Clothing manufacturing industry process compliance detection method and device and production management method and system

Abstract

The invention discloses a garment manufacturing industry process compliance detection method and device, and a production management method and system. The detection method comprises the following steps: acquiring a frame image in a clothing manufacturing process; inputting the frame images into a target detection model based on deep learning to obtain target detection information of each frame image; and according to the target detection information, judging whether the garment manufacturing process is compliant or not. The generation management method comprises the following steps: firstly, performing compliance detection on processes of the clothing manufacturing industry, then performing statistical query on a detection result, and performing man-hour configuration according to a statistical query result; and finally, according to a man-hour configuration result, carrying out maintenance prompting on production equipment of the clothing manufacturing industry so as to realize production management of the clothing manufacturing industry. According to the detection method, the process compliance detection of the garment manufacturing industry can be efficiently and accurately completed, so that the overall production efficiency and the product quality can be improved, and the production management is optimized.

Description

Technical Field

[0001] The present invention relates to a method and device for detecting compliance of a garment manufacturing process, and a production management method and system. Background Art

[0002] In the garment manufacturing process, operators face some significant challenges when conducting quality inspections on cut pieces, semi-finished fabrics, and finished garments. First, the lack of an effective supervision mechanism makes it difficult to ensure the standardization of the quality inspection process; second, the accuracy of the quality inspection results is also difficult to be effectively guaranteed. In the existing technology, these quality inspection operations rely on manual training, and the final quality inspection results can often only be evaluated through after-sales feedback. This reliance on manual labor makes the quality inspection process more subjective and uncertain.

[0003] Therefore, how to efficiently and accurately implement and standardize the quality inspection process of the clothing manufacturing industry has become an important issue that clothing manufacturers need to solve urgently in quality control, especially for the quality inspection process in process compliance testing. Summary of the invention

[0004] The technical problem to be solved by the present invention is to address the above-mentioned deficiencies in the prior art and propose a method, device, production management method and system for detecting process compliance in the garment manufacturing industry. The detection method can efficiently and accurately complete process compliance detection in the garment manufacturing industry during garment production, thereby improving overall production efficiency and product quality.

[0005] In a first aspect, the present invention provides a method for detecting compliance of a garment manufacturing process based on deep learning, the method comprising the following steps:

[0006] Step S1: Acquire frame images in the garment manufacturing process;

[0007] Step S2: inputting the frame image into a target detection model based on deep learning to obtain target detection information for each frame image;

[0008] Step S3: judging whether the garment manufacturing process is compliant based on the target detection information.

[0009] Furthermore, the step S1 specifically includes the following steps:

[0010] Monitor the garment manufacturing process through surveillance cameras;

[0011] Get the video stream of the surveillance camera;

[0012] Frames are extracted from the video stream according to time to obtain frame images in the clothing manufacturing process.

[0013] Furthermore, the step S3 specifically includes the following steps:

[0014] According to the process information in the target detection information, detect whether the process steps and process materials are compliant to obtain a first detection result; and

[0015] According to the clothing information in the target detection information, detect whether the clothing quality and the clothing process are compliant to obtain a second detection result; and

[0016] According to the worker's hand in the target detection information, the compliance of the operation is judged in combination with the hand key point regression model to obtain a third detection result;

[0017] Whether the garment manufacturing process is compliant is determined based on the first test result, the second test result and the third test result.

[0018] Furthermore, the hand key point regression model is obtained by using data enhancement method and wing_loss as regression loss function;

[0019] The data enhancement method includes random color perturbation, random whitening, random rotation, and random mirroring;

[0020] The calculation formula of the regression loss function is as follows:

[0021]

[0022] C=ω-ωln(1+ω / ∈) (1)

[0024] In formula (1), y represents the prediction result, Represents the real coordinate label;

[0025] C is a constant for adjusting the loss value;

[0026] ω represents the dividing point between the linear and nonlinear regions; ∈ represents the curvature of the control curve.

[0027] Furthermore, the target detection model based on deep learning is a YoLov7 target model;

[0028] Before the step S2, the method further includes a step S0: constructing a YoLov7 target model;

[0029] The step S0 specifically includes the following steps:

[0030] Obtain training data; and build the YoLov7 initial model;

[0031] Set a minimum target for total loss value;

[0032] The YoLov7 initial model is trained according to the training data and the total loss value minimum target to obtain the YoLov7 target model.

[0033] Furthermore, the obtaining of training data specifically includes the following steps:

[0034] Collecting video or image data from historical clothing manufacturing; the video or image data includes process video or image data, clothing material video or image data, and worker hand video or image data;

[0035] Labeling the process video or image data, the clothing video or image data, and the worker hand video or image data respectively;

[0036] Performing data enhancement on the labeled process video or image data, the clothing video or image data, and the worker hand video or image data to obtain a video or image data set;

[0037] The video or image data set is divided into a training set and a test set, thereby obtaining training data consisting of the training set and the test set.

