A simple automated clothing sorting method and system based on artificial intelligence

Abstract

This invention discloses a simple automated clothing sorting method and system based on artificial intelligence, comprising: collecting multi-dimensional visual information of clothing, constructing a three-dimensional feature matrix through a clothing texture depth perception model, and generating standardized feature vectors through a cross-scene clothing adaptation learning model; completing category prediction through a clothing sorting intelligent data analysis platform, constructing a multi-objective optimization scheduling model by calling a multi-task sorting scheduling and balancing algorithm, allocating tasks to execution units to complete physical sorting, and finally dynamically updating the parameters of each model through error feedback data. The system achieves accurate and efficient clothing sorting through multi-unit collaborative operation, adapting to sorting needs in multiple scenarios. This invention, through the organic combination of feature depth perception, cross-scene adaptation, intelligent scheduling, and dynamic iteration, ensures both the accuracy of sorting and classification and improves overall sorting efficiency, meeting the needs of large-scale, multi-scenario clothing sorting.

Description

Technical Field

[0001] This invention relates to the field of automatic clothing sorting technology, and in particular to a simple automatic clothing sorting method and system based on artificial intelligence. Background Technology

[0002] With the continuous growth of large-scale production and logistics sorting demands in the apparel industry, the efficiency and accuracy of garment sorting have become key factors affecting the supply chain turnover rate. Traditional garment sorting relies on manual classification, which is affected by subjective judgment, labor intensity, and skill level, making it difficult to adapt to multi-category, large-volume sorting scenarios and prone to classification errors. To address this issue, AI-based automated sorting technology has gradually emerged, achieving unmanned sorting processes through visual recognition, model learning, and intelligent scheduling. However, existing technologies mostly focus on single-scenario adaptation and lack dynamic response capabilities to changes in different garment materials, textures, and sorting environments. Furthermore, there is room for improvement in the synergy of multi-task scheduling and model iterative optimization. There is an urgent need to build a complete technical solution integrating deep feature perception, cross-scenario adaptation, intelligent scheduling, and data feedback to meet the practical needs of simplified deployment and efficient sorting.

[0003] Existing technologies have two significant drawbacks: First, the accuracy of clothing feature recognition is insufficient. Traditional models rely heavily on two-dimensional feature extraction of surface texture, failing to fully integrate three-dimensional features such as thickness distribution and contour edges. This leads to confusion when classifying clothing with similar textures but different materials, and lacks cross-scenario learning capabilities. When the sorting environment or clothing category changes, feature adaptability decreases, and classification errors increase. Second, the balance of multi-task sorting scheduling is lacking. Existing algorithms often only consider single factors such as task priority or equipment load, without establishing a multi-objective optimization mechanism that considers category confidence, real-time equipment status, and task execution efficiency. This results in unreasonable allocation of sorting resources, with some execution units running overloaded while others are idle, limiting overall sorting efficiency. Furthermore, the lack of a linkage feedback mechanism with the feature recognition model makes it impossible to dynamically optimize model parameters based on sorting results, hindering continuous improvement in sorting performance. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a simple automatic clothing sorting method and system based on artificial intelligence.

[0005] The technical solution adopted in this invention is a simple automatic clothing sorting method based on artificial intelligence, comprising the following steps: S1, acquiring multi-dimensional visual information of the clothing to be sorted through an image acquisition module, and transmitting the multi-dimensional visual information to a clothing texture depth perception model, wherein the model constructs a three-dimensional feature matrix based on clothing surface texture features, thickness distribution features, and contour edge features; S2, performing feature mapping processing on the three-dimensional feature matrix using a cross-scene clothing adaptation learning model, and generating a standardized feature vector adapted to the current sorting scene by combining historical feature data of clothing sorting in different scenes; S3, inputting the standardized feature vector into a clothing sorting intelligent data analysis platform, and performing preliminary classification through the feature classification rule library built into the platform. S4. Output the clothing category prediction result and the corresponding confidence index; S5. Call the multi-task sorting scheduling and balancing algorithm to construct a multi-objective optimization scheduling model based on the category prediction result, confidence index and real-time operating status parameters of the sorting equipment; S6. Allocate sorting tasks to the corresponding execution units through the multi-objective optimization scheduling model. The execution units adjust their operating parameters according to the task instructions to complete the physical sorting operation of the clothing; S7. Collect the actual category information of the sorted clothing, compare the actual category information with the prediction result, generate error feedback data, and transmit the error feedback data to the clothing texture depth perception model, the cross-scene clothing adaptation learning model and the multi-task sorting scheduling and balancing algorithm respectively for dynamic updating of model parameters.

[0006] Furthermore, the feature extraction expression of the clothing texture depth perception model is as follows: ,in, Output values ​​for the three-dimensional feature matrix of clothing. Let the two-dimensional distribution function of the surface texture of the clothing be denoted as . Let the thickness distribution function of clothing be . For the three-dimensional feature function of the clothing outline edge, These are the feature weight coefficients. For texture depth activation function, Adjust the parameters for the activation function.

[0007] Furthermore, the scheduling allocation expression of the multi-task sorting scheduling balancing algorithm is as follows: ,in, Let be the scheduling priority coefficient for the i-th sorting task. Predict the confidence level for the category of the i-th task. For the real-time running efficiency of the j-th execution unit, Let be the current task load of the k-th execution unit. Let be the execution path length parameter for the i-th task. For the balance adjustment coefficient, Let be the urgency parameter for the i-th task.

[0008] Furthermore, the feature mapping expression of the cross-scene clothing adaptation learning model is: ,in, To standardize the feature vector, For characteristic nonlinear mapping functions, The initial three-dimensional feature matrix, Adapt the matrix for the current scene. This is a correction factor for historical data. The number of historical scene samples. Let be the feature vector of the nth historical scene.

[0009] Furthermore, the category prediction expression of the intelligent data analysis platform for clothing sorting is: ,in, Predictions for clothing categories, The number of categories in the rule base. Let be the weight coefficient of the k-th category. This is the function for calculating feature similarity. This is the standard feature vector of the k-th category in the rule base.

[0010] Furthermore, the expression for the dynamic update of the model parameters is: ,in, For the updated model parameters, These are the model parameters before the update. For learning rate, This is the gradient value of the loss function. For the actual category label of the clothing, To predict category labels, This is a correction factor for the operating status of the sorting equipment.

[0011] Further, step S3 includes the following sub-steps: S31, extracting the calibration feature dimensions from the standardized feature vector, the calibration feature dimensions including the texture feature dominant dimension, the thickness feature core dimension, and the contour feature calibration dimension, and assigning feature importance weights to each dimension according to a preset weight allocation rule; S32, calling the rule base matching module of the clothing sorting intelligent data analysis platform, comparing the weighted calibration feature dimensions with the feature threshold ranges corresponding to each category in the rule base one by one, and recording the category identifiers and matching degree values ​​of successfully matched categories; S33, sorting according to the matching degree values, selecting the top three categories with the highest matching degree as candidate categories, and calculating the comprehensive confidence of each candidate category, the comprehensive confidence being calculated by weighting the feature matching degree, the historical similar sorting accuracy, and the current scene adaptability; S34, determining the candidate category with the highest comprehensive confidence as the clothing category prediction result, and simultaneously outputting the confidence index corresponding to the prediction result and the relevant information of the other two candidate categories.

