Cloth identification robot

Abstract

The invention discloses a cloth identification robot, and relates to the technical field of cloth identification. The device comprises a shell, a touch display screen, operation keys, a protective cover, a positioning assembly, an identification assembly and an AI processor, wherein the AI processor is integrated with an algorithm processing module. The positioning assembly realizes stable and accurate positioning of the cloth through a positioning frame, an elastic pressing sheet and an edge alignment sensor; after an identification assembly collects an image, the image is subjected to preprocessing, feature extraction and screening, a deep learning classification model fused with an attention mechanism is input, an identification result is output in combination with optimization mechanisms such as cross validation and secondary identification, and the model can also be replaced by a lightweight transfer learning model to adapt to different scenes. An identification result is displayed through the touch display screen and uploaded through the communication module, and the model can be adaptively optimized. According to the scheme, a mechanical structure and an intelligent algorithm are fused, the problems that existing equipment is inaccurate in positioning and weak in interference resistance are solved, the recognition precision and adaptability are improved, and the digital requirement of the textile industry is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloth recognition, and particularly relates to a cloth recognition robot. BACKGROUND

[0002] Under the background of accelerating digital transformation of the textile industry, cloth recognition as the core link of the whole chain such as procurement, production and processing, terminal sales, its efficiency and accuracy directly affect the overall efficiency of the industry. With the market's increasing demand for the diversity and individualization of cloth materials, traditional manual identification methods have been difficult to meet the needs of large-scale and accurate production and operation, and automated and intelligent cloth recognition equipment has become a key support for the industry upgrade and is widely used in textile factories, clothing enterprises and fabric markets.

[0003] However, the technical optimization of existing cloth recognition robots is mostly limited to mechanical structure adjustment, and the core performance of flexible positioning and image acquisition has reached the upper limit of structure optimization, making it difficult to achieve breakthrough improvement. Such equipment generally has low positioning fault tolerance and weak recognition anti-interference ability: cloth is prone to shift or wrinkle after being placed due to material differences, resulting in deviation of image acquisition position; at the same time, there is a lack of targeted algorithm optimization, and when facing cloth with similar colors and blurred textures, it is easy to be disturbed by light and impurities, and the recognition accuracy and adaptability are insufficient, which cannot meet the accurate recognition needs of diversified cloth, so an innovative scheme integrating mechanical structure and intelligent algorithm is urgently needed to break through the existing technical bottleneck. SUMMARY

[0004] The present application relates to the technical field of cloth recognition, and particularly relates to a cloth recognition robot.

[0005] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: a cloth recognition robot, comprising a shell, a touch display screen and operation buttons mounted on the top surface of the shell, a clamping groove is horizontally formed at the lower end of the front and rear of the shell, a protective cover is arranged at the lower end of the shell, a clamping strip is mounted in the protective cover, and the protective cover is clamped at the lower end of the shell through the clamping strip and the clamping groove; two detection openings are formed on one side of the bottom surface of the shell, and a recognition assembly is mounted in the shell; an installation groove is formed on the bottom surface of the shell, a flexible positioning frame is mounted in the installation groove, and elastic pressing pieces are movably hinged on the front and rear ends of the positioning frame; a positioning groove is formed on the diagonal of the bottom surface of the shell, the positioning groove is arranged on the outer side of the installation groove, and an edge alignment sensor is mounted in the positioning groove; the recognition assembly comprises an image acquisition module, an AI processor and a high-speed storage module, the AI processor integrates an algorithm processing module, the algorithm processing module is configured to execute a cloth recognition algorithm, and the cloth recognition algorithm comprises: The recognition assembly further comprises an AI processor in communication connection with the image acquisition module, the AI processor is configured to execute a cloth recognition algorithm, and the execution of the cloth recognition algorithm comprises the following steps: Step S1: Acquire image data of the fabric sample through the image acquisition module; Step S2: Preprocess the image data, including denoising, illumination equalization, and image enhancement; Step S3: Extract the feature vector of the preprocessed image, the features including texture features, color histogram features and local binary pattern features; Step S4: Input the feature vector into the pre-trained deep learning classification model and output the fabric material category recognition result; Step S5: Display the recognition result through the touch screen and / or upload it to an external system through the communication module; The specific implementation of step S4 is as follows: The deep learning classification model adopts a convolutional neural network model with an attention mechanism. After the feature vector is input, it is unified in dimension and integrated in information by the feature fusion layer. Then, it is mined in depth by multi-level convolutional blocks. The attention mechanism module assigns differentiated weights to different types of features to strengthen key recognition information. The model output layer calculates the confidence of various fabric materials through probability distribution, selects the category with the highest confidence as the preliminary recognition result, and verifies the reliability of the preliminary result through the feature consistency verification mechanism. After the verification is passed, the final fabric material category recognition result is output.

