Garment style digital identification method and system

Abstract

The invention relates to the technical field of electrical digital data processing, and discloses a garment style digital identification method and system, and the method comprises the steps: collecting a dynamic visual sequence of a garment in physical driving, extracting a multi-point motion track on the surface of the garment, and carrying out the frequency domain transformation, so as to generate a frequency domain fingerprint vector representing the inherent dynamic response characteristics of the garment, according to the method, the identification basis is converted from the unstable static appearance of the clothes to the inherent and stable dynamic response characteristics of the clothes; the core of the method is that a physical deformation process is converted into an effective signal source containing structural information from an interference item in traditional identification, so that the identification process can get rid of the influence of factors such as an initial random form of clothes and wrinkle shielding.

Description

Technical Field

[0001] This invention relates to a method and system for digitally identifying clothing styles, belonging to the field of electronic digital data processing technology. Background Technology

[0002] Currently, a common technical strategy for the automated identification of flexible objects such as clothing is based on visual analysis of static images. This strategy attempts to capture two-dimensional images of clothing using high-resolution cameras, and then use complex feature extraction algorithms or deep learning models to match and classify key visual elements such as collar, sleeve length, and outline in the image. Its fundamental logic is to reconstruct or infer the internal style structure of clothing from a single frame or a few frames of static spatial form. However, when this static vision-based identification strategy is placed in industrial scenarios such as large-scale, high-throughput automated sorting or recycling, its inherent limitations become apparent due to the pressure of the application environment. In these scenarios, clothing is usually in an imperfect form, such as being randomly grabbed, kneaded, folded, or partially obscured. This results in any static image captured at any single moment being essentially incomplete and highly distorted. This makes it extremely difficult for the algorithm to reliably obtain all the necessary visual features for accurate classification. If physical actions such as shaking or flattening the clothing are added beforehand to improve the recognition rate, the processing efficiency of the entire automated process will be sacrificed, thus causing the system to fall into an inherent contradiction between efficiency and accuracy.

[0003] A major technical challenge of existing technologies lies in treating the inherent flexibility and deformability of clothing as negative interference information that needs to be filtered out, compensated for, or forcibly overcome by algorithms. This design philosophy of resisting deformation not only leads to the increasing complexity and computational dependence of data processing models, but also fundamentally ignores the rich temporal information contained in the dynamic deformation process of an object when it is subjected to external forces, which can reveal its internal physical structure and topological relationships. In other words, existing technologies are trying to solve a misdefined problem; they are committed to analyzing what an object looks like, rather than understanding how an object moves.

[0004] Specifically, existing technologies mainly face the following inherent constraints that are difficult to reconcile: 1. The fundamental instability of the identification basis, that is, using the instantaneously changing static visual form as the basis for identifying the inherent, constant style of flexible objects, resulting in a low signal-to-noise ratio and incomplete information in the input data source itself; 2. The fundamental dilemma of information processing, that is, treating the deformation of the core physical characteristic of flexible objects as negative noise to be counteracted, rather than utilizing it as a signal source containing structural information, which may lead to insufficient information utilization. Therefore, how to extract a digital feature that can stably represent the style structure of flexible objects of any shape by analyzing their dynamic response process without relying on the static visual appearance of clothing, thus achieving efficient and accurate identification of flexible objects of any shape, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a digital recognition method for clothing styles. Its main purpose is to solve the problem of how to extract a digital feature that can stably represent the style structure of clothing by analyzing its dynamic response process without relying on the static visual appearance of the clothing, thereby achieving efficient and accurate recognition of flexible objects of any shape.

[0006] To achieve the above objectives, the present invention provides a method for digitally identifying clothing styles, comprising the following steps:

[0007] Step a: During the process of applying a physical motion to the garment, a dynamic video sequence of the garment is acquired through a visual sensor.

[0008] Step b: Based on the dynamic video sequence, extract the motion trajectory of multiple feature points on the surface of the clothing to generate multi-dimensional time series data representing the motion trajectory;

[0009] Step c: The multidimensional time series data is transformed in the frequency domain to generate a frequency domain fingerprint vector, which represents the inherent vibration mode of the clothing structure under physical induction.

[0010] Step d: Input the frequency domain fingerprint vector into a trained classifier and output the clothing style classification result.