[0038] Furthermore, the total loss value is calculated by a total loss function;

[0039] The total loss function is calculated by the following formula (2):

[0040]

[0041] In formula (2),

[0042] N is the number of detection layers;

[0043] B i is the number of targets whose labels are assigned to the prior box;

[0044] S i ×S i is the number of grids into which the current scale of the i-th target is divided;

[0045] L box is the bounding box regression loss; L obj is the target object loss; L cls is the classification loss;

[0046] L CloUj is the CloU loss of the jth target;

[0047] L objj is the target object loss of the jth target;

[0048] L clsjis the classification loss of the jth target;

[0049] λ 1 is the weight of the bounding box regression loss;

[0050] λ 2 is the weight of the target object loss;

[0051] λ 3 is the weight of classification loss;

[0052] CloU loss L CloU It is calculated by the following formula (3):

[0053]

[0054] In formula (3), b, b gt are the prediction box and label box respectively, w gt 、h gt , w, h are the width and height of the label box and the prediction box respectively, ρ represents the distance between the center points of the two boxes, v is the difference between the aspect ratio of the prediction box and the aspect ratio of the label box, α is the weight coefficient, and IoU represents the intersection over union ratio of the prediction box and the label box;

[0055] Objective loss function L obj It is calculated by the following formula (4):

[0056]

[0057] In formula (4), i represents the sample number, n represents the total number of samples, σ represents the sigmoid function, and y i represents the true label of sample i, x i Represents the output of the model.

[0058] In a second aspect, the present invention provides a production management method for a garment manufacturing industry, the method comprising the following steps:

[0059] According to the clothing manufacturing industry process compliance detection method based on deep learning described in the first aspect, compliance detection is performed on the processes of the clothing manufacturing industry to obtain detection results;

[0060] Performing statistical query on the detection results to obtain statistical query results;

[0061] According to the statistical query results, the working hours are configured to obtain the working hours configuration results;

[0062] According to the working time allocation result, maintenance and servicing reminders are given to the production equipment of the clothing manufacturing industry to realize the production management of the clothing manufacturing industry.

[0063] In a third aspect, the present invention provides a device for detecting compliance of garment manufacturing process based on deep learning, the device comprising:

[0064] An acquisition unit, used for acquiring frame images in the garment manufacturing process;

[0065] An input unit, connected to the acquisition unit, for inputting the frame image into a target detection model based on deep learning to obtain target detection information for each frame image;

[0066] The determination unit is connected to the input unit and is used to determine whether the garment manufacturing process is compliant based on the target detection information.

[0067] In a fourth aspect, the present invention provides a production management system for a garment manufacturing industry, the system comprising:

[0068] A detection device, wherein the detection device is the clothing manufacturing process compliance detection device based on deep learning as described in the third aspect, and the detection device is used to perform compliance detection on the process of the clothing manufacturing industry to obtain a detection result;

[0069] A statistical query device, connected to the detection device, for performing a statistical query on the detection result to obtain a statistical query result;

[0070] A working time configuration device, connected to the statistical query device, is used to perform working time configuration according to the statistical query result to obtain a working time configuration result;

[0071] The equipment management device is connected to the working time configuration device and is used to provide maintenance and servicing prompts for the production equipment of the clothing manufacturing industry according to the working time configuration results, so as to realize the production management of the clothing manufacturing industry.

[0072] The present invention can improve overall production efficiency and product quality and optimize production management by completing process compliance detection in the garment manufacturing industry more efficiently and accurately. The specific beneficial effects are as follows:

[0073] 1. High accuracy

[0074] The deep learning model of the present invention can be trained through a large amount of data and automatically extract features, thereby significantly improving the recognition accuracy of key parts of clothing and reducing human errors.

[0075] 2. Real-time

[0076] The present invention adopts modern deep learning algorithms, which can process large amounts of image data in a short period of time, realize real-time monitoring and detection, and promptly discover non-compliance in the production process.

[0077] 3. High degree of automation

[0078] The present invention reduces the reliance on manual quality inspection and can automatically perform compliance inspection, thereby reducing labor costs and improving production efficiency.

[0079] 4. Strong adaptability

[0080] The present invention is based on a deep learning model and has excellent adaptability. It can effectively cope with complex scenarios and diverse clothing styles and meet the requirements of different production environments and needs.

[0081] 5. Continuous learning and optimization

[0082] The present invention enables the deep learning model to continuously learn and optimize by continuously collecting new data, thereby gradually improving the accuracy and efficiency of detection.

[0083] 6. Data-driven decision making

[0084] The present invention can generate detailed test reports and data analysis to help management make data-based decisions and promote the optimization of production processes.

[0085] 7. Improve product quality

[0086] The present invention can timely discover and correct quality problems in production through efficient compliance testing, thereby improving the quality of the final product and reducing after-sales problems.

[0087] 8. Promote digital transformation

[0088] The deep learning-based detection method of the present invention has promoted the digital transformation of the clothing manufacturing industry and improved the intelligence level and market competitiveness of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 Schematic diagram of a compliance detection framework for garment manufacturing processes based on deep learning in an embodiment of the present invention;

[0090] Figure 2 Schematic diagram of a method for detecting compliance of a garment manufacturing process based on deep learning in an embodiment of the present invention;

[0091] Figure 3 Schematic diagram of a garment manufacturing process compliance detection device based on deep learning in an embodiment of the present invention;

[0092] Figure 4 This is a flowchart of compliance detection of clothing manufacturing process based on deep learning in an embodiment of the present invention;

[0093] Among them, the figure numerals are: 10, acquisition unit, 20, input unit, 30, determination unit. DETAILED DESCRIPTION

[0094] In order to enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0095] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0096] It can be understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments can be combined with each other.

[0097] It can be understood that, for the convenience of description, the drawings of the present invention only show the parts related to the present invention, while the parts irrelevant to the present invention are not shown in the drawings.

[0098] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0099] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0100] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or may be implemented by a combination of hardware and computer instructions.

[0101] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0102] Embodiment 1:

[0103] like Figure 1 and Figure 2 As shown, this embodiment provides a method for compliance detection of garment manufacturing process based on deep learning. This method is a method for compliance detection of garment manufacturing process based on deep learning. It is widely used in cutting quality detection, sewing process monitoring, finished product inspection, fabric quality assessment, production process optimization, intelligent warehouse management, personalized customization and employee training, etc. It can effectively improve production efficiency, ensure product quality, and promote the intelligent and digital transformation of the industry. The method includes the following steps:

[0104] Step S1: Acquire frame images in the garment manufacturing process.