[0006] Preferably, the image acquisition module includes two industrial cameras installed on one side of the top surface inside the housing. The lower end of the industrial cameras is mounted on a bracket, and a supplementary light is installed directly below the bracket. The supplementary light is placed inside the detection port. An L-shaped mounting base is fixed to the other side of the housing. The AI ​​processor and the high-speed storage module are respectively installed on the top surface and the inner side of the mounting base.

[0007] Preferably, a communication module is mounted on the bottom surface of the mounting base, and a rechargeable lithium battery is located below the communication module. The lithium battery is installed on the other side of the bottom surface inside the housing; both the communication module and the lithium battery are electrically connected to the AI ​​processor.

[0008] Preferably, the elastic pressure plate is made of elastic metal and has a torque adjustment structure at its hinge to adaptively adjust the clamping force according to the fabric thickness.

[0009] As a preferred embodiment, another implementation of step S4 is to replace the deep learning classification model with a lightweight neural network model based on transfer learning, using a combination structure of a pre-trained model's feature extraction layer and a custom classification layer, fine-tuning the classification layer parameters with a small number of fabric samples to achieve rapid deployment and low-power operation, and outputting fabric material category recognition results.

[0010] Preferably, the edge alignment sensor communicates with the AI ​​processor in real time. When it detects that the fabric has not fully entered the calibration area, the AI ​​processor controls the touch screen to display a prompt message and pauses the image acquisition process.

[0011] Preferably, the algorithm processing module further includes a feature filtering algorithm, which is executed after step S3 and before step S4. Redundant features are eliminated through a feature importance evaluation mechanism, and core feature vectors that contribute highly to the identification of fabric material are retained and then input into a deep learning classification model for identification.

[0012] As a preferred option, a cross-validation step is added to step S4, where the feature vector is simultaneously input into two deep learning classification models with different structures. When the recognition results output by the two models are consistent, the result is output directly; when the results are inconsistent, a weighted voting mechanism is activated to combine the confidence of the two models and output the final result.

[0013] Preferably, the algorithm processing module also includes a model adaptive optimization algorithm, which periodically calculates the accuracy of the recognition results. When the recognition error exceeds the preset range multiple times in a row, the model parameter fine-tuning process is automatically triggered to optimize the model recognition performance based on the recently collected effective sample data.

[0014] Preferably, in step S4, when the highest confidence level output by the deep learning classification model is lower than a preset threshold, the algorithm processing module automatically starts the secondary recognition process, re-extracts the supplementary feature vector of the image, merges it with the feature vector extracted initially, and then inputs it back into the model to output the secondary recognition result.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This solution achieves stable fabric bearing and precise calibration through the coordinated operation of a flexible positioning frame, elastic pressure plates, and edge alignment sensors, combined with a torque adjustment structure and real-time communication mechanism. The elastic pressure plates can adaptively adjust the clamping force according to the fabric thickness, while the edge alignment sensors can quickly detect whether the fabric has entered the effective acquisition area, providing timely feedback and pausing acquisition. This dual approach, combining mechanical structure and electronic control linkage, ensures the accuracy of image acquisition position, completely avoiding recognition deviations caused by fabric displacement or wrinkles, and significantly improving operational error tolerance.

[0016] 2. Image preprocessing algorithms remove interference and enhance features, while feature selection algorithms retain core and effective information. A deep learning model integrating attention mechanisms achieves deep feature mining and key information enhancement. Combined with optimization mechanisms such as cross-validation and secondary recognition, it effectively handles complex scenarios such as similar colors, blurred textures, and light and shadow interference. It also provides a lightweight model alternative, balancing accurate recognition with rapid deployment and low-power operation, significantly improving the device's adaptability to diverse fabrics.

[0017] 3. This solution achieves intelligent management and control of the entire fabric recognition process through deep integration of the AI ​​processor and various modules. Recognition results can be displayed in real-time on the touchscreen and simultaneously uploaded to external systems via the communication module, supporting data synchronization and remote interaction, thus meeting the digital transformation needs of the textile industry. The built-in adaptive optimization algorithm automatically fine-tunes parameters based on recognition accuracy, continuously adapting to the recognition needs of different batches and types of fabrics, reducing manual intervention and maintenance costs. A rechargeable lithium battery ensures the device's portability, and the quick-release design of the protective cover facilitates operation and component protection, comprehensively improving the device's practicality and ease of maintenance.