[0011] Preferably, the step of extracting the motion trajectory of multiple feature points on the surface of the garment is implemented using an optical flow estimation algorithm; and the step of performing frequency domain transformation on the multidimensional time series data is to perform a fast Fourier transform to generate a frequency domain fingerprint vector. The physical triggering is formed by using a robotic arm to grasp and move the garment in the automated sorting process.

[0012] Preferably, the trained classifier is a support vector machine classifier.

[0013] Preferably, the trained classifier is the K-nearest neighbor classifier.

[0014] Preferably, before inputting the frequency domain fingerprint vector into the trained classifier, the method further includes: calculating a normalization factor corresponding to the total kinetic energy injected by the physical event based on the motion trajectory of multiple feature points; and normalizing the frequency domain fingerprint vector through the following operation to eliminate the influence of the inconsistency of the physical event on the style classification result: ,in, The normalized frequency domain fingerprint vector, The frequency domain fingerprint vector before normalization. This is the normalization factor.

[0015] Preferably, before inputting the frequency domain fingerprint vector into the trained classifier, the method further includes: controlling a device for applying physical impulse to perform a micro-perturbation action with a preset amplitude, preset duration, and preset acceleration; acquiring the dynamic response of the garment after the micro-perturbation action, and determining a material parameter based on the energy decay rate of the dynamic response, the material parameter being used to quantify the elasticity and damping characteristics of the garment fabric; and selecting a matching classifier from a set of classifiers based on the material parameter to perform style classification, the set of classifiers containing multiple classifiers trained for different material characteristics.

[0016] Preferably, before the step of extracting the motion trajectory of multiple feature points on the surface of the garment, the method further includes: determining whether the number of traceable texture features in the dynamic video sequence is lower than a first threshold; if the number is lower than the first threshold, then activating a one-dimensional ranging sensor to collect sequence data of one-dimensional distance changing over time formed by the undulations of the garment surface during the physical activation process, and performing frequency domain transformation based on the sequence data of one-dimensional distance changing over time to generate a frequency domain fingerprint vector.

[0017] Preferably, after obtaining the style classification result, the method further includes: selecting the frame image with the largest clothing outline from the dynamic video sequence and calculating the aspect ratio of the bounding rectangle of the outline; obtaining a pre-stored aspect ratio value range corresponding to the style classification result; and marking the style classification result as pending review when the calculated aspect ratio of the bounding rectangle is not within the aspect ratio value range.

[0018] Preferably, the step of determining the material parameters based on the energy decay rate of the dynamic response includes: obtaining a decay coefficient by performing an exponential function fitting on the decay process of the overall motion energy over time after the micro-perturbation action, and using the decay coefficient as the material parameter.

[0019] A digital identification system for clothing styles, comprising:

[0020] A data acquisition module is configured to acquire a dynamic video sequence of the garment via a visual sensor during the application of a physical motion to the garment.

[0021] A processing module, connected to the acquisition module, is configured to: extract the motion trajectories of multiple feature points on the surface of the garment based on a dynamic video sequence to generate multidimensional time-series data representing the motion trajectories; and perform frequency domain transformation on the multidimensional time-series data to generate a frequency domain fingerprint vector, which represents the inherent vibration mode of the garment structure under physical induction.

[0022] A style classification module is connected to the processing module. The style classification module is configured to receive frequency domain fingerprint vectors and use a built-in trained classifier to classify the styles of clothing to output the style classification results of clothing.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. This invention shifts the basis of clothing style recognition from analyzing the static spatial form of clothing at a specific moment, which is easily distorted by crumpling and folding, to analyzing the inherent dynamic response characteristics determined by its own structure when subjected to physical forces. By capturing the motion trajectory of multiple points on the surface of the clothing and transforming it to the frequency domain, a stable fingerprint that can characterize its internal structural vibration mode is generated. This approach eliminates the attempt to resist or repair the inevitable deformation of flexible objects in the recognition process. Instead, it utilizes the deformation process itself as a signal source containing structural information. Thus, under the premise of using low-cost visual sensors and simple calculation methods, it helps to reduce the impact of the randomness of the initial shape of clothing, wrinkles, and self-occlusion on the recognition results.

