectively.

[0054] Figure 6 A flowchart for classifying quality levels;

[0055] Figure 7 A flowchart for identifying defect-dense areas and generating heatmaps. Detailed Implementation

[0056] 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.

[0057] Please see Figure 1 This invention provides a method for quality inspection of safety shoes based on artificial intelligence visual inspection, the method comprising:

[0058] This process involves multi-angle image acquisition, preprocessing, feature extraction, quality grade classification, defect-dense area identification, and quality decision integration. First, industrial cameras capture images of the safety shoes from the top, sides, and bottom, forming a raw image data stream. Gaussian filtering and histogram equalization are applied to the raw image data to eliminate noise and lighting interference, generating a standardized image set. Then, a pre-trained convolutional neural network is used to extract surface defect features of the safety shoes, including crack depth, seam offset, and material wear areas, generating a set of visual feature vectors. Based on a preset quality threshold range, the safety shoes are divided into multiple quality grade intervals, generating preliminary quality classification results. Defect location information is fused, and a spatial clustering algorithm is used to identify defect-dense areas, calculate severity scores, and generate a quality assessment heatmap. Finally, a decision tree model integrates the quality assessment heatmap and preliminary quality classification results to output a final quality inspection report, including a pass / fail label or rework recommendations.

[0059] Example 1: See Figure 2 The industrial cameras are configured to capture multi-view images of the safety shoes from a fixed angle, including top, side, and bottom views, ensuring comprehensive coverage of the shoe's structural details and preventing the omission of critical areas such as the toe, upper, and sole joints. The industrial cameras utilize high-resolution CMOS sensors, positioned above and to the sides of the production line conveyor belt. The top camera is positioned vertically above, aligned with the overall shape of the shoe; the side cameras are set at a 90-degree angle, facing the length of the shoe, capturing seams and surface texture; the bottom camera requires a reflector or a rotating platform to capture the sole pattern and wear condition. The image capture process automatically begins after the safety shoes enter the inspection area, using photoelectric sensors to trigger the camera sequence, avoiding delays caused by manual intervention. The top view is captured first, obtaining the overall outline of the shoe upper; the side view follows, focusing on the side structure of the upper and toe; the bottom view is captured last, obtained by raising the platform or by the camera looking down. All images are acquired in real-time in RAW format, forming a raw image data stream; the system integrates a time-series stamping mechanism, associating the capture time with the image and recording it as timestamp information. The timestamp accuracy reaches the millisecond level, implemented based on the system's internal clock. Images are strictly arranged in the capture order: top, then side, then bottom, constructing a time-series image library for convenient subsequent processing and traceability. Image metadata includes ambient light intensity and camera shooting distance, measured by the camera's built-in sensor or auxiliary light sensor; light intensity is recorded in lux units, representing the ambient light intensity at the time of each image capture; shooting distance is measured using a laser rangefinder to ensure accurate recording of the distance between the camera and the shoe upper. This metadata serves as a key input for noise filtering, used in subsequent processing to dynamically adjust filtering parameters.

[0060] The raw image data stream enters the preprocessing stage, where noise filtering is performed using a Gaussian filtering algorithm. Gaussian filtering is implemented using a standard image processing library, with the kernel size adaptively adjusted according to the noise type. The process first reads each frame of the time-series image library, identifying outliers in the brightness value distribution, such as random noise or pixel distortion. For common lighting variations in industrial environments, the system calculates a global average brightness value and adjusts the baseline brightness of all images uniformly based on the light intensity in the metadata. The adjustment process refers to a preset brightness target value, calculates the brightness difference of each pixel, and applies a linear transformation to ensure consistent lighting across all images, reducing overexposure or underexposure. After Gaussian smoothing, a smoothed image set is generated, where each image retains its original viewpoint and time sequence to ensure data integrity. Subsequently, a size normalization operation is performed, scaling all images in the smoothed image set to a uniform pixel resolution size, with a target size set at 512 pixels wide by 512 pixels high. The scaling process uses bicubic interpolation to ensure size consistency without destroying image content, preventing scale differences from affecting feature extraction in subsequent analysis. After scaling, the images undergo histogram equalization to enhance contrast. This technique scans the pixel value distribution of each image, identifies low-contrast areas such as shadows or blurred textures, and then redistributes pixel values ​​to even out brightness distribution, ensuring that defects on the shoe surface, such as cracks or wear areas, are more clearly visible. After contrast enhancement, the images are stored in a standardized image group, with each image retaining its original timestamp, viewpoint type, and related metadata. Finally, all standardized images are integrated, grouped chronologically, and stored according to viewpoint type. The time series is ordered by capture time, and viewpoint types are categorized into top, side, and bottom. The data is stored in a structured array format in a database or written to an HDF5 file, forming a preprocessed image dataset. This dataset is directly linked to the feature extraction module, serving as the input source for subsequent convolutional neural network processing. Throughout the implementation, system configuration parameters such as target brightness values, scaling dimensions, and histogram parameters are fixed to avoid runtime variations that introduce processing biases. The image processing workflow employs a multi-threaded parallel processing approach to shorten preprocessing time and improve system throughput. The parallel processing mechanism is designed for multiple image channels to run simultaneously, with each channel corresponding to a sequence of viewpoints, maximizing process efficiency. After the preprocessed dataset is built, the data validation step includes random sampling to check image quality and consistency, preventing abnormal data from flowing into subsequent modules and affecting the overall detection reliability. All settings and metadata are integrated into the log file for system monitoring and error troubleshooting.