[0018] In summary, this solution, through the deep integration of mechanical structure and intelligent algorithms, not only solves the core pain points of existing fabric recognition robots, such as inaccurate positioning and weak anti-interference capabilities, but also achieves a comprehensive improvement in recognition accuracy, adaptability, and intelligence level. It not only breaks through the performance limits of simple mechanical structure optimization, meeting the textile industry's needs for large-scale and precise recognition, but also offers flexible deployment modes and convenient operation and maintenance design. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall three-dimensional structure proposed in this invention; Figure 2 This is a schematic diagram of the overall bottom-view three-dimensional structure proposed in this invention; Figure 3 This is a schematic diagram of the internal structure of the shell proposed in this invention; Figure 4 The present invention proposes Figure 2 Enlarged schematic diagram of the structure at part A in the middle; Figure 5 This is a block diagram showing the connection of the device hardware modules proposed in this invention; Figure 6 This is a flowchart illustrating the core algorithm principle proposed in this invention. Figure 7 This is a block diagram of the model optimization mechanism proposed in this invention.

[0020] The components in the diagram are numbered as follows: 1. Housing; 2. Protective cover; 3. Touch screen; 4. Operation buttons; 5. Positioning frame; 6. Elastic pressure plate; 7. Edge alignment sensor; 8. Industrial camera; 9. Bracket; 10. Fill light; 11. Mounting base; 12. AI processor; 13. High-speed storage module; 14. Communication module; 15. Lithium battery. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] See Figures 1 to 5 This invention discloses a fabric recognition robot, comprising a housing 1, a touch screen display 3 mounted on the top surface of the housing 1, and operation buttons 4. The housing 1 has horizontally opening slots at its lower front and rear ends, and a protective cover 2 at its lower end. A retaining strip is installed inside the protective cover 2, and the protective cover 2 is secured to the lower end of the housing 1 via the retaining strip and the slots. Two detection ports are opened on one side of the bottom surface of the housing 1, and a recognition component is installed inside the housing 1. The housing 1 provides structural support and protection for internal components, while the protective cover 2 prevents the recognition component from being affected by dust or impacts. The bottom surface of the housing 1 has a mounting groove, in which a flexible positioning frame 5 is installed. Elastic pressure plates 6 are laterally hinged to the front and rear ends of the positioning frame 5. The flexible material of the positioning frame 5 facilitates stable support and positioning of the fabric. The elastic pressure plates 6 allow for elastic compression of the fabric edges and corners, preventing fabric displacement during data collection. Positioning grooves are also provided at the diagonal corners of the bottom surface of the housing 1, located outside the mounting groove. Edge alignment sensors 7 are installed within these grooves. The edge alignment sensors 7 accurately detect whether the fabric is placed in the calibration area, ensuring accurate data collection. The edge alignment sensor 7 is an Omron E3Z-T61. The recognition component includes an image acquisition module, an AI processor 12, and a high-speed storage module 13. The AI ​​processor 12 integrates an algorithm processing module, which is configured to execute a fabric recognition algorithm. The fabric recognition algorithm includes: The recognition component also includes an AI processor 12 that is communicatively connected to the image acquisition module. The AI ​​processor 12 is configured to execute a fabric recognition algorithm, which includes the following steps: Step S1: Acquire image data of the fabric sample through the image acquisition module; Step S2: Preprocess the image data, including denoising, illumination equalization, and image enhancement; Step S3: Extract the feature vector of the preprocessed image. The features include texture features, color histogram features, and local binary pattern features, denoted as follows: ,in n, m, and k are the dimensions of the three types of features, respectively; Step S4: Input the feature vector into the pre-trained deep learning classification model and output the fabric material category recognition result; Step S5: Display the recognition results via the touch screen 3 and / or upload them to an external system via the communication module 14; The specific implementation of step S4 is as follows: The deep learning classification model employs a convolutional neural network model with an attention fusion mechanism. First, a feature fusion layer unifies the dimensions and integrates information from the three types of feature vectors. The fusion formula is as follows: In the formula, The fused feature vector These are vectors representing texture features, color histogram features, and local binary pattern features, respectively, after dimensionality unification. These are the feature fusion weights corresponding to texture features, color histogram features, and local binary pattern features, respectively. fused feature vector Deep feature mining is performed by inputting multi-level convolutional blocks. The output formula of the convolutional layer is: In the formula, Y is the output feature vector of the convolutional layer; W is the weight matrix of the convolutional kernel; and b is the bias vector. It is the ReLU activation function; The attention mechanism module assigns differentiated weights to the deep feature vector Y. The weight calculation formula is as follows: , In the formula Let be the attention weight for the i-th feature dimension. The importance score is assigned to feature dimension i, where p is the dimension of the deep feature vector Y. Let Y be the i-th element. The feature vector after attention weighting; The model output layer calculates the confidence scores for various fabric materials using the Softmax function, with the following formula: In the formula Let be the confidence level for the fabric to belong to category c. The output value of the fully connected layer of the model for category c, where q is the total number of cloth material categories; The category with the highest confidence level is selected as the initial identification result, and then verified through a feature consistency check mechanism. The verification formula is as follows: In the formula, Let w be the feature consistency coefficient, and w be the mean attention weight; when If the result is less than the preset threshold, the verification is considered successful, and the preliminary identification result is output as the final fabric material category identification result.