[0105] Step S1 specifically includes the following steps:

[0106] Monitor the garment manufacturing process through surveillance cameras;

[0107] Get the video stream of the surveillance camera;

[0108] Frames are extracted from the video stream according to time to obtain frame images in the clothing manufacturing process.

[0109] Acquiring the video stream of the surveillance camera means capturing and transmitting the image data of the production site in real time through the surveillance camera installed in the garment factory. This process can effectively monitor the production process and ensure the standardization and quality control of the process. The following is a detailed description of a specific embodiment in combination with the production process of down jackets.

[0110] In a down jacket production factory, multiple high-definition surveillance cameras are installed in the factory, arranged in the working areas of key processes, such as cutting, sewing, filling and inspection. Each camera is connected to a central monitoring system that can obtain video streams in real time and process and store them.

[0111] (1) Cutting process monitoring:

[0112] In the cutting area, the camera shoots the cutting machine and the fabric. The video stream clearly records the cutting process of the fabric, ensuring that each piece of down jacket fabric is cut according to the design drawings.

[0113] This embodiment detects the accuracy of cutting in real time by analyzing the video stream. If a cutting size deviation is found, the system will immediately issue an alarm to remind the operator to make adjustments.

[0114] (2) Sewing process monitoring:

[0115] In the sewing workshop, multiple cameras cover the workstations of each sewing machine. The video stream captures each worker's sewing movements to ensure the quality of the stitching and the standardization of the process.

[0116] Through image recognition technology, the system can automatically identify whether the stitching meets the standards, such as the spacing and neatness of the stitching lines. If problems are found, the monitoring system will record the relevant video clips for subsequent quality tracing.

[0117] (3) Filling process monitoring:

[0118] In the down filling area, the camera is facing the filling machine to monitor the filling amount and uniformity of the down in real time. The video stream shows every detail of the filling process to ensure that the warmth performance of each down jacket meets the standard.

[0119] This embodiment automatically calculates the filling weight of each down jacket by analyzing the data in the video stream and compares it with the standard value to ensure the consistency of the product.

[0120] (4) Inspection process monitoring:

[0121] In the finished product inspection area, the camera faces the inspection table to record the final inspection process of each down jacket. The video stream provides a detailed inspection record, including appearance, stitching, filling and other aspects.

[0122] This embodiment will compare the inspection results with the preset standards to ensure that each product meets the quality requirements. If unqualified products are found, the system will automatically mark and archive the relevant videos for subsequent analysis and improvement.

[0123] (5) Results and Effects

[0124] In this way, the factory can obtain the video stream of the surveillance camera in real time and fully monitor the production process of down jackets. The details of each process are recorded, providing reliable data support for quality control. At the same time, the intelligent analysis function of the monitoring system can detect and solve problems in a timely manner, reducing the burden of manual inspection and improving production efficiency and product quality.

[0125] The algorithm of this embodiment uses deep learning technology to design a one-stage target detection model, which is mainly used to locate the fabric, the back, the eye, the bottom edge and the illegally placed items on the workbench, and can also identify the worker's hand area. For example, the model can detect whether a piece of cloth is placed correctly on the workbench, or whether the worker's hands are performing illegal operations.

[0126] In order to improve the accuracy of the model, data enhancement and difficult sample mining were performed. Data enhancement generates more samples by rotating, scaling, and cropping existing samples, thereby improving the model's adaptability to different scenarios. Difficult sample mining focuses on samples that are prone to errors during the model training process, ensuring that the model can better learn these difficult-to-recognize features.

[0127] In order to meet the real-time requirements, the model has been compressed and quantized, distilled, and pruned. Quantization converts model parameters from floating point numbers to integers to reduce storage and computing requirements; distillation trains a small model to imitate the output of a large model, maintaining performance while reducing model size; pruning removes unimportant connections in the model to further reduce computational complexity.

[0128] In addition, a lightweight hand key point regression model is designed to quickly and accurately identify the worker's hand movements and support real-time reasoning. For example, when a worker is operating, the model can detect the position and movement of the hand in real time and determine whether the worker's operation is in compliance with the specification.

[0129] In order to enhance the adaptability of the algorithm in complex scenarios, this application designs a multimodal transformer model by combining the changes in the key points of the hand between frames with the motion characteristics of the entire picture. The model can analyze the dynamic information in the video stream and make video-level judgments on the category and compliance of the current process. For example, when a worker's hand moves to the wrong position during an operation, the model can identify it in time and issue a warning.

[0130] In terms of result display, this embodiment uses the ffmpeg streaming tool to push the algorithm processing results to the front-end page for display in real time, including picture frames with target detection and hand key point recognition. The content displayed on the front end includes the clothing, hand positions and illegal items marked in the real-time monitoring screen. At the same time, through the database and message transmission tools, the system can synchronize the process category and compliance information to the front-end display, so that the monitoring personnel can view the monitoring video and algorithm effects in real time to ensure the compliance and safety of the production process. Such a design realizes the explicit display of the algorithm effect and improves the efficiency and transparency of production management.