[0012] Further, step S4 includes the following sub-steps: S41, collecting real-time operating status parameters of the sorting equipment, including the current task queue length, operating power, response latency, and fault warning status of each execution unit, and constructing an equipment status matrix; S42, fusing the clothing category prediction results, confidence index, and equipment status matrix to determine the constraints of the multi-objective optimization scheduling model, including task execution time constraints, equipment load balancing constraints, and sorting accuracy constraints; S43, setting an objective function based on a multi-task sorting scheduling balancing algorithm, including three sub-objectives: maximizing task completion efficiency, minimizing equipment energy consumption, and minimizing sorting error; S44, solving the multi-objective optimization scheduling model using the gradient descent method to generate the optimal matching scheme between each sorting task and execution unit, and the corresponding task execution parameters.

[0013] Further, S5 includes the following sub-steps: S51, converting the optimal matching scheme into control instructions recognizable by the execution unit, the control instructions including task number, target sorting area coordinates, action execution timing, and force adjustment parameters; S52, after receiving the control instructions, each execution unit initializes and adjusts its own operating parameters, calibrating the detection benchmarks of the position sensor, force sensor, and speed sensor; S53, according to the timing requirements of the control instructions, the execution unit starts the mechanical transmission structure, and completes the physical transfer of clothing through grasping, transporting, and placing actions, collecting action execution data in real time during the process; S54, after the action is completed, the execution unit feeds back the task completion signal to the multi-task sorting scheduling and balancing algorithm, and simultaneously uploads the action execution data to the clothing sorting intelligent data analysis platform for storage.

[0014] A simple automated clothing sorting system based on artificial intelligence is disclosed. This system, applied to an AI-based method for automated clothing sorting, includes: a multi-dimensional visual information acquisition unit for acquiring images and depth information of the clothing to be sorted, establishing a data transmission connection with a clothing texture depth perception model, and transmitting the acquired raw information to the model for feature extraction; a cross-scene feature adaptation processing unit that communicates bidirectionally with the clothing texture depth perception model, receiving the three-dimensional feature matrix output by the model, and completing feature mapping through a built-in cross-scene clothing adaptation learning model to generate standardized feature vectors; and an intelligent data analysis and prediction unit connected to the cross-scene feature adaptation processing unit, integrating an intelligent data analysis platform for clothing sorting, and classifying the standardized feature vectors. The system outputs category prediction results and confidence indices; a multi-task scheduling optimization unit communicates with the intelligent data analysis and prediction unit and the sorting execution unit, respectively, constructs an optimization model based on the multi-task sorting scheduling balancing algorithm, allocates sorting tasks, and outputs control commands; a mechanical sorting execution unit receives control commands from the multi-task scheduling optimization unit, completes the physical sorting operation of clothing through mechanical structures, and simultaneously collects the actual category information after sorting; a model parameter dynamic update unit establishes data interaction channels with the mechanical sorting execution unit, the clothing texture depth perception model, the cross-scene clothing adaptation learning model, and the multi-task scheduling optimization unit, receives error feedback data between the actual category information and the prediction results, adjusts the parameters of each model and algorithm in real time, and continuously optimizes the system's sorting performance.

[0015] Beneficial Effects: This invention proposes a simple automated clothing sorting method and system based on artificial intelligence. At the method level, a clothing texture depth perception model integrates three-dimensional features of clothing surface texture, thickness distribution, and contour edges, overcoming the limitations of traditional two-dimensional feature extraction. Combined with a cross-scene clothing adaptation learning model that leverages historical data, it enhances the flexibility of feature adaptation across different scenarios, significantly reducing the probability of confusion between similar clothing categories. A multi-task sorting scheduling and balancing algorithm integrates multi-dimensional parameters such as category confidence and real-time equipment status to construct an optimized model, achieving precise matching between sorting tasks and execution units, thus resolving efficiency bottlenecks caused by unreasonable resource allocation. At the system level, each unit forms a closed loop through data interaction. The model parameter dynamic update unit continuously optimizes the performance of each model using sorting error feedback, further improving the stability of recognition and scheduling. The overall solution, through the organic combination of feature depth perception, cross-scene adaptation, intelligent scheduling, and dynamic iteration, ensures both the accuracy of sorting and classification and improves overall sorting efficiency, meeting the needs of large-scale, multi-scenario clothing sorting. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention;

[0017] Figure 2 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, a simple automated clothing sorting method based on artificial intelligence includes the following steps:

[0020] S1, acquire multi-dimensional visual information of the clothes to be sorted through the image acquisition module, and transmit the multi-dimensional visual information to the clothing texture depth perception model. The model constructs a three-dimensional feature matrix based on the clothing surface texture features, thickness distribution features and contour edge features.

[0021] Specifically, step S1 involves using an image acquisition module composed of a 12-megapixel high-definition industrial camera with a frame rate of 60 frames per second and a laser depth sensor to capture omnidirectional visual information of the garments to be sorted moving at a constant speed (controlled at 0.5 meters per second) on a conveyor belt. The acquisition range covers the front, back, and sides of the garments. The acquired visual information includes a 4096×3072 resolution two-dimensional texture image, thickness distribution data with an accuracy of 0.1 mm, and contour edge coordinate information with a sampling interval of 0.5 mm. During the acquisition process, a synchronous triggering mechanism ensures that the timestamp error between the image and depth data does not exceed 10 milliseconds. All acquired raw data is transmitted via gigabit Ethernet to the preprocessing module of the garment texture depth perception model, with the transmission delay controlled within 50 milliseconds. The model incorporates a feature extraction network that uses 128 convolutional kernels (3×3, 5×5, and 7×7 sizes) to extract multi-scale features from 2D texture images. It also combines thickness data from a laser depth sensor to construct a texture feature vector encompassing three dimensions: texture grayscale value, texture density, and texture direction. Based on the thickness distribution data, the clothing area is divided into 5×5 mm grids, and the average thickness of each grid is calculated to form a 2D thickness distribution feature matrix. An edge detection operator processes the contour edge coordinate information, extracting feature parameters such as contour curvature, edge slope, and contour closure. Finally, the texture feature vector, thickness distribution feature matrix, and contour edge feature parameters are fused to construct a 2048×1024 3D feature matrix, providing comprehensive and accurate basic data support for subsequent feature processing. This step ensures the integrity and accuracy of the clothing feature information, providing a reliable data foundation for subsequent sorting decisions.

[0022] S2, use the cross-scene clothing adaptation learning model to perform feature mapping processing on the three-dimensional feature matrix, and combine the historical feature data of clothing sorting in different scenarios to generate a standardized feature vector adapted to the current sorting scenario.

[0023] Specifically, during step S2, the cross-scene clothing adaptation learning model first receives the three-dimensional feature matrix output from step S1. This model includes three core modules: a feature mapping layer, a scene adaptation layer, and a feature normalization layer. In the feature mapping layer, a multilayer perceptron structure is used to transform the dimensionality of the three-dimensional feature matrix. Three fully connected layers (each with 4096, 2048, and 1024 neurons respectively) map the high-dimensional feature matrix into a low-dimensional feature vector. Batch normalization is used during the mapping process to ensure that the mean of the feature data is controlled near 0 and the variance is controlled around 1. The scene adaptation layer has a built-in scene feature database. This database stores historical feature data for 100 typical sorting scenarios (including garment factory outbound sorting, express logistics transit sorting, e-commerce return sorting, etc.). Each scenario includes no fewer than 100,000 clothing feature samples, covering 30 common clothing materials such as cotton, linen, silk, and synthetic fibers, and 20 clothing styles such as long-sleeved, short-sleeved, jackets, and pants. The model calculates the similarity between the current 3D feature matrix and feature samples from various historical scenes, selecting the top 5 scenes with the highest similarity as reference scenes. Feature adaptation parameters for these reference scenes are extracted, including feature weight coefficients and feature offsets. Subsequently, the scene adaptation layer adjusts the mapped low-dimensional feature vector based on these parameters. During the adjustment process, a scene adaptation factor (ranging from 0.1 to 0.9, dynamically allocated based on the similarity between the current scene and the reference scenes) is introduced to differentiate the weighting of features in different dimensions. Finally, the feature normalization layer uses the Min-Max normalization method to compress the adapted feature vector to the 0-1 range, generating a standardized feature vector with a dimension of 1024. This vector eliminates the influence of feature scale differences across different scenes, ensuring the consistency and comparability of feature data, providing standardized input for subsequent category prediction, and effectively improving the model's adaptability and robustness in different sorting scenarios.