[0023] It should be noted that the snap-fit ​​between the shell and the protective cover provides reliable protection for the internal components; the synergistic effect of the flexible positioning frame, the elastic pressure plate 6, and the edge alignment sensor 7 achieves stable fabric bearing and accurate calibration, avoiding displacement during data collection; the fabric recognition algorithm integrated into the AI ​​processor 12, through image preprocessing, multi-dimensional feature extraction, deep learning model operation with attention mechanism, and feature consistency verification, deeply mines the essential characteristics of the fabric, strengthens key recognition information, effectively improves the accuracy and reliability of fabric material recognition, and adapts to diverse fabric recognition needs.

[0024] Specifically, the image acquisition module includes two industrial cameras 8 installed on one side of the top surface inside the housing 1. A bracket 9 is mounted on the lower end of each industrial camera 8, and a supplementary light 10 is installed directly below the bracket 9, positioned inside the detection port. An L-shaped mounting base 11 is fixed to the other side of the housing 1. An AI processor 12 and a high-speed storage module 13 are respectively installed on the top and inner sides of the mounting base 11. The industrial cameras 8 are designed to facilitate the acquisition of clear images of the fabric in conjunction with the detection port, providing raw data for recognition. The model of the industrial camera 8 is KnightCamV50C. The supplementary light 10 provides uniform illumination for image acquisition, avoiding shadows and reflections and improving image clarity. The AI ​​processor 12 facilitates the analysis of the fabric image for recognition. The model of the AI ​​processor 12 is i5-12600HE. The high-speed storage module 13 facilitates the storage of acquired image data, recognition parameters, and related information, ensuring rapid data retrieval and storage. The model of the high-speed storage module 13 is PhisonPS5013-E13T.

[0025] Specifically, a communication module 14 is installed on the bottom surface of the mounting base 11, and a rechargeable lithium battery 15 is located below the communication module 14. The lithium battery 15 is installed on the other side of the bottom surface inside the housing 1. Both the communication module 14 and the lithium battery 15 are electrically connected to the AI ​​processor 12. The communication module 14 facilitates data transmission between the device and the outside world, supporting data synchronization and remote interaction. The lithium battery 15 provides rechargeable power, ensuring the device's battery life when disconnected from mains power, and adapting to mobile usage scenarios.

[0026] Specifically, the elastic pressure plate 6 is made of elastic metal, and its hinge has a torque adjustment structure for adaptively adjusting the clamping force according to the fabric thickness. This structure adopts existing adjustable torque hinge technology, and through the cooperation of the built-in elastic element and the adjustment component, it adaptively adjusts the clamping torque according to the fabric thickness to adapt to the clamping force. Since this torque adjustment structure is a conventional prior art in this field, its specific structure and working principle are common knowledge, so there is no need to describe it in detail. It is sufficient to clarify its function and application method to meet the requirements of sufficient patent disclosure.

[0027] Specifically, another implementation of step S4 is as follows: the deep learning classification model is replaced with a lightweight neural network model based on transfer learning, using a combination structure of the feature extraction layer of the pre-trained model and a custom classification layer. The formula for the confidence score output by the custom classification layer is: In the formula: The confidence level that the fabric belongs to category c; This is the output vector of the pre-trained feature extraction layer; , These are the weight vector and bias term for category c corresponding to the custom classification layer, respectively. These represent the weight vector and bias term for the fabric material of the d-th class corresponding to the custom classification layer; fine-tuned using a small number of fabric samples. , By using classification layer parameters, we can achieve rapid deployment and low-power operation, and output the results of fabric material category recognition.

[0028] It should be noted that this implementation method uses a lightweight neural network model based on transfer learning. By leveraging the combination of the feature extraction layer of the pre-trained model and the custom classification layer, it can be quickly deployed without the need for full training with a large number of fabric samples. It only requires fine-tuning the classification layer parameters with a small number of samples, which reduces the model training cost and cycle, and achieves low power consumption operation. While ensuring the accuracy of fabric material recognition, it is suitable for use scenarios that require deployment efficiency and energy consumption, thus expanding the application scope of the device.