[0025] 2. In processing the dynamic response data of clothing, this invention constructs a data normalization processing mechanism. By reusing the extracted motion trajectories of feature points on the clothing surface, it calculates a normalization factor that can quantify the total kinetic energy injected in each physical activation. This factor is then used to process the original frequency domain fingerprint vector. This approach does not rely on external sensors to monitor the repeatability accuracy of the activation device, but rather extracts key perturbation variables from the intermediate data processed by the system itself. This avoids the inconsistency caused by factors such as fluctuations in the working conditions of the robotic arm during the activation process at the algorithm level, ensuring that the feature vector used for classification only reflects the response characteristics of the clothing to unit energy. This significantly enhances the stability and comparability of the features in response to process fluctuations in a real industrial environment.

[0026] 3. This invention establishes an adaptive classification process that decouples material perception and style recognition. Before determining the style, the system can apply a standardized micro-perturbation action and determine a material parameter that quantifies the elasticity and damping characteristics of the clothing fabric based on the energy decay rate of the subsequent dynamic response. Then, based on this parameter, the system selects the classifier that best matches the current clothing material from a set of classifiers trained for different material characteristics to perform subsequent style classification. This mechanism separates the two major variables of material and structure of flexible objects through the progressive logic of identifying the material first and then judging the style, avoiding the ambiguity of recognition where changes in dynamic response caused by material differences are mistakenly attributed to different styles. Attached Figure Description

[0027] Figure 1This is a flowchart illustrating an embodiment of a digital identification method for clothing styles according to the present invention;

[0028] Figure 2 This is an example diagram of the frequency domain fingerprint vectors of different styles of clothing according to the present invention;

[0029] Figure 3 This is a structural block diagram of a clothing style digital recognition system according to the present invention.

[0030] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit the scope of protection of this application.

[0032] A method for digitally recognizing clothing styles begins by acquiring a dynamic video sequence of the clothing using a visual sensor during a physical movement applied to it. In a specific scenario of this application, this physical movement can be naturally generated by a robotic arm on an automated sorting line performing routine tasks of grasping and moving clothing. Given that in similar high-throughput industrial scenarios, clothing often appears in a non-ideal form, such as random kneading, folding, or partial obscuring, any technical strategy attempting to reconstruct its inherent style from a single static image faces the fundamental limitation of incomplete information. The technical solution of this application is based on recognizing the inherent flexibility of clothing and... The easily deformable characteristics are transformed from interference items in traditional identification into effective signal sources containing structural information. To achieve this, after obtaining the dynamic video sequence, a processing module is configured to use an optical flow estimation algorithm to extract and track the motion trajectory of multiple feature points on the surface of the clothing, thereby generating multidimensional time series data that can characterize the motion trajectory. Subsequently, the processing module performs a fast Fourier transform on the multidimensional time series data to generate a frequency domain fingerprint vector. The frequency domain fingerprint vector objectively characterizes the inherent vibration mode of the clothing structure under physical induction. This fingerprint vector is finally input into a trained classifier to output the style classification result of the clothing.

[0033] In actual working conditions, the intensity of physical activation may be inconsistent due to factors such as fluctuations in the robotic arm's operating conditions. This difference directly affects the dynamic response, thus impacting the stability of the final classification. To avoid this impact, this application introduces an intrinsic self-calibration mechanism. Before inputting the frequency domain fingerprint vector into the trained classifier, the system calculates a normalization factor corresponding to the total kinetic energy injected by the physical activation based on the motion trajectories of multiple extracted feature points. The original frequency domain fingerprint vector is then processed through the following operations. Normalization is performed: ,in, The normalized frequency domain fingerprint vector utilizes intermediate data processed by the system itself to quantify and eliminate perturbation variables, ensuring that the final feature vector used for classification only reflects the clothing's response characteristics to unit energy input. This enhances the stability of the features in response to process fluctuations in a real industrial environment. Furthermore, the fabric material of the clothing is also a key variable affecting its dynamic response. To avoid misattributing changes in dynamic response due to material differences to different styles, this application also discloses an adaptive classification process. Before the main initiation of style recognition, the system can control a device for applying physical initiation to perform a micro-perturbation action with a preset amplitude, preset duration, and preset acceleration. The system then collects the dynamic response of the clothing after the micro-perturbation action. By fitting an exponential function to the decay process of the overall kinetic energy of the dynamic response over time, a decay coefficient is obtained. This decay coefficient is used as a material parameter to quantify the elasticity and damping characteristics of the clothing fabric. Subsequently, the system can select a matching classifier from a set of classifiers based on the material parameter to perform style classification. The classifier set contains multiple classifiers trained for different material characteristics, such as support vector machine classifiers or nearest neighbor classifiers.