[0061] See Figure 3This demonstrates the complete image preprocessing workflow. The first row, from left to right, shows: the original image (including noise), the image after Gaussian filtering, and the image after brightness adjustment; the second row, from left to right, shows: the image after size normalization, the image after histogram equalization, and the final preprocessed result. The preprocessing effectively eliminates noise and illumination interference, enhances image contrast, and lays the foundation for subsequent feature extraction.

[0062] Example 2: See Figure 4 The preprocessed image dataset is loaded into a pre-trained convolutional neural network (CNN) model to perform feature extraction. During data loading, a batch input mechanism is used, with each batch containing sixteen standardized images organized into a four-dimensional tensor data structure and fed into the model's input layer. The CNN model architecture uses ResNet50 as its basic framework, and its pre-trained weights are initialized using the ImageNet dataset. The model input layer receives a 512×512 pixel three-channel RGB image, which undergoes feature response mapping through the first convolutional kernel. The convolutional layer is configured with a stride of 2, a kernel size of 7x7, and zero-padding mode to preserve boundary feature information. The ReLU non-linear activation function is used to enhance feature representation. The output feature map of the convolutional layer is passed to a batch normalization layer to adjust the mean and variance of the feature distribution. A max-pooling layer is applied to the normalized feature map, with a pooling window size of 3x3 and a stride of 2, compressing the feature dimension while preserving key texture information.

[0063] Multiple convolutional blocks process the input data sequentially, with each block containing a bottleneck structure. The bottleneck structure uses a 1x1 convolutional kernel to compress the number of channels, a 3x3 convolutional kernel to extract spatial features, and a final 1x1 convolutional kernel to restore the channel dimension. A residual connection mechanism runs through each convolutional block, allowing features from previous layers to be passed to deeper layers via cross-layer paths, avoiding the vanishing gradient problem. The feature maps output by each convolutional block have different receptive fields: shallow convolutional blocks extract edge and corner features, mid-layer convolutional blocks identify texture structures, and deep convolutional blocks capture global semantic features. During forward propagation, the feature activation maps of each convolutional layer are recorded and stored in a cache in real time, serving as the foundational data for the original feature mapping. The batch processing continues until all preprocessed image datasets have completed the model inference process.

[0064] The feature segmentation module locates key areas of the shoe body in the original feature map. Based on a predefined safety shoe component template, the model feature space is divided into the toe area, upper area, and sole area. The toe area is located in the top third of the feature map, corresponding to the actual toe protection area. The upper area covers the middle vertical area of ​​the feature map, mapping the shoe support structure. The sole area occupies the bottom quarter of the feature map, corresponding to the ground contact area of ​​the sole. Region segmentation uses a coordinate boundary cutting method to create a binary mask matrix in the feature tensor space. Each key region feature map is extracted by element-wise multiplication with the mask matrix to generate regularly sized feature sub-blocks. Each region feature sub-block independently enters the subsequent processing flow.

[0065] The region feature calculation performs texture gradient and color distribution analysis on each feature sub-block. Texture gradient calculation applies the Sobel edge detection operator to calculate the gradient components in the horizontal and vertical directions respectively. The gradient magnitudes in the two directions are fused to form the final gradient intensity map, reflecting the degree of surface irregularity. Color distribution parameter extraction is based on HSV color space conversion, segmenting the feature sub-blocks into three channels: hue, saturation, and brightness. The hue channel generates a color histogram to statistically analyze the distribution of major hues; the saturation channel calculates the mean and variance indices; and the brightness channel analyzes illumination consistency. These indices are encapsulated as mathematical vectors, forming a region feature subset data structure. The feature subsets of the three key regions are calculated and stored in parallel, maintaining the original spatial correspondence.