[0029] Specifically, the edge alignment sensor 7 communicates with the AI ​​processor 12 in real time. When it detects that the fabric has not fully entered the calibration area, the AI ​​processor 12 controls the touch screen 3 to display a prompt message and pauses the image acquisition process.

[0030] Specifically, the algorithm processing module also includes a feature selection algorithm, which is executed after step S3 and before step S4. It evaluates feature importance using mutual information entropy, and the calculation formula is as follows: In the formula: Let be the mutual information entropy between the i-th feature and the fabric category C; For the information entropy of category C, Known features Conditional information entropy of category C, discarding Redundant features below a preset threshold are retained, and the core feature vectors are then input into a deep learning classification model for recognition.

[0031] It should be noted that this feature selection algorithm is executed after feature extraction and before model input. It accurately evaluates the correlation importance between each feature and the fabric category through mutual information entropy, effectively eliminating redundant features with low contribution to recognition, and retaining only the core effective feature vectors. This reduces the computational load of the deep learning classification model, improves recognition efficiency, and avoids redundant information interfering with the model's capture of the essential features of the fabric, further ensuring the accuracy and stability of fabric material recognition.

[0032] Specifically, a cross-validation step is added in step S4, where the feature vector Xfusion is simultaneously input into two deep learning classification models with different structures. , Output the confidence scores respectively. When a category exists Make and At that time, output directly As the identification result; when the categories corresponding to the highest confidence scores of the two models are different, a weighted voting mechanism is activated, and the comprehensive confidence score formula is: In the formula: The overall confidence level for category c. The weighting coefficients are (0 < λ < 1); [Select] The largest category is used as the final identification result.

[0033] It should be noted that the cross-validation process effectively avoids the identification bias and limitations of a single model by simultaneously inputting the fused feature vector into two deep learning classification models with different structures, and by using the consistency of the results to directly output or by using a weighted voting mechanism to integrate the confidence level. This enhances the reliability and credibility of the identification results. At the same time, by flexibly balancing the identification advantages of different models through weight coefficients, the accuracy and stability of fabric material identification in complex scenarios are further improved.

[0034] Specifically, the algorithm processing module also includes a model adaptive optimization algorithm and periodically calculates the accuracy of the recognition results. The accuracy calculation formula is: In the formula: The number of correctly identified samples. This represents the total number of samples within the statistical period. When t consecutive statistical analyses are performed When all values ​​are below the preset accuracy threshold, the model parameter fine-tuning process is automatically triggered. Based on the recently collected valid sample data, the convolution kernel weight matrix W and attention weight calculation parameters are updated to optimize the model's recognition performance.

[0035] It should be noted that the adaptive optimization algorithm periodically calculates the recognition accuracy. When the accuracy falls below the preset threshold multiple times in a row, the parameter fine-tuning process is automatically triggered. The core parameters of the model are updated based on recent valid sample data, enabling the model to continuously adapt to the recognition needs of different batches and types of fabrics, dynamically compensate for recognition deviations, avoid model performance degradation, ensure the accuracy and stability of fabric material recognition in the long term, and reduce the cost of manual intervention optimization.

[0036] Specifically, in step S4, when the deep learning classification model outputs the highest confidence level... When the image quality falls below a preset threshold, the algorithm processing module automatically initiates a secondary recognition process to re-extract the image's shape feature vector. The primary and supplementary features are fused using the following formula: In the formula This is a secondary fusion feature vector. For shape features via dimensions The unified vector; δ is the fusion coefficient (0 < δ < 1); Input the data into the model again, and the model will output the secondary recognition result.

[0037] It should be noted that the secondary recognition process is automatically initiated when the highest confidence level of the model output does not reach the threshold. By supplementing the extraction of image shape features and weighting the fusion with the initial fusion features, the feature dimensions are enriched and the essential feature representation of the fabric is strengthened. This effectively makes up for the limitations of single feature extraction, avoids recognition deviations caused by insufficient feature information, significantly improves the recognition accuracy of fabrics with indistinct features or complex materials, and ensures the reliability and comprehensiveness of the recognition results.