[0034] To ensure the universality and robustness of the solution, this application also considers the boundary case where the surface of the clothing to be identified lacks stable texture features. Before the step of extracting the motion trajectory of multiple feature points on the clothing surface, the system also includes a step of determining whether the number of trackable texture features in the dynamic video sequence is lower than a first threshold. If the number is lower than the first threshold, the system activates a one-dimensional ranging sensor to collect the sequence data of one-dimensional distance changing over time formed by the undulations of the clothing surface during the physical motion process, and performs frequency domain transformation based on the sequence data of one-dimensional distance changing over time to generate a frequency domain fingerprint vector. This primary and backup data acquisition method... The path settings ensure that the core functions of the system can still be reliably implemented even when the preferred path is blocked. Finally, after obtaining the style classification results, in order to further verify the rationality of the classification results, the system can also select the frame image with the largest clothing outline from the dynamic video sequence and calculate the aspect ratio of the bounding rectangle of the outline. At the same time, a pre-stored aspect ratio value range corresponding to the style classification results is obtained. When the calculated aspect ratio of the bounding rectangle is not within the aspect ratio value range, the style classification result is marked as pending review. This step provides a verification link based on macroscopic shape constraints for the automated process.

[0035] Example 1: In a continuously operating large-scale automated textile sorting facility, the system faces a specific working condition: a pure cotton long-sleeved shirt and a polyester-cotton blend long-sleeved shirt appear on two adjacent conveyor belts. Due to long-term use and transportation, the two garments exhibit highly wrinkled and partially flipped random shapes. Visually, their outlines, colors, and textures are extremely similar under the sensor. However, according to the sorting procedure, they must be precisely sent to different recycling channels due to their different materials. When the first pure cotton shirt enters the robotic arm's working range, the system first drives the robotic arm to perform a micro-perturbation motion with a preset amplitude and acceleration. The visual sensor simultaneously collects the energy attenuation process of the garment after this micro-perturbation and determines a material parameter characterizing its high-damping properties through exponential function fitting. Immediately afterwards, the robotic arm applies a complete physical motion to the garment. After extracting the multi-point motion trajectory on its surface, the processing module generates the original frequency domain fingerprint vector. On the other hand, the normalization factor corresponding to the total kinetic energy injected in this tactic is calculated based on the same set of motion trajectory data. And then through calculation A normalized frequency domain fingerprint vector was obtained. Finally, based on the material parameters obtained in the previous step, the system calls a support vector machine classifier trained on high-damping materials such as pure cotton from the classifier set, and normalizes the frequency domain fingerprint vector. Input the classifier to get the classification result.

[0036] Subsequently, when the second polyester-cotton blended top entered the working area, the system completely repeated the above procedure. The difference was that the system determined a material parameter characterizing its low-damping properties through micro-perturbation response, and selected another classifier trained for low-damping materials such as synthetic fibers from the classifier set based on this parameter to complete the final style determination. In this process, even if the two physical actions injected different total kinetic energy due to the difference in the working conditions of the robotic arm, the existence of the normalization factor ensured that the normalized frequency domain fingerprint vectors on which the two recognitions were based had a unified scale, and the material parameter pre-... First, it is determined that a decision model with specific discrimination ability is assigned to the two classification tasks. Thus, the two originally coupled technical problems of material recognition and style recognition are decomposed into two recognition steps that have a sequence and are independent of each other. Finally, pure cotton long-sleeved tops and polyester-cotton blend long-sleeved tops are sent to their respective recycling channels. The throughput efficiency of the entire sorting process is not affected. The basis of this recognition process has changed from analyzing the uncertain spatial shape of the object to analyzing its inherent physical characteristics determined by both material and structure under energy dissipation and forced vibration.