[0131] ffmpeg streaming tool is a powerful open source multimedia framework that can process and transmit audio and video data in real time. It supports pushing processed audio and video content to streaming media servers or front-end pages, so that users can view monitoring videos and algorithm effects in real time. With ffmpeg, users can easily record, convert and stream audio and video, providing an effective solution for real-time monitoring and data display. Combining the changes in hand key points between frames with the motion characteristics of the entire picture, the designed multimodal transformer model (transformer model is a deep learning model for processing sequence data) aims to improve the accuracy of target detection and behavior recognition in complex scenes. The model first captures the key point positions of workers' hands, such as the coordinates of fingers and palms, in real time through hand key point detection technology, and then analyzes the changes of these key points between consecutive frames to extract the dynamic features of hand movements. These changes not only reflect the specific movements of the hands, but also reveal the behavior patterns of workers when performing tasks. At the same time, the model also extracts the motion features of the entire picture, including the movement of the background and other objects. This information provides important context for understanding the dynamic state of the current scene. For example, when a worker moves his hand, other objects in the background may change, which may affect the worker's operation. By fusing the changes in the key points of the hand with the motion characteristics of the picture, the model uses the self-attention mechanism in the transformer architecture to effectively capture the relationship between different modalities, pay attention to the mutual influence between hand movements and scene dynamics, and then judge the category and compliance of the current process at the video level. Ultimately, this multimodal fusion design not only improves the recognition accuracy of hand movements, but also enhances the model's adaptability to complex scenes, enabling it to more accurately perform target detection and behavior analysis in practical applications, thereby achieving more efficient production management and safety monitoring.

[0132] Step S2: Input the frame image into a target detection model based on deep learning to obtain target detection information for each frame image.

[0133] The deep learning-based target detection model is a YoLov7 target model;

[0134] Before the step S2, the method further includes a step S0: constructing a YoLov7 target model;

[0135] The step S0 specifically includes the following steps:

[0136] Obtain training data; and build the YoLov7 initial model;

[0137] Set a minimum target for total loss value;

[0138] The YoLov7 initial model is trained according to the training data and the total loss value minimum target to obtain the YoLov7 target model.

[0139] YOLOv7 is a new version of the target detection model, which belongs to the YOLO (You Only Look Once) series and is designed to achieve efficient and real-time target detection. This model significantly improves the detection speed and accuracy by optimizing the network architecture and introducing innovative training strategies, especially in complex scenes. YOLOv7 supports multi-scale detection and can effectively identify targets of different sizes, especially the detection ability of small targets has been enhanced. In addition, YOLOv7 uses improved data enhancement technology and loss function to improve the generalization ability of the model, making it more stable in various application scenarios. This model is widely used in video surveillance, autonomous driving, industrial inspection, drone monitoring and other fields, with strong adaptability and high practicality. As an open source project, YOLOv7 has active community support. Users can easily obtain models, codes and documents for secondary development and personalized applications, which promotes the further development of computer vision technology.

[0140] As a specific implementation, the obtaining of training data specifically includes the following steps:

[0141] Collecting video or image data from historical clothing manufacturing; the video or image data includes process video or image data, clothing material video or image data, and worker hand video or image data;

[0142] Labeling the process video or image data, the clothing video or image data, and the worker hand video or image data respectively;

[0143] Performing data enhancement on the labeled process video or image data, the clothing video or image data, and the worker hand video or image data to obtain a video or image data set;

[0144] The video or image data set is divided into a training set and a test set, thereby obtaining training data consisting of the training set and the test set.

[0145] As a specific implementation manner, the total loss value is calculated by a total loss function;

[0146] The total loss function is calculated by the following formula:

[0147]

[0148] In the above calculation formula,

[0149] N is the number of detection layers;

[0150] B i is the number of targets whose labels are assigned to the prior box;

[0151] S i ×S i is the number of grids into which the current scale of the i-th target is divided;

[0152] L box is the bounding box regression loss; L obj is the target object loss; L cls is the classification loss;

[0153] L CloUj is the CloU loss of the jth target;

[0154] L objj is the target object loss of the jth target;

[0155] L clsj is the classification loss of the jth target;

[0156] λ 1 is the weight of the bounding box regression loss;

[0157] λ 2 is the weight of the target object loss;

[0158] λ 3 is the weight of classification loss;

[0159] CloU loss L CliU It is calculated by the following formula:

[0160]

[0161] In this calculation formula, b, b gt are the prediction box and label box respectively, w gt 、h gt , w, h are the width and height of the label box and the prediction box respectively, ρ represents the distance between the center points of the two boxes, v is the difference between the aspect ratio of the prediction box and the aspect ratio of the label box, α is the weight coefficient, and IoU represents the intersection over union ratio of the prediction box and the label box;

[0162] Objective loss function L obj It is calculated by the following formula:

[0163]

[0164] In this calculation formula, i represents the sample number, n represents the total number of samples, σ represents the sigmoid function, and y i represents the true label of sample i, x i Represents the output of the model.

[0165] In the target detection model, each weight (λ 1 , 2 , 3 ) usually needs to be adjusted according to the specific task and model requirements. These weights control the contribution ratio of different loss items to the total loss, thereby affecting the focus of model training. This embodiment adopts the following methods and considerations:

[0166] (1) Empirical adjustment: The initial value of the weight is usually set based on experience and experimental results. For example, if the bounding box regression loss shows a large impact on model performance in experiments, then λ may be 1 Assign a larger value.

[0167] (2) Cross-validation: Cross-validation is used to evaluate the impact of different weight combinations on model performance. By testing different weight settings on the validation set, the weight combination that optimizes model performance can be selected.

[0168] (3) Properties of loss functions: Consider the characteristics of different loss functions and their impact on model training. For example, classification loss usually has a greater impact on the classification ability of the model, so λ 3 May need to be set relatively high.

[0169] (4) Balancing loss terms: In the early stages of training, the contribution of each loss term needs to be balanced to prevent a single loss term from dominating the entire training process. As training progresses, the weights can be gradually adjusted based on model performance.