[0038] The working principle of the fabric recognition robot proposed in this invention is as follows: Equipment startup and preparation: The rechargeable lithium battery 15 provides stable power to all electrical components of the entire device, including the industrial camera 8, fill light 10, AI processor 12, high-speed storage module 13, communication module 14, edge alignment sensor 7, touch display screen 3, and operation buttons 4. After the operator starts the device using the operation button 4 on the top surface of the housing 1, they can pinch the protective cover 2 and use the locking mechanism between the internal locking strip of the protective cover 2 and the horizontally opened slots at the front and rear ends of the housing 1 to remove the protective cover 2 from the bottom of the housing 1. This prevents the protective cover 2 from obstructing the two detection ports opened on one side of the bottom surface of the housing 1, thus clearing the way for subsequent image acquisition.

[0039] Fabric positioning and calibration testing: The operator places the fabric to be identified onto the flexible positioning frame 5 within the mounting groove on the bottom surface of the housing 1. The flexible positioning frame 5 provides a stable support base and positioning reference for the fabric, ensuring that the fabric is initially placed in a neat position. After placement, the elastic pressure plates 6, which are laterally hinged at the front and rear ends of the positioning frame 5, automatically adhere to and press the edges and corners of the fabric due to their elasticity. The elastic pressure plates 6 are made of elastic metal, and their torque adjustment structure at the hinges adaptively adjusts the pressing force according to the thickness of the fabric. This prevents the fabric from shifting or deviating during subsequent image acquisition and also prevents wrinkles caused by excessive pressing force, ensuring the integrity of image acquisition. Simultaneously, the edge alignment sensor 7 installed in the positioning groove (located outside the mounting groove) on the bottom of the housing 1 starts working synchronously, maintaining real-time communication with the AI ​​processor 12 to accurately detect whether the fabric has completely entered the preset effective acquisition calibration area. If the fabric is detected to be not fully within the calibration area, the AI ​​processor 12 will immediately control the touch screen 3 to display a clear prompt message, and at the same time suspend the subsequent image acquisition process until the operator adjusts the fabric position to meet the requirements, ensuring the accuracy of the image acquisition position and improving the operational error tolerance.

[0040] Image acquisition and preprocessing: Once the edge alignment sensor 7 detects that the fabric has fully entered the calibration area and sends a pass signal, the recognition component inside the housing 1 begins image acquisition. The image acquisition module within the recognition component consists of two industrial cameras 8, a bracket 9, and a supplementary light 10. The two industrial cameras 8 are mounted on one side of the top inner surface of the housing 1 via the bracket 9, and the supplementary light 10, directly below the bracket 9, is positioned within the detection port. During acquisition, the supplementary light 10 is simultaneously activated, providing uniform light supplementation for the image acquisition by the industrial cameras 8, effectively avoiding the influence of environmental factors such as shadows and reflections on image quality, and improving the clarity of the fabric image. The industrial cameras 8 are aligned with the detection port, continuously acquiring image data of the fabric to be identified, and transmitting the acquired raw image data in real time to the AI ​​processor 12 mounted on the top surface of an L-shaped mounting base 11 fixed to the other side of the housing 1. After receiving the raw image data, the AI ​​processor 12 first executes the image preprocessing process, which performs noise reduction, illumination equalization, and image enhancement on the image through preset algorithms: noise reduction can remove interference impurities such as dust and fabric lint in the image; illumination equalization can correct the image brightness imbalance caused by uneven lighting or differences in ambient light; and image enhancement further enhances the key features such as the texture and color of the fabric, laying a high-quality data foundation for subsequent feature extraction and recognition work.

[0041] Feature extraction and filtering: After image preprocessing, the algorithm processing module integrated in the AI ​​processor 12 initiates the feature extraction process, extracting three types of core feature vectors from the clear preprocessed image: texture features, color histogram features, and local binary pattern features. Texture features reflect the differences in the fabric's surface texture structure, color histogram features accurately capture the color distribution patterns of the fabric, and local binary pattern features effectively characterize the local texture details of the fabric surface. These three types of features comprehensively depict the material properties of the fabric from different dimensions. To avoid redundant features affecting recognition efficiency and accuracy, after feature extraction and before inputting into the deep learning classification model, the algorithm processing module also initiates a feature filtering algorithm. This algorithm uses a feature importance evaluation mechanism to filter the extracted feature vectors, eliminating redundant features that contribute little to fabric material recognition and retaining only the core feature vectors, ensuring the efficiency and accuracy of the subsequent recognition process.