[0024] S3, input the standardized feature vector into the intelligent data analysis platform for clothing sorting, perform preliminary classification through the feature classification rule library built into the platform, and output the clothing category prediction result and the corresponding confidence index.

[0025] Specifically, during step S3, the standardized feature vector generated in step S2 is first transmitted to the intelligent data analysis platform for clothing sorting via a data interface. This platform is equipped with a high-performance processor (3.5GHz, 8 cores, 16 threads) and a 128GB high-speed cache to ensure that data processing latency does not exceed 30 milliseconds. The platform's built-in feature classification rule library includes a set of multi-classification models built based on decision trees, random forests, and support vector machines. The rule library stores feature judgment thresholds for different clothing categories, including texture feature thresholds, thickness feature thresholds, contour feature thresholds, and more than 500 judgment rules. Each rule corresponds to a specific combination of features for a particular clothing category. After the standardized feature vector is input, the platform first compares the data of each dimension of the feature vector with the judgment rules in the rule library one by one through the rule matching engine. The comparison order is executed from high to low rule priority (priority is set according to historical matching accuracy, with rules with an accuracy of ≥95% set as the highest priority). During the matching process, the matching result (successful or unsuccessful matching) and matching score (score range 0-100 points, calculated based on feature fit) of each rule are recorded. Subsequently, the platform performs a comprehensive calculation on all matching results, using a weighted summation method (60% weighting for rule priority and 40% weighting for matching score) to obtain a comprehensive matching value for each clothing category. The top three categories with the highest comprehensive matching values ​​are then selected as candidate categories. Simultaneously, the platform calculates a confidence index based on the comprehensive matching value of each candidate category and its historical classification accuracy (the ratio of the number of correct classifications for that category to the total number of classifications in historical data). The confidence index is normalized to ensure its value ranges between 0 and 1, with the comprehensive matching value accounting for 70% and the historical classification accuracy accounting for 30%. Finally, the platform outputs the clothing category prediction result (i.e., the candidate category with the highest comprehensive matching value) and its corresponding confidence index, while also storing the comprehensive matching values ​​and confidence index data for the other two candidate categories. This step, through multi-rule collaborative judgment and quantitative evaluation of confidence, improves the accuracy of category prediction, providing a reliable decision-making basis for subsequent sorting and scheduling, and effectively reducing the risk of sorting errors caused by misjudgments.

[0026] S4, invoke the multi-task sorting scheduling and balancing algorithm, and construct a multi-objective optimization scheduling model based on the category prediction results, confidence index and real-time operating status parameters of the sorting equipment;

[0027] Specifically, after step S4 is initiated, the system first calls the multi-task sorting scheduling and balancing algorithm. This algorithm runs on a server with parallel computing capabilities (equipped with 4 GPUs, each with 16GB of video memory) to ensure the algorithm's solution efficiency. The algorithm first receives the clothing category prediction results and confidence index output from step S3. At the same time, it collects the real-time operating status parameters of the sorting equipment through the industrial bus at a frequency of 10 times per second. The collected parameters include the current task queue length (range 0-20 tasks), operating power (range 300-800 watts), response latency (range 5-50 milliseconds), and fault warning status (0 indicates normal, 1 indicates level 1 warning, and 2 indicates level 2 warning) of each of the 8 execution units. These parameters are then organized into an 8×4 equipment status matrix. Subsequently, based on the category prediction results, the algorithm determines the sorting process requirements corresponding to different categories of clothing (e.g., fragile clothing needs to be handled gently, and heavy clothing needs to be gripped with greater force). It then sets sorting accuracy thresholds based on confidence levels (99% for confidence levels ≥ 0.9; 97% for confidence levels between 0.7 and 0.9; and 95% for confidence levels < 0.7). Based on this information, the algorithm constructs a multi-objective optimization scheduling model. The model's objective function includes three core dimensions: first, maximizing task completion efficiency, measured by the average execution time of all tasks (target value ≤ 3 seconds); second, balancing equipment load, measured by the variance of the task queue length of each execution unit (target value ≤ 5); and third, minimizing energy consumption, measured by the total operating power of all execution units (target value ≤ 4000 watts). The model's constraints include a maximum load limit for the execution unit (a single execution unit can process a maximum of 20 tasks simultaneously), a task execution time limit (the execution time of a single task ≤ 5 seconds), and a sorting accuracy constraint (it must meet the accuracy threshold at the corresponding confidence level). The algorithm solves the problem iteratively (the number of iterations is set to 100, and the convergence accuracy is set to 10^10). -4 The optimal allocation scheme for each sorting task is determined, and the execution unit, execution order and execution parameters corresponding to each task are clarified. This step achieves reasonable allocation of sorting resources through multi-dimensional parameter fusion and multi-objective optimization design, avoids equipment overload or idleness, and provides algorithmic support for improving overall sorting efficiency.

[0028] S5, the sorting tasks are assigned to the corresponding execution units through the multi-objective optimization scheduling model, and the execution units adjust the running parameters according to the task instructions to complete the physical sorting operation of the clothing;

[0029] Specifically, step S5 begins with the sorting task allocation scheme output by the multi-objective optimization scheduling model. This scheme is transmitted via industrial Ethernet to the controllers of each execution unit (using PLC controllers with a response time ≤10 milliseconds). The allocation scheme includes task number, target sorting area coordinates (accuracy ±1 mm), action execution sequence (error ≤5 milliseconds), and force adjustment parameters (value range 5-50 Newtons, dynamically adjusted according to the fabric texture). After receiving the task instruction, each execution unit first initializes its own operating parameters, including calibrating the initial position of the robotic arm to the origin coordinates (error ≤0.5 mm), setting the initial force of the gripping mechanism to a reference value (10 Newtons), and adjusting the operating speed of the conveyor to match the task allocation scheme (0.3-0.8 meters per second). Subsequently, the execution unit activates the position sensor, force sensor, and speed sensor (sensor sampling frequency of 100 times per second, measurement accuracy of ±0.1 mm, ±0.5 Newtons, and ±0.01 meters per second, respectively) to monitor the operating status in real time. The robotic arm plans its movement path according to the coordinates of the target sorting area, and uses a trapezoidal acceleration / deceleration algorithm to control its speed. During acceleration, the acceleration is set to 0.5 m / s², during constant speed, the speed is adjusted according to the task distance (0.5-1.0 m / s), and during deceleration, the acceleration is set to -0.5 m / s² to ensure smooth, impact-free movement. The gripping mechanism adjusts the clamping force according to the force adjustment parameters. For soft clothing such as cotton and silk, the clamping force is controlled between 5-15 Newtons; for heavy clothing such as denim and canvas, the clamping force is controlled between 30-50 Newtons. After gripping, the robotic arm transfers the clothing to the target sorting area according to the planned path. During the transfer, it receives real-time sensor feedback data. If a force deviation exceeds ±2 Newtons or a position deviation exceeds ±1 mm, the controller immediately adjusts the operating parameters for correction. Upon reaching the target area, the gripping mechanism releases the clothing to complete the placement operation. During placement, the release speed is controlled to ≤0.2 m / s to avoid damage to the clothing. Throughout the entire physical sorting operation, each time the execution unit completes an action step, it sends a status signal back to the multi-task sorting scheduling and balancing algorithm to ensure the traceability of task execution. This step, through precise instruction execution and real-time status monitoring, ensures the stability and security of clothing sorting, and achieves efficient implementation from task allocation to physical sorting.