[0066] The feature reduction stage merges all regional feature subsets into a single feature matrix. Rows in the feature matrix correspond to different image samples, and columns contain all texture and color parameters for each region. Principal component analysis (PCA) processes the feature matrix, calculating the eigenvalues ​​and eigenvectors of the covariance matrix. Based on a preset cumulative variance contribution rate threshold, the number of principal components is selected to retain 95% of the information. The dimensionality-reduced features are projected into a lower-dimensional space to form compressed feature vectors. Each compressed feature vector contains a fixed-length numerical sequence that compactly represents the key features of each region of the shoe.

[0067] The defect location coordinates are recorded during the feature vector association stage. The original image coordinates of each feature point are extracted from the original feature map and mapped back to the original image pixel position using the feature map resolution. The feature point coordinate data is represented using a normalized coordinate system, with the X-axis corresponding to the proportional position in the image width direction and the Y-axis corresponding to the proportional position in the image height direction. The feature vector is compressed to expand the data dimension, and a position coordinate field is added to form a complete visual feature vector. The final generated set of visual feature vectors is output as a structured dataset, with each record containing feature vector data, position coordinate information, and an original image index identifier. The entire implementation utilizes a GPU parallel computing architecture for accelerated processing. The cuDNN library is used to optimize convolutional operation efficiency during the feature extraction stage, and the data pipeline uses DMA transfer to reduce host device communication overhead. The visual feature vector set is persistently stored in a distributed file system, allowing direct access and use by the subsequent quality level classification module.

[0068] See Figure 5 This demonstrates the key steps of feature extraction and dimensionality reduction. The left image is a schematic diagram of key region division, dividing the work boot into three key detection regions: the toe, the upper, and the sole. The right image is a feature dimensionality reduction distribution map, showing the distribution of feature points in two-dimensional space for the three regions. Circles represent the toe region, squares represent the upper region, and triangles represent the sole region. The feature points in different regions exhibit different distribution patterns, indicating that the extracted features can effectively distinguish between different regions.

[0069] Example 3: See Figure 6 The visual feature vector set enters the quality grade classification process, first by parsing the feature value data structure. Each visual feature vector contains multiple dimensions of numerical features, including key parameters such as crack depth index, suture offset, and material wear area. The crack depth index is extracted from texture gradient features, reflecting the severity of surface fracture; the suture offset is calculated through edge detection features, representing the precision of the sewing process; and the material wear area is derived based on color distribution features, quantifying the range of surface damage. These feature values ​​are matched and compared with preset quality threshold ranges, which are stored in a relational database and managed according to the safety shoe model. The database table structure includes fields such as shoe model number, maximum allowable crack length, minimum suture accuracy, and acceptable wear area ratio. The query process quickly retrieves the corresponding threshold parameters through the shoe model number index, establishing a mapping relationship between feature values ​​and quality standards.

[0070] The Support Vector Machine (SVM) classifier is configured for multi-class classification, and the radial basis function (RBM) is used as the kernel function. The classifier's input space consists of dimensionality-reduced feature vectors, and the output space defines five quality levels: excellent, acceptable, reworkable, borderline, and defective. The classification decision function is expressed as follows:

[0071]

[0072] in: This represents the input feature vector. For support vectors, It's a category label. Representing the Lagrange multipliers, For bias terms, It's a kernel function. During the training phase, historical sample data is used to optimize the model parameters, and the position of the classification hyperplane is adjusted in each update iteration. The classifier outputs the predicted label for each safety shoe sample, generating a list of raw classification results. The list record format includes fields such as sample ID, prediction quality level, and classification confidence.

[0073] The defect quantity statistics module analyzes the defect distribution within the visual feature vector set. For each sample, it calculates the frequency of occurrence of three main types of defects: crack quantity statistics are achieved through connected component analysis to identify discrete crack regions; suture line offset point statistics use edge continuity detection to record the number of positions deviating from the baseline; and wear area statistics are based on a clustering algorithm to distinguish independent wear patches. The statistical results are stored numerically and merged with the original classification labels to form an expanded dataset. New fields added to the dataset include quantitative indicators such as the number of crack regions, the number of offset points, and the number of wear patches.

[0074] The sample distribution density is calculated using kernel density estimation, constructing a probability density function in the feature space. The quality level intervals are divided into five dimensions, each corresponding to a feature distribution of a quality level. Gaussian kernel smoothing is used for density function calculation, with the bandwidth parameter determined empirically. The density value calculation process iterates through all sample points, statistically analyzing the point distribution concentration within each quality level region. The output results form a three-dimensional density distribution surface, visually representing the clustering characteristics of different quality levels in the feature space. Density data is stored in matrix format, with rows and columns corresponding to the feature dimensions and quality levels, respectively.