[0037] Example 2: To objectively verify the recognition stability of the technical solution of this application when faced with physical inconsistency disturbances, especially the effectiveness of the normalization process, the following experimental platform was built. This platform uses a six-axis industrial robotic arm to apply physical motion, and a high-speed vision sensor to acquire dynamic video sequences. The test subjects were four different styles of standard clothing: T-shirts, long-sleeved tops, trousers, and dresses, with one hundred pieces of each style, totaling four hundred pieces. To isolate the interference of fabric material variables on this experiment, all clothing of the same style used the same batch of pure cotton fabric. The core of this experiment is to quantitatively evaluate the impact of changes in physical motion energy on the classification results. Regarding the ability of the normalization mechanism proposed in this application to suppress this influence, the driving energy is set as a key experimental parameter to simulate the fluctuations in equipment operating conditions that may occur in industrial settings. The fundamental technical consideration here is to ensure that the energy change is sufficient to significantly affect the unprotected feature vectors, while also ensuring that the range of change is within a reasonable range for typical industrial applications. Based on this, three driving energy levels were set in the experiment: low energy, medium energy, and high energy, which were achieved by setting the maximum movement speed of the robotic arm end effector to 0.5 m / s, 1.0 m / s, and 1.5 m / s, respectively; among them, 1.0 m / s was set as the baseline driving speed for training the classifier.

[0038] The experimental procedure is as follows: First, a K-nearest neighbor classifier was trained using data collected from all 400 garments under medium-energy conditions. Then, each garment was tested for identification under low, medium, and high energy levels of physical conditions, generating an original frequency domain fingerprint vector for each test. and a normalized frequency domain fingerprint vector The two types of vectors were input into the previously trained classifier, and the accuracy of style classification was recorded. The experiment observed that when using unnormalized frequency domain fingerprint vectors… When performing classification, its recognition accuracy peaks under medium-energy propulsion, but significantly decreases under both low-energy and high-energy propulsion; in contrast, using a normalized frequency domain fingerprint vector... When classifying, its recognition accuracy remained at a similarly high level across all three energy levels, as shown in Table 1.

[0039] Table 1: Comparison of classification accuracy under different motivation energy levels.

[0040]

[0041] The data trends shown in Table 1 stem from the following underlying mechanism: the total kinetic energy injected by physical initiation. Changes in this will directly cause the vibration amplitude of the clothing to change in the same direction, thereby affecting the frequency domain fingerprint vector. The amplitudes of each frequency component are scaled proportionally. When the driving energy deviates from the baseline energy used during classifier training, the position of the feature vector in multidimensional space will shift as a whole, thus exceeding the learned classification boundary and leading to misclassification. However, through computation... The normalization process essentially utilizes the total kinetic energy calculated in real time from the trajectory. By removing the energy-dependent scaling factor from the feature vectors, the feature vectors generated under different driving energies can be mapped back to a compact and stable region in the feature space, thus ensuring the stability of the classifier's decision. The experimental results show that by introducing a normalization process based on the total kinetic energy of the physical driving force, the performance fluctuations caused by the inconsistency of the driving source can be effectively suppressed. This mechanism provides performance consistency support for the deployment of this recognition method in real industrial environments with inherent process fluctuations.

[0042] Example 3: This example combines Figures 1 to 3 A description of a digital identification method and system for clothing styles, such as... Figure 1 As shown, in the input acquisition layer, the process begins with physical actuation, such as a robotic arm grasping moving clothing. During this process, a visual sensor continuously records to acquire dynamic video sequences. Then, a decision node is reached where the system determines whether the number of texture features is below a threshold. If not, the system follows the main technical path and uses an optical flow estimation algorithm to extract the motion trajectory of feature points. If yes, a backup plan is activated, where a one-dimensional ranging sensor acquires surface undulation data. Additionally, this layer includes a separate micro-perturbation material recognition step to perform standardized micro-perturbations. In the feature processing layer, the trajectory data extracted on the main path is converted into multi-dimensional time-series data to characterize the motion trajectory, and then processed by Fast Fourier Transform. The system uses leaf transform to generate frequency domain fingerprint vectors. Simultaneously, one-dimensional data collected via backup paths can also be directly transformed in the frequency domain to generate fingerprint vectors. This layer also includes three key processing steps: energy normalization to avoid inconsistencies; material parameter calculation based on micro-perturbation response to calculate energy attenuation coefficients; and adaptive classifier selection based on material parameter matching. Finally, in the decision output layer, the processed frequency domain fingerprint vector is fed into a selected classifier, such as a support vector machine (SVM) K-nearest neighbor, to obtain a style classification output. This output also needs to undergo aspect ratio verification to check the rationality of the outline shape. Based on all the above information, the system finally outputs the clothing style recognition result.