[0030] S6. Collect the actual category information of the sorted clothing, compare the actual category information with the prediction result, generate error feedback data, and transmit the error feedback data to the clothing texture depth perception model, the cross-scene clothing adaptation learning model and the multi-task sorting scheduling balance algorithm respectively for dynamic updating of model parameters.

[0031] Specifically, step S6 begins by using RFID readers (with a recognition distance of 0-10 cm and a recognition accuracy of 99.9%) installed at the exits of each sorting area to collect the actual category information of the sorted clothing. These readers communicate with the RFID tags (storing core information such as clothing category, material, and size) pre-attached to the clothing, with a reading rate of 10 times per second and a reading latency of ≤5 milliseconds, ensuring accurate capture of information for each sorted garment. Subsequently, the system transmits the collected actual category information and the predicted results output in step S3 to the data comparison module. This module uses a string matching algorithm to compare the two types of information field by field. The comparison dimensions include three core attributes: clothing category, material, and size. If all three attributes are completely identical, the match is considered successful; if any attribute is inconsistent, the match is considered unsuccessful, and the inconsistent attribute type and its specific value are recorded. Based on the comparison results, the system generates error feedback data, which includes the matching success rate (the ratio of successfully matched clothing to the total number of clothing sorted), the misjudgment rate for each category of clothing (the ratio of misjudged clothing of a certain category to the total number of clothing sorted in that category), the distribution of misjudged attributes (the proportion of misjudgments for each attribute to the total number of misjudgments), and the corresponding feature deviation value (the difference between the actual features and the features used for prediction). The error feedback data is transmitted via dedicated data channels to the clothing texture depth perception model, the cross-scene clothing adaptation learning model, and the multi-task sorting scheduling and balancing algorithm, with a transmission frequency of once after every 100 pieces of clothing are sorted. The clothing texture depth perception model adjusts the convolution kernel weights and feature extraction thresholds based on the feature deviation value; the cross-scene clothing adaptation learning model updates the scene adaptation factor and historical feature database based on the misjudgment rate; and the multi-task sorting scheduling and balancing algorithm adjusts the weight coefficients of the optimization target and the constraint thresholds based on the matching success rate. During parameter updates, an incremental update strategy is adopted, with the parameter change range controlled within 5% for each update to avoid drastic fluctuations in model performance. Through continuous error feedback and parameter optimization, the performance of each model and algorithm is dynamically improved, ensuring that the system maintains stable sorting accuracy and efficiency during long-term operation.

[0032] Preferably, the feature extraction expression of the clothing texture depth perception model is: ,in, Output values ​​for the three-dimensional feature matrix of clothing. Let the two-dimensional distribution function of the surface texture of the clothing be denoted as . Let the thickness distribution function of clothing be . For the three-dimensional feature function of the clothing outline edge, These are the feature weight coefficients. For texture depth activation function, Adjust the parameters for the activation function.

[0033] Specifically, the feature extraction expression of the clothing texture depth perception model achieves accurate output of the three-dimensional feature matrix of clothing through multi-dimensional feature fusion. Its implementation revolves around three core features: clothing surface texture, thickness distribution, and contour edges. In the feature acquisition stage, a high-definition image sensor captures the two-dimensional distribution information of the clothing surface texture, a laser depth sensor acquires clothing thickness data, and a visual imaging system acquires the three-dimensional coordinate information of the contour edges. The acquisition frequency of the three types of raw data is kept consistent to ensure feature synchronization. The feature weight coefficients in the model are dynamically allocated according to the feature importance of different clothing categories. After determining the initial values ​​through training with a large number of samples, adjustments can be made based on error feedback from actual sorting scenarios. The texture depth activation function is used to enhance the response intensity of effective features. The adjustment parameters are determined through iterative optimization algorithms, improving the model's ability to recognize subtle texture differences. During implementation, the three types of feature data are first preprocessed by substituting them into their corresponding functions, then weighted and summed according to the weight coefficients. The activation function performs a non-linear transformation on the summation result, ultimately outputting a three-dimensional feature matrix that comprehensively characterizes the physical features of the clothing. This expression breaks through the limitations of traditional single feature extraction. It improves the completeness and accuracy of feature representation through the synergistic effect of multiple features, providing high-quality feature input for subsequent cross-scene adaptation and category prediction. It ensures that the features of clothing of different materials and styles can be fully captured, laying the foundation for sorting accuracy.

[0034] Preferably, the scheduling allocation expression of the multi-task sorting scheduling balance algorithm is: ,in, Let be the scheduling priority coefficient for the i-th sorting task. Predict the confidence level for the category of the i-th task. For the real-time running efficiency of the j-th execution unit, Let be the current task load of the k-th execution unit. Let be the execution path length parameter for the i-th task. For the balance adjustment coefficient, Let be the urgency parameter for the i-th task.

[0035] Specifically, the scheduling allocation expression of the multi-task sorting scheduling balancing algorithm achieves scientific determination of sorting task priorities through multi-dimensional parameter integration and balancing calculation. Its implementation relies on the real-time data acquisition and processing capabilities of the sorting system. In the parameter acquisition phase, the category prediction confidence level directly comes from the output results of the clothing sorting intelligent data analysis platform, reflecting the reliability of the category judgment in real time. The real-time operating efficiency of the execution unit is calculated by monitoring indicators such as the task completion rate and failure rate of the equipment. The current task load is determined comprehensively based on the number and complexity of tasks being processed by each execution unit. The execution path length parameter is obtained by calculating the straight-line distance between the task starting point and the target sorting area and the path obstacle coefficient using the system's built-in spatial positioning module. The urgency parameter is set based on factors such as the time requirements of clothing orders and the connection requirements of the sorting process. The balancing adjustment coefficient is determined through training with a multi-objective optimization algorithm to balance the influence of each parameter on the scheduling priority, avoiding scheduling imbalance caused by a single factor. During implementation, various parameters are first collected synchronously and standardized to eliminate dimensional differences. Then, according to the calculation logic of the expression, the numerator integrates parameters related to task importance, and the denominator integrates parameters related to equipment status. The ratio of the two is corrected by combining the square root of the urgency parameter, ultimately obtaining the scheduling priority coefficient for each sorting task. This expression establishes a linkage mechanism between task characteristics and equipment status, achieving optimal allocation of sorting tasks through quantitative calculation. This ensures that high-confidence tasks are executed first, equipment load is evenly distributed, path loss is minimized, and the overall operating efficiency and resource utilization of the sorting system are improved.

[0036] The scheduling and allocation expression of the multi-task sorting and scheduling balancing algorithm has consistent dimensions. All parameters involved in the calculation are standardized, and the calculation logic conforms to the dimensional coordination rule. First, the category prediction confidence in the numerator is a dimensionless index (range 0-1) calculated based on feature matching degree and historical accuracy. The real-time operating efficiency of the execution unit is a relative value (dimensionless) obtained by the ratio of actual operating data to rated efficiency. After weighted summation by balancing adjustment coefficients (dimensionless weighting factors that only adjust the degree of influence), the overall result still retains the dimensionless attribute. Second, the current task load of the execution unit in the denominator is a relative value (dimensionless) normalized to the maximum load threshold. The task execution path length parameter is a standardized value (dimensionless) obtained by conversion with the system's maximum path length. After weighted summation by another set of dimensionless balancing adjustment coefficients, the result is also dimensionless. Finally, the urgency parameter is a standardized priority quantification value (dimensionless, range 0-1). Its square root operation does not introduce a new dimension, but only performs nonlinear correction on the value. In summary, dividing the dimensionless numerator by the dimensionless denominator results in a dimensionless result. Multiplying this result by the square root of the dimensionless urgency coefficient yields a scheduling priority coefficient that remains dimensionless throughout the process. Standardization eliminates the original dimensional differences among the parameters, and weighted, multiplicative, and division operations do not disrupt dimensional coordination, ensuring dimensional uniformity in the expression.