[0075] The classification result correction mechanism handles outliers in the original classification. Outliers are defined as isolated points in the feature space that are far from the cluster centers of samples of the same class. The detection algorithm calculates the Mahalanobis distance of each sample to the center of its quality class; samples exceeding three standard deviations are marked as points to be verified. A majority voting mechanism organizes an expert evaluation panel to manually review the outlier samples. The review is based on procedures including original image inspection and secondary verification of feature values. The final classification result update mechanism adopts a version control strategy, retaining a history of each correction. The corrected data generates a new classification label distribution map, showing the re-classification of samples in the feature space.

[0076] The quality threshold range database employs a dynamic update strategy. Database maintenance is set to weekly incremental updates, with new data sources including production line inspection results, customer feedback reports, and laboratory test records. The update process includes data cleaning steps to remove duplicate records and invalid measurements. Field validation rules ensure that numerical parameters are within reasonable ranges, such as crack length not exceeding 20% ​​of the total shoe length. Database version management uses timestamps and supports historical data backtracking. The query interface provides multiple access methods, including batch data export and real-time API calls.

[0077] The feature value and threshold comparison process is optimized for parallel computation. The comparison algorithm is designed in MapReduce mode, where each feature value is distributed to computing nodes during the mapping phase, and all comparison results are aggregated during the reduction phase. The computing nodes are configured with a load balancing strategy to dynamically allocate task queues. The real-time comparison engine processes thousands of samples per second, meeting the production line's inspection rate requirements. A comparison result caching mechanism stores recently queried records, reducing database access frequency. The output data structure includes fields such as matching status code, deviation from standard value, and pass / fail flags, encoded in binary format to improve transmission efficiency.

[0078] Feature engineering is implemented during the support vector machine (SVM) training data preparation phase. Historical sample datasets undergo preprocessing steps including missing value imputation, outlier removal, and numerical standardization. Feature selection algorithms evaluate the importance ranking of features across dimensions, removing redundant features with low relevance. The training and test sets are randomly partitioned in a 7:3 ratio, with a 50% cross-validation fold. Confusion matrix analysis is used to calculate classification performance parameters such as precision, recall, and F1 score. A grid search strategy is employed during hyperparameter tuning to explore the optimal combination of kernel function parameters. The trained model is serialized into a binary file and deployed to the production environment for classification inference.

[0079] The classification density distribution data visualization system features an interactive user interface. The visualization tool utilizes WebGL technology for 3D rendering and supports functions such as view rotation, zoom, and profile viewing. A color coding scheme maps different quality grades to distinct color schemes; for example, superior products are represented by a blue gradient, while defective products are marked with a red warning color. The distribution surface can be exported as a high-resolution image for use as an attachment in quality analysis reports. The user interface provides data filtering controls and supports dynamic view updates based on time period, production line number, shoe type, and other dimensions.

[0080] The classification result storage system adopts a distributed database architecture. Database nodes are deployed across multiple physical servers, and a data sharding strategy is implemented to improve access performance. Write operations follow the WAL log protocol to ensure transaction integrity. The index structure is optimized by creating a B+ tree index for the quality grade field, accelerating range query efficiency. Data backup strategy includes daily incremental backups and weekly full backups, with backup files encrypted and stored in an off-site disaster recovery center. Access control implements a role-based management mechanism, differentiating between different permission levels such as quality engineers, system administrators, and production line operators. Data export formats support multiple industry-standard formats such as CSV, JSON, and Parquet.

[0081] Example 4: See Figure 7 The defect location coordinate extraction process in the visual feature vector set is read from the structured data fields. Each feature vector record contains X and Y position values ​​in a normalized coordinate system, ranging from 0 to 1, representing the relative position of the defect point in the width and height directions of the image, respectively. Taking a batch of safety shoe inspection data as an example, the extracted defect coordinate data is shown in Table 1.

[0082] Table 1: Extracted defect coordinate data.

[0083]

[0084] The coordinate system of the standard 3D model for safety shoes is constructed based on computer-aided design drawings. The model origin is set at the geometric center of the shoe body, with the positive X-axis pointing towards the toe, the Y-axis extending along the width of the shoe body, and the Z-axis pointing vertically upwards. The model mesh achieves millimeter-level precision, containing approximately 50,000 triangular faces to fully represent the curved surface features of the shoe body. The coordinate transformation process maps the 2D image coordinates to the 3D model surface, using a ray casting algorithm to solve for the intersection points. For coordinate points in the sole area, additional perspective projection transformation compensation is considered to eliminate the effects of perspective distortion. The converted spatial position data recording format adds a 3D coordinate field while retaining the original image coordinates as a reference.