[0170] In the target detection task of this embodiment, the width and height of the label box (also called the real box or ground truth box) are known because they are manually annotated or automatically generated according to the real size of the actual target in the image. These annotations provide real information about the position and size of the target object in the image, which is usually provided in the training data set to guide the model to learn how to accurately detect and locate the target.

[0171] The width and height of the prediction box are predicted by the model based on the input image data. The model tries to predict the position and size of each object in the image by learning the patterns and features in the training data set. The accuracy of the prediction box depends on the performance of the model and the data used in the training process.

[0172] During training, the model tries to minimize the difference between the predicted box and the label box, which is usually achieved by calculating the loss function. The loss function evaluates the error between the width and height of the predicted box and the width and height of the label box, as well as other related errors (such as position error, classification error, etc.). By optimizing these losses, the model can produce more accurate prediction boxes in subsequent predictions.

[0173] In the CloU loss function, α is a weight coefficient used to balance the contribution of different components to the total loss. The value of this coefficient usually depends on the specific application scenario and experimental results. In practical applications, the value of α needs to be determined through experiments to find the best model performance.

[0174] In this embodiment, the value range of α can be any value between 0 and 1. If α is 0, then the v term does not contribute to the total loss; if α is 1, then the contribution of the v term is the same as the IoU term and the center point distance term, which means that the model will pay equal attention to the consistency of the aspect ratio of the predicted box and the label box during training, thereby improving the accuracy of the box prediction, because the model needs to pay attention not only to the position and size of the box, but also to the shape characteristics of the box.

[0175] Step S3: judging whether the garment manufacturing process is compliant based on the target detection information.

[0176] Step S3 specifically includes the following steps:

[0177] According to the process information in the target detection information, detect whether the process procedures and process materials are compliant; and,

[0178] Detecting the quality and process of the clothing material according to the clothing material information in the target detection information; and,

[0179] Based on the worker’s hands in the target detection information and combined with the hand key point regression model, the compliance of the operation is judged and the process flow is optimized.

[0180] The hand key point regression model is obtained by using data augmentation and wing_loss as the regression loss function;

[0181] The data enhancement method includes random color perturbation, random whitening, random rotation, and random mirroring;

[0182] The calculation formula of the regression loss function is as follows:

[0183]

[0184] C=ω-ωln(1+ω / ∈)

[0185] In this calculation formula, y represents the predicted result, Represents the real coordinate label;

[0186] C is a constant for adjusting the loss value;

[0187] ω represents the dividing point between the linear and nonlinear regions; ∈ represents the curvature of the control curve.

[0188] C is calculated by the above formula, and its value depends on the values ​​of ω and ∈. In theory, C should be a non-negative number, because if C is negative, then when the error exceeds ω, the value of the loss function will be lower than the maximum value in the nonlinear region, which is mathematically unreasonable.

[0189] In this embodiment, the selection of ω and ∈ usually needs to be determined through experiments to achieve the best model performance. Once ω and ∈ are determined, the value of C can be calculated by the above formula. This process may require multiple iterations and adjustments to find the parameter combination that best suits a specific task.

[0190] Therefore, the value of the constant C is determined by the values ​​of ω and ∈. Its purpose is to ensure that the loss function is continuous and differentiable in the entire domain of definition, and to maintain a certain penalty when the error is large.

[0191] The specific process of this embodiment is as follows Figure 1 and Figure 4 As shown, it specifically includes the following processes 1 to 6:

[0192] Process 1. Data collection and model training: Data preprocessing is performed based on the collected video data and image data of workers' real operations in the enterprise production environment, including data cleaning, labeling, and data enhancement. The one-stage target detection model and lightweight hand key point regression model of the deep learning algorithm are trained, and the multimodal transformer model is trained based on the changes in hand key points between frames and the motion characteristics of the entire picture to make video-level judgments on the category and compliance of the current process. A variety of model compression technologies, including quantization, distillation, and pruning, are used to compress and optimize the above algorithm models.

[0193] Process 2. Initialization operation, initialize key point coordinate storage queue, initialize multimodal process recognition model intermediate state storage queue, initialize process list and working time statistics record, initialize target detection parameters, initialize clothing ID, initialize current frame process type and previous frame process type, initialize effective process start time, initialize clothing detection start time, etc.

[0194] Process 3. Get real-time video stream from surveillance camera;

[0195] Process 4. Extract video frames and perform target recognition on each frame based on the trained one-stage target detection model, including workers' hands, clothing, illegal items and other targets. This patent designs a lightweight network structure based on the yolov7 algorithm. Based on a small sample size, it uses data enhancement methods such as color perturbation, random cropping, random mirroring, random rotation of left and right mirroring, mosic, and small target copypaste. The swish activation function is used to improve the nonlinear expression ability of the model. The positive and negative sample matching strategy of simOTA is adopted, and formula (1) is the total loss function of target detection. Formula (2) is the CIoU loss. In this patent, CIoU is used as the border regression loss. Formula (3) represents the target loss and classification loss. Finally, the accuracy on the test set reached 0.981 and the recall rate reached 0.853.

[0196]

[0197] In formula (1), N is the number of detection layers, B is the number of targets whose labels are assigned to the prior box, and S i ×S i is the number of grids into which the current scale of the i-th target is divided. box is the bounding box regression loss, calculated for each object; L obj is the target object loss, calculated for each grid; L cls is the classification loss, and λ is also calculated for each target 1 , 2 , 3 are the weights of these three losses respectively.