[0037] Preferably, the feature mapping expression of the cross-scene clothing adaptation learning model is: ,in, To standardize the feature vector, For characteristic nonlinear mapping functions, The initial three-dimensional feature matrix, Adapt the matrix for the current scene. This is a correction factor for historical data. The number of historical scene samples. Let be the feature vector of the nth historical scene.

[0038] Specifically, the feature mapping expression of the cross-scene clothing adaptation learning model achieves the adaptation transformation from the initial three-dimensional feature matrix to standardized feature vectors through feature nonlinear transformation and historical data correction. This implementation relies on a rich historical scene feature database and efficient feature processing algorithms. In the feature mapping stage, the feature nonlinear mapping function adopts a multilayer perceptron structure, using multiple fully connected layers to compress the dimensionality of the initial three-dimensional feature matrix and reorganize its features, transforming the high-dimensional, sparse feature matrix into a low-dimensional, dense feature vector. During the transformation, key distinguishing features are retained, and redundant information is removed. The current scene adaptation matrix is ​​constructed based on environmental parameters and clothing category distribution information of the real-time sorting scene, representing the feature adaptation requirements of the current scene. Historical data correction coefficients are used to control the influence of historical scene samples on the current feature mapping, avoiding adaptation deviations caused by excessive differences between historical data and the current scene. In implementation, the initial 3D feature matrix is ​​first input into a nonlinear mapping function to obtain basic feature vectors. Then, element-wise multiplication is performed with the current scene adaptation matrix to achieve scene adaptation adjustment of features. Simultaneously, the sum of squared differences between the initial feature matrix and the feature vectors of each historical scene is calculated. After weighting with historical data correction coefficients, this sum is superimposed on the scene-adapted feature vectors to generate standardized feature vectors. This expression addresses the problem of feature distribution differences under different sorting scenarios. By combining historical data with the current scenario, the scene adaptability of feature vectors is improved, enabling the model to maintain stable classification performance in different scenarios such as garment factory outbound, express logistics transfer, and e-commerce return sorting, thus enhancing the system's environmental adaptability and robustness.

[0039] Preferably, the category prediction expression of the intelligent data analysis platform for clothing sorting is: ,in, Predictions for clothing categories, The number of categories in the rule base. Let be the weight coefficient of the k-th category. This is the function for calculating feature similarity. This is the standard feature vector of the k-th category in the rule base.

[0040] Specifically, the category prediction expression of the intelligent data analysis platform for clothing sorting achieves accurate classification of clothing categories through weighted calculation of multi-category feature similarity and maximum value filtering. Its implementation relies on a comprehensive feature classification rule base and an efficient similarity calculation mechanism. During the rule base construction phase, the number of categories in the rule base is determined based on the sorting categories supported by the system, including common clothing material, style, size, and other classification dimensions. Each category corresponds to a set of standard feature vectors, which are generated by collecting a large amount of feature data of similar clothing and performing statistical averaging and outlier removal. Category weight coefficients are set based on factors such as sorting frequency, classification difficulty, and misclassification loss for each category of clothing. Categories with high sorting frequency and high classification difficulty are given higher weights to ensure the classification accuracy of key categories. The feature similarity calculation function uses the cosine similarity algorithm to quantify the degree of matching between the standardized feature vector and the standard feature vector of each category. The closer the calculation result is to 1, the higher the feature fit. In implementation, the standard feature vectors corresponding to all categories are first retrieved from the rule base. The similarity between each standardized feature vector and each standard vector is calculated. Then, the similarity is multiplied by the corresponding category weight coefficient. All products are summed and averaged. Finally, the category with the largest weighted similarity sum is determined using a maximum value filtering function; this is the clothing category prediction result. This expression reduces the randomness of single feature matching through multi-category comparison and weighted decision-making, improving the reliability of category prediction. Simultaneously, the standardized design of the rule base makes the classification logic more interpretable and adjustable, providing a clear and reliable basis for sorting decisions.

[0041] Preferably, the dynamic update expression for the model parameters is: ,in, For the updated model parameters, These are the model parameters before the update. For learning rate, This is the gradient value of the loss function. For the actual category label of the clothing, To predict category labels, This is a correction factor for the operating status of the sorting equipment.

[0042] Specifically, the dynamic update expression for model parameters achieves iterative optimization of the parameters of each model and algorithm through error feedback and state correction, thus constructing a closed-loop linkage mechanism between sorting results and model parameters. During the data acquisition phase required for parameter updates, the model parameters before the update are the effective parameter set of each currently running model and algorithm. The learning rate is set according to the convergence speed and stability of model training, employing an adaptive adjustment strategy. A larger value is set in the early stages of model training to accelerate convergence, and then gradually decreased as training iterations proceed to avoid parameter oscillations. The gradient of the loss function is calculated by comparing the difference between the actual category label and the predicted category label of the clothing. The cross-entropy loss function is used to quantify the classification error, and the gradient calculation is implemented through the backpropagation algorithm, accurately reflecting the direction and magnitude of parameter adjustments. The sorting equipment operating status correction factor is calculated based on the real-time operating parameters of the equipment. When the equipment is in a stable operating state, the correction factor approaches 1, showing no significant inhibition on the parameter update magnitude. When the equipment experiences operational abnormalities, the correction factor is less than 1, reducing the parameter update magnitude and avoiding model performance fluctuations caused by abnormal data. In implementation, the gradient of the loss function is first calculated, and then multiplied by the learning rate to obtain the basic parameter update amount. Next, the basic update amount is adjusted using a device operating status correction factor. Finally, the adjusted update amount is added to the original model parameters to obtain the updated model parameters. This expression establishes a correlation between model performance and actual sorting results and device operating status. Through continuous parameter iteration and optimization, each model and algorithm can adapt to the dynamic changes in the sorting scenario, correct deviations in the classification and scheduling process, gradually improve the system's sorting accuracy and operational stability, and achieve long-term optimization of system performance.

[0043] Preferably, step S3 includes the following sub-steps: S31, extracting the calibration feature dimensions from the standardized feature vector, the calibration feature dimensions including the texture feature dominant dimension, the thickness feature core dimension, and the contour feature calibration dimension, and assigning feature importance weights to each dimension according to a preset weight allocation rule; S32, calling the rule base matching module of the clothing sorting intelligent data analysis platform, comparing the weighted calibration feature dimensions with the feature threshold ranges corresponding to each category in the rule base one by one, and recording the category identifiers and matching degree values ​​of the successfully matched categories; S33, sorting according to the matching degree values, selecting the top three categories with the highest matching degree as candidate categories, and calculating the comprehensive confidence of each candidate category, the comprehensive confidence being calculated by weighting the feature matching degree, the historical sorting accuracy of similar categories, and the current scene adaptability; S34, determining the candidate category with the highest comprehensive confidence as the clothing category prediction result, and simultaneously outputting the confidence index corresponding to the prediction result and the relevant information of the other two candidate categories.