[0085] The quality grade labels in the preliminary quality classification results are associated with spatial location data using a hash index method. A unique identifier for each sample serves as the key, and a fast lookup table is built in memory. The label association process iterates through all defect records, matching the sample number with the corresponding quality grade. A quality grade field is added to the associated dataset, forming spatial distribution data with quality evaluation. A data validation step checks for any unmatched records to ensure that all defect points receive the correct quality label.

[0086] Density peak clustering algorithm processes spatial data with quality labels. The algorithm first calculates the local density value of each point, defined as the number of neighboring points within a sphere with a radius of 0.1 meters. Then, it determines the distance from the minimum distance of each point to a higher density point; the product of these two parameters serves as the clustering decision value. Cluster centers are automatically identified as points with prominent decision values, forming the initial cluster cores. Taking the inspection data of a pair of safety shoes as an example, seven defect points were detected in the toe area, three of which had cracks concentrated within a 2-centimeter radius, reaching the highest density value in the region and were marked as a densely defective area. The clustering process is executed iteratively until all defect points are assigned to a specific cluster or marked as noise points.

[0087] After generating the list of densely populated defect regions, the geometric features of each region are calculated. Region boundaries are determined using a convex hull algorithm, calculating the smallest convex polygon enclosing all member points. Region area is calculated based on the vertex coordinates of the polygon, accurate to the square millimeter. The average defect size is taken as the median of the eigenvalues ​​of all defects within the region, avoiding the influence of extreme values. Defect frequency is calculated by counting the number of defect points per unit area, reflecting the distribution density. These calculated parameters are recorded in the region description data structure as fundamental indicators for subsequent evaluation.

[0088] The severity score is calculated using a weighted scoring model. Different weighting coefficients are assigned to different defect types: cracks are weighted at 0.5, seam offset at 0.3, and wear at 0.2. Location importance coefficients are adjusted based on the functional position of the area within the shoe: toe protection area at 1.2, upper support area at 1.0, and sole contact area at 0.8. The scoring formula combines these parameters with regional feature values ​​to generate standardized scores ranging from 0 to 100. The scoring process batch processes all densely populated areas, and the results are sorted in descending order to generate a priority list.

[0089] The heatmap overlay image generation is based on 3D model surface rendering technology. The rendering engine divides the 3D model mesh into uniform pixel units, and each unit is assigned a color intensity based on the defect score within the area. The color mapping uses a red-yellow-green gradient spectrum, with red representing high-severity areas and green representing low-risk areas. The rendering resolution is set to four pixels per millimeter to ensure clear visibility of details. The overlay process preserves the geometric features of the original 3D model, and the heatmap is displayed as a semi-transparent layer over the model surface, supporting multi-angle viewing. The image output format is PNG, including an alpha channel to achieve transparency.

[0090] The data structure of the completed quality assessment heatmap contains multi-dimensional information. In addition to the basic image file, metadata records detailed information such as the actual physical location of each heatmap pixel, its severity score, and the combination of defect types involved. An interactive viewing tool enables click-to-search functionality, allowing users to select any area on the heatmap to view detailed assessment data. The heatmap version management system records the modification history of each update, supporting comparative analysis of detection results from different periods. The data export function supports the generation of standardized quality assessment reports, including heatmap thumbnails and key area analysis summaries.

[0091] A case study analyzing densely populated defect areas illustrates a typical application scenario. During the inspection of a batch of shoes, two densely populated defects were found in the toe area: the first contained three radial cracks, the longest being 8 mm, distributed within a 1.5 square centimeter area; the second was a series of stitch line offset points, involving five adjacent stitch lengths, with an offset of up to 2 mm. The system calculated the severity score for the first area to be 78 points, and for the second area 65 points. The heat map showed a distinct red area at the front of the toe, consistent with the actual defect location confirmed by visual inspection. Quality engineers located the problem area based on the heat map and further analysis revealed a correlation between this defect pattern and wear on the stamping die.

[0092] The maintenance and update mechanism for the 3D model coordinate system is regularly optimized for accuracy. Every quarter, a 3D scanner is used to digitize the physical shoe pattern, generating point cloud data for comparison with the existing model. When deviations exceed tolerance thresholds, the model's geometric parameter adjustment process is initiated. A coordinate reference system calibration report records the adjustments made during each maintenance, ensuring the continuity of historical data. Semantic version control is used for the model, and all relevant systems are notified to upgrade synchronously during major updates. This mechanism guarantees the consistency of spatial location data during long-term testing, avoiding analytical errors caused by model deviations.