[0198]

[0199] In formula (2), b, b gt are the prediction box and label box respectively, w gt 、h gt , w, h are the width and height of the label box and the prediction box respectively, ρ represents the distance between the center points of the two boxes, α is the weight coefficient, and IoU represents the intersection over union ratio of the prediction box and the label box.

[0200]

[0201] In formula (3), i represents the sample number, n represents the total number of samples, σ represents the sigmoid function, and y i represents the true label of sample i, x i Represents the output of the model.

[0202] Process 5. Update the target detection parameters of the current frame. If a process non-compliance mark is detected, send a process non-compliance message to the database and upload a process non-compliance image; otherwise, continue to detect whether there is clothing in the current frame. If no clothing is detected, determine whether clothing is detected in the previous frame. If clothing is detected in the previous frame, update the clothing detection end time and clothing ID, and update the process to detection end, query the process time record, send a working time message and record it in the database, otherwise update the process to identification.

[0203] Process 6. After the clothing is detected, the employee's hands are detected. If the employee's hands are not detected, the process is updated to being identified.

[0204] Process 7. If a hand is detected, the hand key point coordinates are output based on the trained lightweight hand key point regression model, and the key point coordinate storage state is updated and the intermediate state storage sequence of the multimodal process recognition model is queried. The process type is output and the current process state is updated. For the hand key point regression task, data enhancement methods such as random color perturbation, random whitening, random rotation, and random mirroring are used, and wing_loss is used as the loss function (as shown in formulas (4) and (5)), where y represents the prediction result, Represents the true coordinate label. The average mean error nme on the test set is as low as 0.034, and it runs at nearly 200FPS on a 3090 graphics card.

[0205]

[0206] C=ω-ωln(1+ω / ∈) (5)

[0208] In formulas (4) and (5), ω represents the boundary between the linear and nonlinear regions, which is usually set to 10.0, ∈ represents the curvature of the control curve, which is usually set to 2.0, and y, represent the predicted value and the true value respectively.

[0209] Process 8. Determine whether the current process is a valid process (valid processes refer to processes other than the four processes of start, end, other, and flip). If so, determine whether the process of the previous frame is the same as the current process. If not, determine whether the previous frame is a valid process. If so, accumulate the working hours of the current process and record it in the list of completed processes, update the process duration record, synchronize it to the database, and update the multimodal process recognition model output intermediate state storage queue, update the process type and the amount of clothing in the previous frame. If the process of the current frame is the same as the previous frame, directly update the multimodal process recognition model output intermediate state storage queue, and update the process type and the amount of clothing in the previous frame. If the current process is not a valid process, directly update the multimodal process recognition model output intermediate state storage queue, and update the process type and the amount of clothing in the previous frame. The process recognition model uses the transformer architecture to design the network, generates embeddings based on the frame image and the coordinates of the key points of the hand, and then performs feature fusion. The image uses the same patch method as the visual transformer, and combines the multi-modal features of multiple frames for joint judgment, and finally achieves a top 1 accuracy of 92.55 on the test set. Formula (6) is the self-attention calculation formula, where Q, K, and V represent the Qurey, Key, and Value matrices in the original self-attention calculation, respectively.

[0210]

[0211] Process 9. The process-related data is stored in the database, and the system page obtains the data from the database for display, including four page functions: process compliance, statistical query, working time configuration, and equipment management. The process compliance page provides sub-functions such as equipment selection, real-time process progress status display, real-time working time information display, real-time monitoring video reasoning result display, real-time data statistics, and real-time alarm events; the statistical query page provides non-compliant historical data tracing according to the monitoring equipment name, non-compliant type, and time; the working time configuration page supports users to configure standard working hours according to the process, and compare with the real-time operation working hours to determine the compliance of working hours; the equipment management page provides the functions of adding, editing, deleting monitoring equipment information and opening / closing the algorithm push stream.

[0212] The specific steps of this embodiment are as follows: Figure 4 As shown, the specific steps include:

[0213] (1) Obtain real-time images from the surveillance camera. If the acquisition fails, the process ends directly. If the acquisition succeeds, the process proceeds to step (2).

[0214] (2) The target detection model locates targets such as workers’ hands, clothing, and illegal items.

[0215] (3) Update the current frame target detection parameters;

[0216] After updating the current frame target detection parameters, determine whether the process non-compliance mark is detected? If so, send an industrial non-compliance message and record it in the database, then upload the process non-compliance image and go to step (4); if not, go directly to step (4);

[0217] (4) Is clothing detected in the current frame? If yes, go to step (5); if no, go to step (17);

[0218] (5) Determine whether clothing was detected in the previous frame?

[0219] If yes, proceed to step (6); if no, update the start time of the clothing detection and then proceed to step (6);

[0220] (6) Determine whether both hands are detected

[0221] If yes, proceed to step (7); if no, update the process to recognition and reacquire the real-time image from the surveillance camera.

[0222] (7) The regression model outputs the coordinates of the first key point.

[0223] (8) Query the intermediate state storage queue of the multimodal process recognition model and update the key point coordinate storage queue.

[0224] (9) The multimodal process identification model outputs the process type and updates the current process status.

[0225] (10) Determine whether the current process is a valid process?

[0226] If yes, go to step (11); if no, go to step (15).

[0227] (11) Determine whether the previous frame process is the same as the current process?

[0228] If yes, go to step (15); if no, go to step (12).

[0229] (12) Is the previous frame a valid process?

[0230] If yes, the working hours of the current process are accumulated and recorded in the list of completed processes, the process duration record is updated, and step (13) is entered; if no, step (13) is entered directly.

[0231] (13) Send process type message and record it in the database.

[0232] (14) Update the current effective process start time.