[0044] Specifically, the category prediction process in step S3 improves classification accuracy through four sub-steps. S31 extracts the dominant dimension of texture features, the core dimension of thickness features, and the key dimension of contour features from the standardized feature vector. These three dimensions include over 200 feature items. Each dimension is assigned a feature importance weight of 0.3 to 0.5 according to a preset weighting rule. The dominant dimension of texture features has a weight of 0.5, while the core dimension of thickness features and the key dimension of contour features have weights of 0.3 and 0.2 respectively, ensuring the dominant role of the core features. S32 calls the rule base matching module of the intelligent data analysis platform for clothing sorting. This module has over 500 built-in feature threshold range rules. It compares the weighted key feature dimensions with the corresponding feature threshold ranges for each category in the rule base one by one, at a comparison rate of 1000 times per second. It records the category identifier of a successful match and a matching score from 0 to 100. S33 sorts the categories by matching degree from high to low, selecting the top three categories with the highest matching degree as candidate categories. The overall confidence score is calculated by weighting feature matching degree (weight 0.5), historical sorting accuracy of similar categories (weight 0.3), and current scenario adaptability (weight 0.2), with a value range of 0 to 1. S34 determines the candidate category with the highest overall confidence score as the clothing category prediction result, and outputs the corresponding confidence index of the prediction result, as well as the category identifier, matching degree value, and overall confidence score of the other two candidate categories. The entire process, through hierarchical screening and multi-dimensional weighted evaluation, improves the accuracy of category prediction to over 98%, providing a reliable basis for subsequent sorting scheduling.

[0045] Preferably, step S4 includes the following sub-steps: S41, collecting real-time operating status parameters of the sorting equipment, including the current task queue length, operating power, response latency, and fault warning status of each execution unit, and constructing an equipment status matrix; S42, fusing the clothing category prediction results, confidence index, and equipment status matrix to determine the constraints of the multi-objective optimization scheduling model, including task execution time constraints, equipment load balancing constraints, and sorting accuracy constraints; S43, setting an objective function based on a multi-task sorting scheduling balancing algorithm, including three sub-objectives: maximizing task completion efficiency, minimizing equipment energy consumption, and minimizing sorting error; S44, solving the multi-objective optimization scheduling model using the gradient descent method to generate the optimal matching scheme between each sorting task and execution unit, and the corresponding task execution parameters.

[0046] Specifically, the scheduling model construction process in step S4 achieves multi-objective optimization through four sub-steps. S41 collects real-time operating status parameters of the sorting equipment at a frequency of 10 times per second, including the current task queue length (0 to 20 tasks), operating power (300 to 800 watts), response latency (5 to 50 milliseconds), and fault warning status (levels 0, 1, and 2) of the eight execution units. These parameters are then organized into an 8×4 equipment status matrix to ensure real-time parameter accuracy. S42 fuses the clothing category prediction results, the confidence index in the 0-1 range, and the equipment status matrix to determine the constraints of the multi-objective optimization scheduling model. The task execution time constraint is set to ≤5 seconds for a single task; the equipment load balancing constraint is set to ≤5 variance of the task queue length for each execution unit; and the sorting accuracy constraint is set to three levels based on confidence level: 95%, 97%, and 99%. S43 is based on a multi-task sorting scheduling equilibrium algorithm, setting three sub-objective functions: maximizing task completion efficiency, minimizing equipment energy consumption, and minimizing sorting error. The weights of each sub-objective are allocated as 0.4, 0.3, and 0.3, respectively. S44 solves the multi-objective optimization scheduling model using the gradient descent method, with 100 iterations and a convergence accuracy of 10⁻⁶. -4 This process generates the optimal matching scheme for each sorting task and execution unit, as well as the corresponding execution order, running speed, and other parameters. This process achieves a reasonable allocation of sorting resources, increasing equipment utilization to over 90%.

[0047] Preferably, step S5 includes the following sub-steps: S51, converting the optimal matching scheme into control instructions recognizable by the execution unit, the control instructions including task number, target sorting area coordinates, action execution sequence, and force adjustment parameters; S52, after receiving the control instructions, each execution unit initializes and adjusts its own operating parameters, calibrating the detection benchmarks of the position sensor, force sensor, and speed sensor; S53, according to the timing requirements of the control instructions, the execution unit starts the mechanical transmission structure, and completes the physical transfer of clothing through grasping, transporting, and placing actions, collecting action execution data in real time during the process; S54, after the action is completed, the execution unit feeds back the task completion signal to the multi-task sorting scheduling and balancing algorithm, and simultaneously uploads the action execution data to the clothing sorting intelligent data analysis platform for storage.

[0048] Specifically, the physical sorting process in step S5 ensures sorting stability and accuracy through four sub-steps. S51 converts the optimal matching scheme into control commands recognizable by the execution units. These commands include the task number, target sorting area coordinates with an accuracy of ±1 mm, action execution timing with an error ≤5 milliseconds, and force adjustment parameters of 5 to 50 Newtons, ensuring command accuracy. S52, after receiving the control commands, each execution unit initializes and adjusts its own operating parameters, calibrating the robotic arm's initial position to the origin coordinates (error ≤0.5 mm), setting the initial force of the gripping mechanism to 10 Newtons, and simultaneously calibrating the detection benchmarks of the position sensor (accuracy ±0.1 mm), force sensor (accuracy ±0.5 Newtons), and speed sensor (accuracy ±0.01 m / s), with the sampling frequency set to 100 times per second. The S53 execution unit activates the mechanical transmission structure according to the timing requirements of the control commands, and uses a trapezoidal acceleration and deceleration algorithm to control the movement of the robotic arm. The acceleration during both acceleration and deceleration phases is set to 0.5 m / s², while the speed during the constant speed phase is adjusted from 0.5 to 1.0 m / s based on the task distance. The gripping mechanism adjusts the clamping force according to the force adjustment parameters. During the transfer process, motion execution data is collected in real time, and if a deviation exceeds a threshold, it is immediately corrected. Finally, the physical transfer of clothing is completed through gripping, transfer, and placement actions. After the S54 action is completed, the execution unit sends a task completion signal to the multi-task sorting scheduling and balancing algorithm within 10 milliseconds. Simultaneously, the motion execution data is uploaded to the intelligent data analysis platform for clothing sorting for storage. The entire process ensures that the positional accuracy of clothing sorting is ≤1 mm, and the damage rate is controlled below 0.1%.

[0049] The clothing texture depth perception model takes multi-dimensional visual information of the clothing to be sorted as input, including two-dimensional images of surface texture, thickness distribution data, and contour edge coordinates; the output is a three-dimensional feature matrix fused from these three types of features. The processing flow is as follows: First, visual information is simultaneously acquired using a high-definition industrial camera and a laser depth sensor. After preprocessing, this information is fed into the model. Multi-scale feature extraction is performed on the texture image using 128 convolutional kernels of different sizes. A distribution matrix is ​​constructed by combining this with thickness data, and contour edge parameters are extracted. Finally, the features are fused to generate a three-dimensional feature matrix. The model structure includes a preprocessing module and a feature extraction network. The training data consists of visual samples of clothing covering 30 materials and 20 styles, along with their corresponding feature labels. The training method adjusts the weight coefficients ω1, ω2, and ω3 through error feedback and adjusts the activation function parameter θ. Gradient descent is used to optimize the feature extraction accuracy.