[0093] The parameters of the spatial clustering algorithm can be adjusted according to actual inspection needs. For delicate components such as the toe cap protective steel plate area, a smaller cluster radius is set to identify clusters of minor defects; for large areas such as the sole, the radius is appropriately increased to capture distribution trends. The parameter adjustment interface provides a visual aid, displaying real-time previews of the clustering effects under different settings. Configuration schemes are saved according to shoe type and automatically match the preset parameters of the current inspection product upon loading. The historical parameter recording function supports retrospective analysis to help optimize algorithm performance in specific scenarios.

[0094] Performance optimization measures during heatmap generation improve processing efficiency. Image partitioning rendering technology divides the 3D model into multiple blocks, with distributed rendering nodes processing different areas in parallel. A caching mechanism stores intermediate calculation results, avoiding the repeated generation of the same heatmap layer. Dynamic detail level adjustment automatically selects the rendering precision based on view distance, and uses a simplified model to accelerate display when viewed from a distance. These technologies enable the average time to generate a complete heatmap on a standard workstation to be controlled within three seconds, meeting the response requirements of real-time detection. The system resource monitoring module dynamically adjusts the computing load, automatically queuing requests during peak periods to ensure service stability.

[0095] Example 5: The input interface of the decision tree model receives two data sources: a severity score sequence derived from a quality assessment heatmap and quality grade classification data from the preliminary quality classification results. The severity score sequence contains assessment parameters across multiple dimensions, each corresponding to a comprehensive score for a defect-dense region, and the data structure is a floating-point array. The quality grade classification data uses a discrete value encoding scheme, with superior to defective products labeled with numerical tags one through five. During data preprocessing, parameters of different dimensions are normalized, and all values ​​are scaled to the range of zero to one. The number of input layer nodes is dynamically configured based on the parameter dimensions, supporting simultaneous processing of up to one hundred input features. A data buffering mechanism sets queue capacity to adapt to changes in the inspection cycle on the production line.

[0096] A recursive segmentation algorithm is used to construct a decision tree structure. The root node selects the quality level as the initial segmentation feature, and the splitting threshold is searched for the optimal split point using a greedy algorithm. Each split calculates the decrease in impurity of the child nodes, and the feature with the largest information gain is selected for branching. The node segmentation strategy uses a binary search: the left subtree handles branches with parameters below the threshold, and the right subtree handles branches with parameters above the threshold. The tree depth is controlled to a preset maximum of ten levels to prevent overfitting. During the branching process, the severity score parameter is decomposed: the crack area score, the stitching offset score, and the wear area score are used to establish feature judgment nodes. Node decision rules are stored in inequality form; for example, "a crack score in the toe area greater than or equal to 0.7" triggers entry into a high-risk branch path.

[0097] The critical node analysis module traverses decision points along the decision path. The defect repairability node checks three elements: material damage depth, structural deformation amplitude, and defect location operability. The system internally defines a material damage depth tolerance index; leather damage exceeding 40% of its original thickness is considered irreparable. Structural deformation detection parameters measure key dimensional deviations in the shoe body; for example, forefoot width deformation exceeding three millimeters is marked as uncorrectable. The location operability database stores experienced repair data, marking special structural areas, such as the steel toe cap joint, as areas with high repair failure rates. The safety risk coefficient node integrates two types of parameters: defect impact intensity and protective function failure probability. Defect impact intensity is mapped based on laboratory pressure test data, establishing a correlation between severity scores and impact resistance attenuation values. The protective function failure probability uses the Monte Carlo method to simulate usage scenarios and calculate safety risk values ​​under different environmental conditions.

[0098] The overall quality score is calculated at the leaf nodes. The scoring model includes three factors: a quality grade base multiplied by 0.6, a defect repair potential coefficient multiplied by 0.25, and a safety risk index multiplied by 0.15. The quality grade base corresponds to a preset standard value: 100 for excellent products, 80 for qualified products, 60 for reworkable products, 40 for critical products, and 20 for scrap products. The defect repair potential coefficient ranges from 0 to 1.5, with the maximum value for fully repairable defects and zero for completely irreparable defects. The safety risk index is calculated using the reciprocal of a percentage system, with a risk assessment score of 90 or above counted as 1, 80 as 0.8, and so on. The weighted calculation result is rounded to the nearest integer to generate a decision score ranging from 0 to 100.