[0043] like Figure 2As shown, although all curves are relatively similar in the low-frequency region and generally show a decreasing trend with increasing frequency, in the mid-to-high frequency region, different styles of clothing exhibit amplitude response curves with different shapes and obvious distinctions. These differentiated curve shapes constitute a digital fingerprint that can stably characterize the internal physical structure of each clothing style, thus providing a reliable basis for subsequent classifier decisions.

[0044] like Figure 3 As shown, the system's input sources include a visual sensor and a backup one-dimensional ranging sensor, which provide dynamic video sequences and one-dimensional distance sequences to the acquisition module, respectively. The acquisition module transmits the processed raw dynamic data to the processing module, which interacts with a knowledge base storing prior data such as material / aspect ratio, and sends micro-perturbation control commands to the physical actuator robotic arm. The core output of the processing module is a frequency domain fingerprint vector, which is sent to the style classification module. This module then uses the material information obtained from the knowledge base to call a matching classifier stored in the classifier set for decision-making. Finally, the style classification module outputs two types of commands: one is a style classification result containing sorting instructions, which is sent to the physical actuator robotic arm to execute the corresponding physical action; the other is a verification command, used to import garments with logically contradictory recognition results into the verification channel.

[0045] Example 4: Before the method and system of this application are put into actual operation, an offline system calibration and knowledge base construction process needs to be performed to determine a series of key operating parameters and judgment criteria, so as to deal with specific challenges that may be encountered in actual operation, including the lack of stable texture features on the surface of the clothing to be identified, or the logical contradiction between the final classification result and the macroscopic shape of the clothing; in order to solve the problem of optical flow estimation algorithm failure caused by the lack of texture on the surface of clothing, the system needs to pre-calibrate a decision threshold, i.e., the first threshold, to activate the backup sensor channel; the calibration procedure is as follows: prepare a calibration clothing set containing two types of samples, the first type is clothing with rich texture on the surface, and the second type is clothing with solid color or no obvious texture on the surface; the system Each garment in the calibration set undergoes a standard physical motion, during which dynamic video sequences are acquired. The number of feature points that can be stably tracked by the optical flow estimation algorithm is calculated for each video sequence. After processing all garments, the system obtains two sets of statistical distributions of feature point counts, corresponding to garments with rich and lacking textures, respectively. A first threshold is set as the boundary value that distinguishes these two statistical distributions. To ensure that the acquisition of material parameters in the adaptive classification process has physical meaning and reproducibility, the step of determining material parameters based on the energy decay rate of the dynamic response is refined into the following operations: After the system applies a standardized micro-perturbation to the garment, the processing module starts from the acquired dynamic video sequence... Extract all from the frame image Instantaneous velocity of each tracking feature point and according to The total kinetic energy of the system corresponding to this frame is calculated, where the mass of each feature point is... It can be assumed to be uniform and normalized to 1; by performing this calculation on a series of consecutive frames after the micro-perturbation action ends, the system can obtain a time series of data characterizing the decay of total kinetic energy over time. Subsequently, the system uses the least squares method to fit the time series data to the exponential decay model. The attenuation coefficient obtained by the solution This refers to the material parameters used to quantify the damping properties of the garment fabric.