[0050] The cross-scene clothing adaptation learning model takes as input the initial 3D feature matrix output from the clothing texture depth perception model and historical feature data from 100 typical scenes; the output is a standardized feature vector adapted to the current scene. The processing flow is as follows: First, the 3D feature matrix is ​​transformed and batch normalized using three fully connected layers. The five historical reference scenes with the highest similarity are selected. A dynamic adaptation factor is introduced to adjust the feature weights. Finally, the target vector is generated through Min-Max standardization. The model structure consists of a feature mapping layer, a scene adaptation layer, and a feature standardization layer. The training data consists of at least 100,000 clothing feature samples for each scene. The training method calculates the feature similarity between the current and historical scenes and iteratively optimizes the scene adaptation matrix M. snene The correction coefficient λ is used to improve the robustness of feature adaptation across different scenarios.

[0051] The multi-task sorting scheduling balancing algorithm takes into account category prediction results, confidence indices, and real-time operating parameters of the sorting equipment (task queue length, operating power, etc.). The output is the scheduling priority coefficient for each sorting task and the optimal task-execution unit matching scheme. The processing flow is as follows: First, equipment status parameters are collected at a frequency of 10 times / second to construct an equipment status matrix. Constraints (time, load, accuracy) and objective functions (efficiency, energy consumption, error) are set based on the prediction results. The model is solved iteratively 100 times using the gradient descent method. The algorithm structure includes modules for parameter acquisition, constraint setting, objective function construction, and solution. Training data consists of equipment operating data and task execution records under different scenarios. The training method uses a multi-objective optimization algorithm to train the balancing adjustment coefficients α, β, γ, and δ, and dynamically adjusts the objective function weights based on sorting error feedback.

[0052] The texture depth activation function, denoted by σ(θ), primarily enhances the response intensity of effective features in clothing by adjusting the parameter θ to optimize feature extraction. From a technical perspective, this function is a standard feature activation component in deep learning. Its core functionality (enhancing effective features and suppressing redundant information) aligns with the design logic of well-known activation functions such as Sigmoid and ReLU, and the parameter adjustment method conforms to industry-standard design principles. This function falls within the realm of common knowledge and is a standard technique used in feature extraction models to improve feature representation capabilities.

[0053] The formulas in this invention can integrate different scalar and vector parameters for unified calculation by constructing a standardized feature representation and parameter adaptation mechanism. Each formula, through explicit parameter definitions and mathematical transformations, incorporates vector features such as clothing texture, thickness, and contour, along with scalar parameters such as weighting coefficients and correction factors, into the same computational framework. For example, in the clothing texture depth perception model, the three-dimensional feature function in vector form is weighted and summed with scalar weighting coefficients to convert the vector to a specific dimension output value; the cross-scene adaptation model uses a feature nonlinear mapping function to map the high-dimensional vector feature matrix to a low-dimensional vector, and then performs calculations with scalar correction coefficients and vector adaptation matrices, ensuring consistency in the mathematical dimensions of different types of parameters and laying the foundation for unified calculation.

[0054] Through the rational design of mathematical operators, the formula establishes a linked calculation logic for scalar and vector parameters, resolving the compatibility issue of different parameter types. In the multi-task scheduling and balancing algorithm, the scalar confidence and urgency parameters are linearly combined with the vector-based equipment operating efficiency and load parameters through a fractional structure of numerator and denominator, and then nonlinearly corrected using the square root operator. This preserves the physical meaning of various parameters while achieving self-consistency in the calculation logic. The category prediction and parameter update formulas organically integrate scalar weights, vector eigenvectors, and scalar loss values ​​using operators such as summation, maximum value selection, and gradient calculation. This allows the adjustment effect of scalars on vectors and the interaction between vectors to be precisely reflected through mathematical operations, ensuring the rationality and effectiveness of the calculation results.

[0055] Each formula, based on the physical meaning of the model and the actual needs of the sorting scenario, constructs semantic relationships and quantitative mappings between parameters, giving the integration of different types of parameters practical engineering significance. For example, scalar weight coefficients correspond to the importance of different features or targets, while vector parameters correspond to specific physical quantities such as clothing features and equipment status. The formulas, through operations such as weighting, multiplication, and summation, transform semantic-level relationships into mathematical-level computational relationships. Simultaneously, components such as activation functions and correction factors constrain and calibrate the computational results, ensuring that the combined calculation of scalar and vector parameters not only meets mathematical rules but also aligns with the practical needs of feature extraction, scheduling optimization, and model iteration in clothing sorting. This achieves a unity of mathematical logic and engineering application, enabling different types of parameters to work synergistically within the same formula.

[0056] like Figure 2As shown, a simple automated clothing sorting system based on artificial intelligence is described. This system, applied to an AI-based method for automated clothing sorting, includes: a multi-dimensional visual information acquisition unit for acquiring images and depth information of the clothing to be sorted, establishing a data transmission connection with a clothing texture depth perception model, and transmitting the acquired raw information to the model for feature extraction; a cross-scene feature adaptation processing unit that communicates bidirectionally with the clothing texture depth perception model, receiving the three-dimensional feature matrix output by the model, and completing feature mapping through a built-in cross-scene clothing adaptation learning model to generate standardized feature vectors; and an intelligent data analysis and prediction unit connected to the cross-scene feature adaptation processing unit, integrating an intelligent data analysis platform for clothing sorting, and classifying the standardized feature vectors. The system outputs category prediction results and confidence indices; the multi-task scheduling optimization unit communicates with the intelligent data analysis and prediction unit and the sorting execution unit respectively, constructs an optimization model based on the multi-task sorting scheduling balancing algorithm, allocates sorting tasks and outputs control commands; the mechanical sorting execution unit receives control commands from the multi-task scheduling optimization unit, completes the physical sorting operation of clothing through mechanical structure, and collects the actual category information after sorting; the model parameter dynamic update unit establishes data interaction channels with the mechanical sorting execution unit, the clothing texture depth perception model, the cross-scene clothing adaptation learning model and the multi-task scheduling optimization unit, receives error feedback data between the actual category information and the prediction results, adjusts the parameters of each model and algorithm in real time, and continuously optimizes the system sorting performance.

[0057] A simplified automated clothing sorting method and system based on artificial intelligence is proposed. At the feature recognition level, it abandons the traditional single two-dimensional feature extraction mode and comprehensively integrates three-dimensional features of clothing surface texture, thickness distribution, and contour edges using a clothing texture depth perception model. Simultaneously, it mines the value of historical data through a cross-scene clothing adaptation learning model, achieving dynamic feature adaptation under different sorting scenarios. This completely solves the problems of confusion in classifying similar clothing and poor scenario adaptability, significantly improving classification accuracy. At the task scheduling level, a multi-task sorting scheduling balancing algorithm integrates multi-dimensional parameters such as category confidence, real-time equipment status, and execution path to construct a multi-objective optimization mechanism. This achieves scientific matching of sorting tasks and execution units, effectively solving the drawbacks of unreasonable resource allocation and overload or idleness of some units in traditional scheduling, significantly improving overall sorting efficiency.

[0058] At the system level, a complete technical closed loop is formed through the organic linkage of six functional units, further enhancing the practical application of the advantages. The multi-dimensional visual information acquisition unit provides comprehensive data support for feature extraction, the intelligent data analysis and prediction unit achieves accurate feature classification and confidence assessment, the mechanical sorting execution unit ensures the stable execution of physical sorting, and the model parameter dynamic update unit uses sorting error feedback data to continuously optimize the parameters of each model and algorithm, enabling the system performance to continuously improve as the application scenarios expand. This end-to-end design of "perception-adaptation-analysis-scheduling-execution-iteration" overcomes the shortcomings of insufficient feature recognition accuracy through three-dimensional feature fusion and cross-scenario adaptation, solves the problem of uneven resource allocation through multi-objective scheduling optimization, and ensures the long-term stability and adaptability of the system with a closed-loop iteration mechanism, meeting the actual needs of large-scale, multi-scenario clothing sorting.