[0099] The pass / fail threshold comparison module is configured with dynamic judgment rules. The basic pass / fail score is set at 70 points, and the threshold can be adjusted according to different customer standards. The threshold comparison adopts a hierarchical decision-making mechanism: scores above 90 points are judged as ready for direct shipment, scores between 70 and 90 points trigger an additional inspection process, and scores below 60 points automatically enter the rework channel. The ambiguity zone where the score is close to the threshold is set to a width of 5 points. When the score falls into this range, a historical data review program is initiated. The review program searches for past handling records of similar defects and selects the most frequently occurring handling recommendations. The marking rules for rework recommendations include various handling instructions: minor rework recommendations include surface polishing or glue repair, moderate rework requires component replacement, and severe rework indicates complete shoe disassembly and reconstruction or scrapping. The judgment result status is stored in the central database, and control signals of the physical actuators are triggered simultaneously.

[0100] The structured report generation system organizes multi-source output information. The report template contains six standard paragraphs: the first paragraph states the inspection conclusion, clearly indicating "qualified" or "rework"; the second paragraph lists the locations of major defects using a three-dimensional coordinate description method; the third paragraph explains the basis for quality grade determination; the fourth paragraph analyzes the causes of key defects; the fifth paragraph proposes specific remedial measures; and the final paragraph includes the quality inspection certificate number. Remedial measures recommendations are generated as professional guidelines based on defect type: for crack defects, rubber hot-melt repair technology is recommended; for suture misalignment defects, re-suturing operation specifications are provided; and for wear defects, the wear-resistant coating spraying process is specified. The report layout system automatically adjusts the text and image layout, linking the defect location description paragraph to a three-dimensional heat map thumbnail, and inserting standardized operation diagrams in the remedial measures paragraph.

[0101] The system output interface supports multiple delivery modes. The main channel transmits data to the production line control system via industrial Ethernet. Qualified products trigger an automatic sorting mechanism, while reworked products activate alarm indicator lights. The human-machine interface displays a real-time summary of the inspection report, and operators can access detailed parameter data. A mobile device push mechanism sends inspection results to the quality engineer's workstation, supporting remote review and the issuance of emergency response instructions. The data archiving module creates permanent storage records, linking each inspection result to a complete chain of evidence, including original image data, feature vector data, and decision path logs, meeting quality traceability requirements. Periodic statistical functions generate batch quality analysis charts to help identify systemic production problems and optimize process parameter settings.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] 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 quality inspection of work shoes based on artificial intelligence visual inspection, characterized in that, The method includes: acquiring multi-angle image data of safety shoes, performing noise filtering and size standardization on the multi-angle image data to eliminate lighting interference and scale differences, and generating a preprocessed image dataset; Based on a preprocessed image dataset, a convolutional neural network is used to extract surface defect features of work shoes. The surface defect features include crack depth, stitch line offset, and material wear area, generating a set of visual feature vectors. Based on the set of visual feature vectors, safety shoes are divided into multiple quality level intervals, each of which corresponds to a preset quality threshold range, generating preliminary quality classification results. By integrating the preliminary quality classification results with the defect location information in the visual feature vector set, a spatial clustering algorithm is used to identify defect-dense regions, and the severity score of each defect-dense region is calculated to generate a quality assessment heatmap. Based on the quality assessment heat map and preliminary quality classification results, a decision tree model is used to integrate quality decisions and output a final quality inspection report for safety shoes. The final quality inspection report includes a pass mark or rework suggestion. The method based on the preprocessed image dataset uses a convolutional neural network to extract surface defect features of work shoes. The surface defect features include crack depth, suture line offset, and material wear area. A set of visual feature vectors is generated, including: loading the preprocessed image dataset into a pre-trained convolutional neural network model, extracting deep features of the image through multiple convolutional layers and pooling layers, and generating the original feature map. Key regions are segmented from the original feature map. The key regions include the toe, upper and sole. Texture gradient and color distribution parameters are calculated for each key region to generate a subset of region features. Merge the regional feature subsets of all key regions, apply principal component analysis algorithm to reduce dimensionality, retain the main feature components, and form a compressed feature vector; The compressed feature vectors are associated with the positional coordinates in the original feature maps to construct a set of visual feature vectors, which is then passed as input to the quality level classification module.

2. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 1, characterized in that, The process involves collecting multi-angle image data of the safety shoes, performing noise filtering and size standardization on the multi-angle image data to eliminate lighting interference and scale differences, and generating a preprocessed image dataset. This includes: acquiring top, side, and bottom view images of the safety shoes using an industrial camera to form a raw image data stream. The Gaussian filtering algorithm is applied to smooth the original image data stream, remove random noise points, and adjust the image brightness to a uniform reference value to generate a smooth image set. Each image in the smoothed image set is scaled to a fixed pixel size to ensure that all images have a consistent resolution. Then, histogram equalization is used to enhance image contrast, generating a standardized image set. A standardized image set is integrated, and the data is organized according to time sequence and viewpoint type to construct a preprocessed image dataset, which is directly used in subsequent feature extraction steps.

3. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 1, characterized in that, The process involves dividing work shoes into multiple quality grade intervals based on a set of visual feature vectors, with each quality grade interval corresponding to a preset quality threshold range, and generating preliminary quality classification results. This includes parsing the feature values ​​in the set of visual feature vectors and matching them with a preset quality threshold range, where the quality threshold range includes acceptable crack length and maximum wear area. A support vector machine classifier is used to divide the feature values ​​into multiple classes, outputting the quality level of each safety shoe sample and generating the original classification label; By aggregating the defect count statistics from the original classification labels and visual feature vector sets, the sample distribution density of each quality level interval is calculated to generate classification density distribution data; Based on the classification density distribution data, outliers in the original classification labels are corrected, and the final classification result is determined through a majority voting mechanism to form a preliminary quality classification result. This preliminary quality classification result is used to fuse defect location information.

4. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 1, characterized in that, The process of fusing preliminary quality classification results with defect location information from a visual feature vector set, identifying densely populated defect areas using a spatial clustering algorithm, calculating a severity score for each densely populated defect area, and generating a quality assessment heatmap includes: extracting defect location coordinates from the visual feature vector set and mapping them to the coordinate system of the workwear 3D model to generate a spatial location dataset; combining the quality level labels from the preliminary quality classification results, performing density peak clustering on the spatial location dataset to identify high-density defect clusters and generate a list of densely populated defect areas; calculating the average defect size and frequency for each densely populated defect area, and applying a weighted formula to obtain a severity score, whereby the severity score is based on defect type weight and location importance. By integrating the severity scores and spatial coordinates of all defect-dense regions, a heatmap overlay image is generated, and a quality assessment heatmap is constructed, which serves as the input source for the decision tree model.

5. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 1, characterized in that, Based on the quality assessment heatmap and preliminary quality classification results, a decision tree model is used to integrate quality decisions and output a final quality inspection report for the safety shoes. The final quality inspection report includes a pass mark or rework suggestion, including: inputting the severity score in the quality assessment heatmap and the quality level in the preliminary quality classification results into the decision tree model, and generating a decision path through a recursive segmentation algorithm; The key nodes in the decision-making path are evaluated, including the defect repairability and safety risk coefficients. The overall quality score is calculated, and decision scoring data is generated. Compare the decision score data with the preset pass threshold. If the overall quality score is lower than the pass threshold, it is marked as a rework suggestion; otherwise, it is marked as pass. The decision-making results are compiled into a structured report, which includes details of the defect location and recommended treatment measures, and the final quality inspection report of the safety shoes is output.

6. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 2, characterized in that, The process of acquiring top, side, and bottom view images of work shoes using an industrial camera to form a raw image data stream includes: setting the industrial camera to capture images of work shoes at a fixed angle to ensure coverage of all key viewpoints and generating a multi-view image sequence; Multi-view image sequences are timestamped, arranged in the order of capture, and a time-series image library is established. Image metadata, including light intensity and shooting distance, is extracted from a time-series image library for subsequent noise filtering.

7. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 1, characterized in that, The process of loading the preprocessed image dataset into the pre-trained convolutional neural network model and extracting deep features of the image through multiple convolutional and pooling layers to generate the original feature map includes: initializing the weights of the convolutional neural network model and optimizing the feature extraction layer using transfer learning techniques. The preprocessed image dataset is input into the model in batches, forward propagation calculation is performed, and feature activation maps are output. Aggregate feature activation maps to generate multi-scale feature representations, which serve as the basis for the original feature mapping.

8. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 3, characterized in that, The feature values ​​in the set of analytical visual feature vectors are matched with a preset quality threshold range, which includes acceptable crack length and maximum wear area. This includes: defining a quality threshold range database to store upper limit values ​​for different defect types. Query the feature values ​​in the visual feature vector set and compare them with the quality threshold range database in real time; The output comparison results are used as training data for the support vector machine classifier.

9. The method for quality inspection of safety shoes based on artificial intelligence visual inspection according to claim 4, characterized in that, The step of extracting the defect location coordinates from the visual feature vector set and mapping them to the coordinate system of the 3D model of the work shoe to generate a spatial location dataset includes: constructing a standard 3D model of the work shoe, wherein the standard 3D model includes a coordinate reference system; Transform the defect location coordinates to the 3D model coordinate system to ensure spatial consistency; Verify the accuracy of coordinate transformation and generate a calibrated spatial location dataset.

Images