[0233] (15) Update the intermediate state storage queue output by the multimodal process identification model.

[0234] (16) Update the process type of the previous frame and the amount of clothing in the previous frame, and then re-acquire the real-time image from the monitoring camera.

[0235] (17) Determine whether clothing was detected in the previous frame?

[0236] If so, proceed to step (18); if not, update the process to identification and reacquire the real-time image from the surveillance camera.

[0237] (18) Update the end time of clothing inspection.

[0238] (19) Update the material ID.

[0239] (20) Update the process to end inspection.

[0240] (21) Query the process duration records, send working time messages, and record them in the database.

[0241] (22) Check the process records to determine whether the process is complete?

[0242] If so, proceed to step (23); if not, send a process incomplete message and proceed to step (23).

[0243] (23) Reset the material ID, target detection parameters, process list, cumulative process hours, effective process start time, and material detection start time; then reacquire the real-time image from the surveillance camera.

[0244] Specific examples of this embodiment

[0245] Example 1: Normal operation process

[0246] Scenario: Workers are doing standard sewing work in a sewing workshop without any violation.

[0247] 1. Get real-time images: The surveillance camera successfully captures the worker’s operation images.

[0248] 2. Target detection: The system recognizes that the worker’s hands are manipulating the fabric and does not detect any illegal items.

[0249] 3. Update target detection parameters: Adjust the detection parameters based on the detection results of the current frame, and no process non-compliance marks are found.

[0250] 4. Detect clothing: Clothing is detected in the current frame and proceed to the next step.

[0251] 5. Determine the cloth detection in the previous frame: Clothes were also detected in the previous frame, so there is no need to update the start time.

[0252] 6. Detect both hands: Successfully detect the worker’s hands and enter the key point coordinate output.

[0253] 7. Output key point coordinates: The regression model outputs the key point coordinates of the worker’s hand.

[0254] 8. Update storage queue: Update the intermediate state storage queue and key point coordinate queue of the multimodal recognition model.

[0255] 9. Identify process type: The model identifies the current process as "sewing" and updates the process status.

[0256] 10. Determine the validity of the process: Confirm that the "sewing" process is a valid process.

[0257] 11. Compare the process of the previous frame: The previous frame is also a "sewing" process, skip the accumulated working hours and continue.

[0258] 12. Update intermediate state storage.

[0259] 13. Update the previous frame information: record the current process type and fabric quantity, reacquire the real-time picture, and return to the starting point of the process.

[0260] Result: The worker’s normal operation was successfully identified and recorded, and the monitoring status was maintained.

[0261] Example 2: Process non-compliance detected

[0262] Scenario: A worker violates safety regulations during operation, such as not wearing required protective gloves.

[0263] 1. Get real-time images: The surveillance camera successfully captures the worker’s operation images.

[0264] 2. Target detection: The system identifies that the worker’s hands are not wearing protective gloves and considers it as a violation.

[0265] 3. Update target detection parameters: adjust the detection parameters according to the detection results of the current frame and detect process non-compliance marks.

[0266] 4. Handle non-compliance marks: Send industrial non-compliance messages, record them in the database and upload relevant images to continue the process.

[0267] 5. Detect clothing: Clothing is detected in the current frame and the process continues.

[0268] 6. Determine the cloth detection in the previous frame: Clothes were also detected in the previous frame, so continue.

[0269] 7. Detect both hands: Both hands are detected (although not compliant), and key point coordinates are output.

[0270] 8. Output key point coordinates: The regression model outputs the key point coordinates of the worker’s hand.

[0271] 9. Update storage queue: Update the intermediate state storage queue and key point coordinate queue of the multimodal recognition model.

[0272] 10. Identify process type: The model identifies the current process as "operation" and updates the process status.

[0273] 11. Determine the validity of the process: Confirm that the "Operation" process is a valid process and continue.

[0274] 12. Compare the process of the previous frame: The previous frame is also an "operation" process, skip the accumulated working hours and continue.

[0275] 13. Update intermediate state storage: Update the intermediate state storage queue.

[0276] 14. Update the previous frame information: record the current process type and fabric quantity, reacquire the real-time picture, and return to the starting point of the process.

[0277] Result: Violations are promptly identified, details are recorded, and relevant departments are notified for handling.

[0278] Example 3: Failed to obtain the surveillance camera

[0279] Scenario: The surveillance camera fails to obtain real-time images due to a network failure.

[0280] Get real-time images: If the surveillance camera fails to obtain the images, the process ends directly.

[0281] Result: The system failed to monitor and manual inspection of the camera equipment was required.

[0282] Example 4: Insufficient process completeness

[0283] Scenario: A worker does not complete all necessary sewing steps.

[0284] 1. Get real-time images: The surveillance camera successfully acquires the images and the system starts detection.

[0285] 2. Target detection and process identification: The system identifies some sewing steps but misses some key processes.

[0286] 3. Update target detection parameters: adjust the parameters based on the detection results of the current frame, and no non-compliant markers are detected.

[0287] 4. Detect clothing: Clothing is detected in the current frame and proceed to the next step.

[0288] 5. Determine the cloth detection in the previous frame: Clothes were also detected in the previous frame, so continue.

[0289] 6. Detect both hands: Successfully detect the worker’s hands and enter the key point coordinate output.

[0290] 7. Output key point coordinates: The regression model outputs the key point coordinates of the worker’s hand.

[0291] 8. Update storage queue: Update the intermediate state storage queue and key point coordinate queue of the multimodal recognition model.

[0292] 9. Identify process type: The model identifies the current process as "sewing" and updates the process status.

[0293] 10. Determine the validity of the process: Confirm that the "Sewing" process is a valid process and continue.