[0042] Deep learning model identification and optimization: (1) Core identification process: The core feature vectors, after being filtered, are input into a pre-trained deep learning classification model in the AI ​​processor 12. This model employs a convolutional neural network model with an attention mechanism. First, the feature vectors enter a feature fusion layer for dimensional unification and information integration, transforming the three different dimensional core feature vectors into a unified dimensional fused feature, achieving complementarity and enhancement of feature information. Then, the fused feature vectors are input into multi-level convolutional blocks, where deep feature mining is performed through multi-layer convolution operations to further extract the essential features of the fabric material. During this process, the attention mechanism module assigns differentiated weights to the deep features, emphasizing and strengthening the feature information that plays a crucial role in material identification while weakening the influence of secondary features. Finally, the model output layer calculates the confidence level of each fabric material using probability distributions, selecting the category with the highest confidence level as the preliminary identification result. Simultaneously, a feature consistency verification mechanism is activated to verify the reliability of the preliminary identification result. After successful verification, the final fabric material category identification result is output.

[0043] (2) Multiple optimization mechanisms: Cross-validation optimization: To further improve the credibility of the recognition results, a cross-validation step is added to the process. The feature vector is simultaneously input into two deep learning classification models with different structures. If the recognition results output by the two models are consistent, the result is output directly. If the results are inconsistent, a weighted voting mechanism is activated to calculate the overall confidence score by combining the confidence scores of the two models. The category with the highest overall confidence score is selected as the final recognition result.

[0044] Secondary recognition optimization: If the highest confidence score output by the deep learning classification model is lower than the preset confidence threshold, it means that the currently extracted features are insufficient to accurately determine the fabric material. The algorithm processing module will automatically start the secondary recognition process, re-extract the supplementary feature vectors of the image (such as shape features), fuse the supplementary feature vectors with the core feature vectors extracted in the first extraction, and then input them into the model again for recognition, outputting the secondary recognition result to ensure accurate recognition of fabrics with inconspicuous features.

[0045] Model Adaptive Optimization: The algorithm processing module also has a built-in model adaptive optimization algorithm, which will periodically calculate the accuracy of the recognition results. When the accuracy of the calculated results is lower than the preset accuracy threshold for multiple consecutive times, the model parameter fine-tuning process will be automatically triggered. Based on the recently collected effective sample data, the model's convolution kernel weight matrix, attention weight calculation parameters, etc. will be updated to continuously optimize the model's recognition performance and adapt to the recognition needs of different batches and different types of fabrics.

[0046] Lightweight Model Alternative: For use cases requiring rapid deployment and low power consumption, the steps also provide another implementation method, replacing the deep learning classification model with a lightweight neural network model based on transfer learning. This model adopts a combination structure of a pre-trained model's feature extraction layer and a custom classification layer. It does not require a large number of samples for full training. Only a small number of cloth samples are used to fine-tune the classification layer parameters, which can achieve rapid deployment and low power consumption, while ensuring that the recognition accuracy meets basic usage requirements.

[0047] Results output, storage, and device cleanup: After the AI ​​processor 12 outputs the final fabric material category recognition result, it is displayed in real time on the touch screen 3 on the top surface of the housing 1 for easy viewing by the operator. Simultaneously, the recognition result is uploaded to an external system via the communication module 14 installed on the bottom of the mounting base 11, supporting data synchronization and remote interaction to meet the digital management needs of the textile industry in procurement, production, and sales. At the same time, the collected raw image data, preprocessed image data, extracted feature vectors, recognition parameters, and final recognition results are synchronously transmitted to the high-speed storage module 13 installed on the inner side of the mounting base 11 for storage, ensuring rapid data retrieval, stable storage, and subsequent traceability. After the fabric recognition is completed, the operator turns off the device using the operation button 4, disconnecting power to all electrical components. Then, the protective cover 2 is re-attached to the lower end of the housing 1 using its internal locking strip and the slot on the housing 1. The protective cover 2 provides dust and impact protection for the detection port at the lower end of the housing 1 and the internal recognition components and electrical control module, extending the equipment's lifespan. This completes the entire fabric recognition operation process.