[0059] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent 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 simple automatic clothing sorting method based on artificial intelligence, characterized in that, Includes the following steps: S1. Multi-dimensional visual information of the clothing to be sorted is acquired through an image acquisition module and transmitted to a clothing texture depth perception model. The model constructs a three-dimensional feature matrix based on clothing surface texture features, thickness distribution features, and contour edge features. S2. A cross-scene clothing adaptation learning model is used to perform feature mapping processing on the three-dimensional feature matrix. Combined with historical feature data of clothing sorting in different scenarios, a standardized feature vector adapted to the current sorting scenario is generated. S3. The standardized feature vector is input into a clothing sorting intelligent data analysis platform. Preliminary classification is performed using the platform's built-in feature classification rule library, and the clothing category prediction result and corresponding confidence index are output. S4, invoke the multi-task sorting scheduling and balancing algorithm to construct a multi-objective optimization scheduling model based on the category prediction results, confidence index, and real-time operating status parameters of the sorting equipment; S5, allocate sorting tasks to corresponding execution units through the multi-objective optimization scheduling model, and the execution units adjust their operating parameters according to the task instructions to complete the physical sorting operation of the clothing; S6, collect the actual category information of the sorted clothing, compare the actual category information with the prediction results, generate error feedback data, and transmit the error feedback data to the clothing texture depth perception model, the cross-scene clothing adaptation learning model, and the multi-task sorting scheduling and balancing algorithm respectively for dynamic updating of model parameters; The feature extraction expression of the clothing texture depth perception model is: ,in, Output values ​​for the three-dimensional feature matrix of clothing. Let the two-dimensional distribution function of the surface texture of the clothing be denoted as . Let the thickness distribution function of clothing be . For the three-dimensional feature function of the clothing outline edge, These are the feature weight coefficients. For texture depth activation function, Adjust the parameters for the activation function; The scheduling allocation expression of the multi-task sorting scheduling and balancing algorithm is as follows: ,in, Let be the scheduling priority coefficient for the i-th sorting task. Predict the confidence level for the category of the i-th task. For the real-time running efficiency of the j-th execution unit, Let be the current task load of the k-th execution unit. Let be the execution path length parameter for the i-th task. For the balance adjustment coefficient, Let be the urgency parameter for the i-th task; S4 includes the following sub-steps: S41, collecting real-time operating status parameters of the sorting equipment, including the current task queue length, operating power, response latency, and fault warning status of each execution unit, and constructing an equipment status matrix; S42, fusing the clothing category prediction results, confidence index, and equipment status matrix to determine the constraints of the multi-objective optimization scheduling model, including task execution time constraints, equipment load balancing constraints, and sorting accuracy constraints; S43, setting an objective function based on a multi-task sorting scheduling balancing algorithm, including three sub-objectives: maximizing task completion efficiency, minimizing equipment energy consumption, and minimizing sorting error; S44, solving the multi-objective optimization scheduling model using the gradient descent method to generate the optimal matching scheme between each sorting task and execution unit, and the corresponding task execution parameters.

2. The simplified automatic clothing sorting method based on artificial intelligence according to claim 1, characterized in that, The feature mapping expression of the cross-scene clothing adaptation learning model is: ,in, To standardize the feature vector, For characteristic nonlinear mapping functions, The initial three-dimensional feature matrix, Adapt the matrix for the current scene. This is a correction factor for historical data. The number of historical scene samples. Let be the feature vector of the nth historical scene.

3. The simplified automatic clothing sorting method based on artificial intelligence according to claim 1, characterized in that, The category prediction expression of the intelligent data analysis platform for clothing sorting is: ,in, Predictions for clothing categories, The number of categories in the rule base. Let be the weight coefficient of the k-th category. This is the function for calculating feature similarity. This is the standard feature vector of the k-th category in the rule base.

4. The simplified automatic clothing sorting method based on artificial intelligence according to claim 1, characterized in that, The expression for the dynamic update of the model parameters is: ,in, For the updated model parameters, These are the model parameters before the update. For learning rate, This is the gradient value of the loss function. For the actual category label of the clothing, To predict category labels, This is a correction factor for the operating status of the sorting equipment.

5. The simplified automatic clothing sorting method based on artificial intelligence according to claim 1, characterized in that, S3 includes the following sub-steps: S31, extracting the calibration feature dimensions from the standardized feature vector, the calibration feature dimensions including the texture feature dominant dimension, the thickness feature core dimension, and the contour feature calibration dimension, and assigning feature importance weights to each dimension according to a preset weight allocation rule; S32, calling the rule base matching module of the clothing sorting intelligent data analysis platform, comparing the weighted calibration feature dimensions with the feature threshold ranges corresponding to each category in the rule base one by one, and recording the category identifiers and matching degree values ​​of successfully matched categories; S33, sorting according to the matching degree values, selecting the top three categories with the highest matching degree as candidate categories, and calculating the comprehensive confidence of each candidate category, the comprehensive confidence is calculated by weighting the feature matching degree, the historical similar sorting accuracy, and the current scene adaptability; S34, determining the candidate category with the highest comprehensive confidence as the clothing category prediction result, and outputting the confidence index corresponding to the prediction result and the relevant information of the other two candidate categories.

6. The simplified automatic clothing sorting method based on artificial intelligence according to claim 1, characterized in that, S5 includes the following sub-steps: S51, converting the optimal matching scheme into control instructions recognizable by the execution unit, the control instructions including task number, target sorting area coordinates, action execution sequence, and force adjustment parameters; S52, after receiving the control instructions, each execution unit initializes and adjusts its own operating parameters, calibrating the detection benchmarks of the position sensor, force sensor, and speed sensor; S53, according to the timing requirements of the control instructions, the execution unit starts the mechanical transmission structure, and completes the physical transfer of clothing through grasping, transporting, and placing actions, collecting action execution data in real time during the process; S54, after the action is completed, the execution unit feeds back the task completion signal to the multi-task sorting scheduling and balancing algorithm, and simultaneously uploads the action execution data to the clothing sorting intelligent data analysis platform for storage.

7. A simple automatic clothing sorting system based on artificial intelligence, characterized in that, This system is applied to a simple automatic clothing sorting method based on artificial intelligence as described in claim 1, comprising: a multi-dimensional visual information acquisition unit, used to acquire images and depth information of the clothing to be sorted, establish a data transmission connection with a clothing texture depth perception model, and transmit the acquired raw information to the model for feature extraction; a cross-scene feature adaptation processing unit, which communicates bidirectionally with the clothing texture depth perception model, receives the three-dimensional feature matrix output by the model, completes feature mapping through a built-in cross-scene clothing adaptation learning model, and generates standardized feature vectors; and an intelligent data analysis and prediction unit, connected to the cross-scene feature adaptation processing unit, integrating a clothing sorting intelligent data analysis platform, classifying the standardized feature vectors, and outputting a category prediction. Results and confidence index; Multi-task scheduling optimization unit, which communicates with intelligent data analysis and prediction unit and sorting execution unit respectively, constructs optimization model based on multi-task sorting scheduling balancing algorithm, allocates sorting tasks and outputs control commands; Mechanical sorting execution unit, which receives control commands from multi-task scheduling optimization unit, completes physical sorting of clothing through mechanical structure, and collects actual category information after sorting; Model parameter dynamic update unit, which establishes data interaction channels with mechanical sorting execution unit, clothing texture depth perception model, cross-scene clothing adaptation learning model and multi-task scheduling optimization unit, receives error feedback data between actual category information and prediction results, adjusts the parameters of each model and algorithm in real time, and continuously optimizes the system sorting performance.

Images