[0046] To ensure consistency and comparability in material parameter measurements, when the system is deployed on a specific production line, the device used to perform micro-perturbation actions has a series of actuation parameters, such as preset amplitude, preset duration, and preset acceleration, which are ultimately determined through a set of on-site calibration procedures. This procedure begins with a set of calibration garments covering a typical fabric to be processed on the production line. Then, each calibration garment is systematically perturbed within a preset parameter space, and its energy attenuation coefficient, i.e., the material parameter λ, is measured. This leads to the introduction of a material discrimination index. This index is defined as the ratio of the between-group variance to the total within-group variance of the λ value distribution corresponding to different fabric groups. Its physical meaning lies in quantifying the degree of separation of the dynamic response signals of each material. Ultimately, by scanning and optimizing across various dimensions of the driving parameters, the material discrimination index will be improved. The set of activation parameters that reaches the maximum value is set as the fixed standard for executing all subsequent micro-perturbation actions under this production line environment. In addition, in order to establish a database of aspect ratio numerical ranges for cross-validation of classification results, the system will simultaneously perform the following steps when training models such as the nearest neighbor classifier offline. For each garment sample of the same style in the training set, after extracting its frequency domain fingerprint vector, the system backtracks its dynamic video sequence, selects the frame image with the largest outline, and calculates the aspect ratio of its bounding rectangle. After processing all training samples under this style, the system will obtain a statistical distribution of the aspect ratio of this style, and calculate the mean and standard deviation of the distribution. Finally, the pre-stored aspect ratio numerical range corresponding to the classification result of this style is defined as the interval of plus or minus two standard deviations of this mean. By executing the above offline calibration and construction process, before entering the real-time recognition task, all the key decision thresholds and judgment knowledge bases in the system have obtained a deterministic source based on statistical and physical models, thereby excluding the operator's experience judgment from the system configuration process.

[0047] Example 5: Before the system is formally deployed and performs online recognition tasks, a structured benchmark dataset covering various materials and styles, along with a set of classifiers, needs to be constructed. This procedure begins by collecting a benchmark library containing various clothing samples, covering all styles to be identified, and each style includes garments made of different fabrics. For each clothing sample in the library, the system first records its accurate style classification result as the benchmark ground truth by manually labeling or reading tags. Subsequently, using the above procedure, through standardized micro-perturbation actions and energy decay fitting, its unique material parameters are accurately measured and recorded. .

[0048] After completing the data collection for all samples, the system processes all recorded material parameters. Cluster analysis is performed on the values ​​to divide their continuous numerical distribution into several discrete material intervals with well-defined boundaries. Each interval corresponds to a macroscopic material dynamic response characteristic. Based on this partitioning result, the system categorizes the entire benchmark dataset according to material parameters. The system automatically divides the data into multiple subsets and trains a dedicated style classifier for each subset. Specifically, for a given material range, the system trains an independent classifier using only the frequency domain fingerprint vectors and style baseline ground truth values ​​corresponding to clothing samples falling within that range. The trained classifier is then stored in a classifier set and indexed and associated with that material range. After this process, the system possesses a decision core consisting of multiple expert classifiers optimized for the dynamic response characteristics of specific materials. In the real-time identification phase, the system can call the matching classifier to perform style classification based on the measured material parameters.

[0049] Example 6: Before deploying the clothing style digital recognition system of this application on a new automated sorting production line, in order to match the system performance with the specific mechanical characteristics of the production line and the characteristics of the materials to be sorted, a standardized on-site pre-deployment calibration and parameter optimization procedure needs to be executed. This procedure first determines the optimal parameters of the micro-perturbation motion used to extract material parameters. The system engineer uses a set of calibration garments containing a variety of typical fabrics and measures the material parameters corresponding to each fabric under a series of preset perturbation amplitude and acceleration combinations. And select those that allow different fabrics to work together. The amplitude and acceleration that have the highest distinguishability in the statistical distribution of values ​​are used as the fixed parameters for standard micro-disturbance actions in this specific production line environment.

[0050] Based on this, the optimal hyperparameters are determined for the nearest neighbor classifier. For each data subset divided according to material range, the system uses cross-validation within a preset integer range to optimize the hyperparameter values ​​before training the final dedicated classifier. The value that achieves the highest average cross-validation accuracy on that data subset is set as the final value of the dedicated classifier for that material range. At the same time, to solidify the anomaly handling logic, the style classification result is marked as pending review. At the system controller level, it is configured to trigger a specific sorting instruction. This instruction guides the robotic arm to place the corresponding garment into a dedicated review channel or container. Finally, a one-dimensional ranging sensor is installed directly above the physical trajectory, with its measuring beam perpendicularly aligned with the geometric center area of ​​the garment during movement to capture the maximum surface undulation dynamic data. After completing this set of procedures, all the core control parameters, model hyperparameters, and anomaly handling logic of the system are given deterministic values ​​and rules bound to the specific deployment environment, and the system then enters online operation.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for digitally recognizing clothing styles, characterized in that, Includes the following steps: Step a: During the process of applying a physical motion to the garment, a dynamic video sequence of the garment is acquired through a visual sensor. Step b: Based on the dynamic video sequence, extract the motion trajectory of multiple feature points on the surface of the clothing to generate multi-dimensional time series data representing the motion trajectory; Step c: The multidimensional time series data is transformed in the frequency domain to generate a frequency domain fingerprint vector, which represents the inherent vibration mode of the clothing structure under physical induction. Step d: Input the frequency domain fingerprint vector into a trained classifier and output the clothing style classification result.