[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fabric recognition robot, comprising a housing (1), a touch screen (3) mounted on the top surface of the housing (1), and operation buttons (4), characterized in that: The housing (1) has horizontal slots at the lower front and rear ends, and a protective cover (2) is provided at the lower end of the housing (1). A retaining strip is installed inside the protective cover (2), and the protective cover (2) is secured to the lower end of the housing (1) by the retaining strip and the slot. Two detection ports are provided on one side of the bottom surface of the housing (1), and an identification component is installed inside the housing (1). An installation groove is provided on the bottom surface of the housing (1), and a flexible positioning frame (5) is installed in the installation groove. An elastic pressure plate (6) is horizontally hinged at the front and rear ends of the positioning frame (5). A positioning groove is provided at each corner of the bottom surface of the housing (1), and the positioning groove is located outside the installation groove. An edge alignment sensor (7) is installed in the positioning groove. The identification component includes an image acquisition module, an AI processor (12), and a high-speed storage module (13). The AI ​​processor (12) integrates an algorithm processing module, which is configured to execute a fabric identification algorithm. The fabric identification algorithm includes: The recognition component further includes an AI processor (12) communicatively connected to the image acquisition module. The AI ​​processor (12) is configured to execute a fabric recognition algorithm, which includes the following steps: Step S1: Acquire image data of the fabric sample through the image acquisition module; Step S2: Preprocess the image data, including denoising, illumination equalization, and image enhancement; Step S3: Extract the feature vector of the preprocessed image, the features including texture features, color histogram features and local binary pattern features; Step S4: Input the feature vector into the pre-trained deep learning classification model and output the fabric material category recognition result; Step S5: Display the recognition result via the touch screen (3) and / or upload it to an external system via the communication module (14); The specific implementation of step S4 is as follows: The deep learning classification model adopts a convolutional neural network model with an attention mechanism. After the feature vector is input, it is unified in dimension and integrated in information by the feature fusion layer. Then, it is mined in depth by multi-level convolutional blocks. The attention mechanism module assigns differentiated weights to different types of features to strengthen key recognition information. The model output layer calculates the confidence of various fabric materials through probability distribution, selects the category with the highest confidence as the preliminary recognition result, and verifies the reliability of the preliminary result through the feature consistency verification mechanism. After the verification is passed, the final fabric material category recognition result is output.

2. The fabric recognition robot according to claim 1, characterized in that: The image acquisition module includes two industrial cameras (8) installed on one side of the top surface inside the housing (1). A bracket (9) is installed at the lower end of the industrial camera (8), and a fill light (10) is installed directly below the bracket (9). The fill light (10) is placed inside the detection port. An L-shaped mounting base (11) is fixed to the other side inside the housing (1). The AI ​​processor (12) and the high-speed storage module (13) are respectively installed on the top surface and the inner side of the mounting base (11).

3. The fabric recognition robot according to claim 1, characterized in that: The mounting base (11) has a communication module (14) installed on its bottom surface. A rechargeable lithium battery (15) is located below the communication module (14). The lithium battery (15) is installed on the other side of the bottom surface inside the housing (1). The communication module (14) and the lithium battery (15) are both electrically connected to the AI ​​processor (12).

4. The fabric recognition robot according to claim 1, characterized in that: The elastic pressure plate (6) is made of elastic metal and has a torque adjustment structure at its hinge to adaptively adjust the pressing force according to the thickness of the fabric.

5. A fabric recognition robot according to claim 1, characterized in that: Another implementation of step S4 is to replace the deep learning classification model with a lightweight neural network model based on transfer learning, using a combination structure of a pre-trained model's feature extraction layer and a custom classification layer, fine-tuning the classification layer parameters with a small number of fabric samples to achieve rapid deployment and low-power operation, and outputting fabric material category recognition results.

6. The fabric recognition robot according to claim 1, characterized in that: The edge alignment sensor (7) communicates with the AI ​​processor (12) in real time. When it detects that the fabric has not fully entered the calibration area, the AI ​​processor (12) controls the touch screen (3) to display a prompt message and pauses the image acquisition process.

7. A fabric recognition robot according to claim 1, characterized in that: The algorithm processing module also includes a feature filtering algorithm, which is executed after step S3 and before step S4. Redundant features are eliminated through a feature importance evaluation mechanism, and core feature vectors that contribute highly to the identification of fabric material are retained. These vectors are then input into a deep learning classification model for identification.

8. A fabric recognition robot according to claim 1, characterized in that: In step S4, a cross-validation step is added. The feature vector is simultaneously input into two deep learning classification models with different structures. When the recognition results output by the two models are consistent, the result is output directly. When the results are inconsistent, a weighted voting mechanism is activated to combine the confidence of the two models and output the final result.

9. A fabric recognition robot according to claim 1, characterized in that: The algorithm processing module also includes a model adaptive optimization algorithm, which periodically calculates the accuracy of the recognition results. When the recognition error exceeds the preset range multiple times in a row, it automatically triggers the model parameter fine-tuning process to optimize the model recognition performance based on recently collected effective sample data.

10. A fabric recognition robot according to claim 1, characterized in that: In step S4, when the highest confidence level output by the deep learning classification model is lower than the preset threshold, the algorithm processing module automatically starts the secondary recognition process, re-extracts the supplementary feature vector of the image, merges it with the feature vector extracted initially, and then inputs it back into the model to output the secondary recognition result.

Images