2. The method for digitally recognizing clothing styles according to claim 1, characterized in that, The step of extracting the motion trajectory of multiple feature points on the surface of the garment is implemented using an optical flow estimation algorithm; and the step of performing frequency domain transformation on multidimensional time series data is to perform a fast Fourier transform to generate a frequency domain fingerprint vector. The physical drive is formed by using a robotic arm to grasp and move the garment in the automated sorting process.

3. The method for digitally recognizing clothing styles according to claim 1, characterized in that, The trained classifier is a support vector machine classifier.

4. The method for digitally recognizing clothing styles according to claim 1, characterized in that, The trained classifier is the K-nearest neighbor classifier.

5. The method for digitally recognizing clothing styles according to claim 1, characterized in that, Before inputting the frequency domain fingerprint vector into the trained classifier, the process includes: calculating a normalization factor corresponding to the total kinetic energy injected by the physical event based on the motion trajectory of multiple feature points; and normalizing the frequency domain fingerprint vector through the following operation to eliminate the influence of the inconsistency of the physical event on the style classification result: ,in, The normalized frequency domain fingerprint vector, The frequency domain fingerprint vector before normalization. This is the normalization factor.

6. The method for digitally recognizing clothing styles according to claim 1, characterized in that, Before inputting the frequency domain fingerprint vector into the trained classifier, the method further includes: controlling a device for applying physical impulses to perform a micro-perturbation action with a preset amplitude, preset duration, and preset acceleration; acquiring the dynamic response of the garment after the micro-perturbation action and determining a material parameter based on the energy decay rate of the dynamic response, the material parameter being used to quantify the elasticity and damping characteristics of the garment fabric; and selecting a matching classifier from a set of classifiers based on the material parameter to perform style classification, the set of classifiers containing multiple classifiers trained for different material characteristics.

7. The method for digitally recognizing clothing styles according to claim 1, characterized in that, Before the step of extracting the motion trajectory of multiple feature points on the surface of the garment, the method further includes: determining whether the number of traceable texture features in the dynamic video sequence is lower than a first threshold; if the number is lower than the first threshold, activating a one-dimensional ranging sensor, collecting sequence data of one-dimensional distance changing over time formed by the undulations of the garment surface during the physical motion process, and performing frequency domain transformation based on the sequence data of one-dimensional distance changing over time to generate a frequency domain fingerprint vector.

8. The method for digitally recognizing clothing styles according to claim 1, characterized in that, After obtaining the style classification result, the process also includes: selecting the frame image with the largest clothing outline from the dynamic video sequence and calculating the aspect ratio of the bounding rectangle of the outline; obtaining a pre-stored aspect ratio value range corresponding to the style classification result; and marking the style classification result as pending review when the calculated aspect ratio of the bounding rectangle is not within the aspect ratio value range.

9. The method for digitally recognizing clothing styles according to claim 7, characterized in that, The steps for determining material parameters based on the energy decay rate of the dynamic response include: obtaining a decay coefficient by performing an exponential function fitting on the decay process of the overall motion energy over time after the micro-perturbation action, and using the decay coefficient as the material parameter.

10. A digital recognition system for clothing styles, characterized in that, include: A data acquisition module is configured to acquire a dynamic video sequence of the garment via a visual sensor during the application of a physical motion to the garment. A processing module, connected to the acquisition module, is configured to: extract the motion trajectory of multiple feature points on the surface of clothing based on a dynamic video sequence, in order to generate multidimensional time series data representing the motion trajectory; A frequency domain transformation is performed on the multidimensional time series data to generate a frequency domain fingerprint vector, which represents the inherent vibration mode of the clothing structure under physical induction. A style classification module is connected to the processing module. The style classification module is configured to receive frequency domain fingerprint vectors and use a built-in trained classifier to classify the styles of clothing to output the style classification results of clothing.

Images