[0294] 11. Compare the process of the previous frame: The previous frame is also a "sewing" process, skip the accumulated working hours and continue.

[0295] 12. Update intermediate state storage: Update the intermediate state storage queue.

[0296] 13. Update the previous frame information: record the current process type and fabric quantity, reacquire the real-time picture, and return to the starting point of the process.

[0297] 14. Query process duration records: Query process duration records, send working time messages, and record them in the database.

[0298] 15. Determine process completeness: Query the process records that have been made. If the process is found to be incomplete, send a process incomplete message.

[0299] 16. Reset system status: reset the material ID, target detection parameters, process list, accumulated working hours, effective process start time and material detection start time, and re-acquire the real-time image.

[0300] Result: The system identified incomplete processes, provided timely feedback and recorded them, and ensured the integrity of the production process.

[0301] Example 5: Process switching and accumulated working hours

[0302] Scenario: After completing one process, a worker immediately starts the next process.

[0303] 1. Get real-time images: The surveillance camera successfully captures the worker’s operation images.

[0304] 2. Object detection and process identification: The system recognizes that the worker switches from "sewing" to "quality inspection".

[0305] 3. Update target detection parameters: adjust the detection parameters according to the detection results of the current frame, and no non-compliant markers are detected.

[0306] 4. Detect clothing: Clothing is detected in the current frame and proceed to the next step.

[0307] 5. Determine the cloth detection in the previous frame: Clothes were also detected in the previous frame, so there is no need to update the start time.

[0308] 6. Detect both hands: Successfully detect the worker’s hands and enter the key point coordinate output.

[0309] 7. Output key point coordinates: The regression model outputs the key point coordinates of the worker’s hand.

[0310] 8. Update storage queue: Update the intermediate state storage queue and key point coordinate queue of the multimodal recognition model.

[0311] 9. Identify process type: The model identifies the current process as "quality inspection" and updates the process status.

[0312] 10. Determine the validity of the process: Confirm that the "Quality Inspection" process is a valid process and continue.

[0313] 11. Compare the process of the previous frame: the previous frame is "sewing", the current process is different and the previous frame is valid.

[0314] 12. Determine the validity of the previous frame process: the previous frame is a valid process "sewing", accumulate the "sewing" working hours, record them in the list of completed processes, update the process duration record, and go to step 13.

[0315] 13. Record process: Send a "sewing" process type message and record it in the database.

[0316] 14. Update process start time: Update the start time of the current effective process "Quality Inspection".

[0317] 15. Update intermediate state storage: Update the intermediate state storage queue output by the multimodal process recognition model.

[0318] 16. Update the previous frame information: record the current process type "quality inspection" and the amount of clothing, re-acquire the real-time picture, and return to the starting point of the process.

[0319] Result: The system accurately records process switching and labor time accumulation, ensuring that the time and completion status of each process are recorded in detail.

[0320] Example 6: Process status updated to "Inspection completed"

[0321] Scenario: After the worker completes all necessary sewing and quality inspection steps, the process enters the end state.

[0322] 1. Get real-time images: The surveillance camera successfully captures the worker’s operation images.

[0323] 2. Object detection: The system recognizes that the worker’s hands are completing the final quality inspection step and no illegal items are detected.

[0324] 3. Update target detection parameters: Adjust the detection parameters based on the detection results of the current frame, and no non-compliant markers are found.

[0325] 4. Detect clothing: Clothing is detected in the current frame and proceed to the next step.

[0326] 5. Determine the cloth detection in the previous frame: Clothes were also detected in the previous frame, so there is no need to update the start time.

[0327] 6. Detect both hands: Successfully detect the worker’s hands and enter the key point coordinate output.

[0328] 7. Output key point coordinates: The regression model outputs the key point coordinates of the worker’s hand.

[0329] 8. Update storage queue: Update the intermediate state storage queue and key point coordinate queue of the multimodal recognition model.

[0330] 9. Identify process type: The model identifies the current process as "quality inspection" and updates the process status.

[0331] 10. Determine the validity of the process: Confirm that the "Quality Inspection" process is a valid process and continue.

[0332] 11. Compare the process of the previous frame: the previous frame is the "Quality Inspection" process, if they are the same, skip the accumulated working hours and continue.

[0333] 12. Update intermediate state storage: Update the intermediate state storage queue.

[0334] 13. Update the previous frame information: record the current process type and fabric quantity, reacquire the real-time picture, and return to the starting point of the process.

[0335] 14. Clothes not detected in the current frame:

[0336] 15. Determine whether clothing was detected in the previous frame.

[0337] 16. Update the end time of clothing inspection.

[0338] 17. Update fabric ID.

[0339] 18.Update process status: Update the process status to "Inspection completed".

[0340] 19. Send working time message: query the process duration record, send working time message, and record it in the database.

[0341] 20. Determine process completeness: Check the process records, confirm that the process is complete, and go to step 23.

[0342] 21. Reset system status: reset the material ID, target detection parameters, process list, accumulated working hours, effective process start time and material detection start time, re-acquire the real-time image, and return to the starting point of the process.

[0343] Result: The system recognizes that the process has been completed, records the working hours and resets the system status, preparing for the monitoring of the next process.

[0344] Through the above detailed description and multiple examples, this embodiment can effectively monitor the workers' process execution in various operation scenarios, promptly identify and handle abnormal situations, and ensure the compliance and efficiency of the production process. Whether in normal operation, violation detection, equipment failure, process switching, or process completion, this embodiment can efficiently and accurately complete the process compliance detection of the apparel manufacturing industry, thereby improving overall production efficiency and product quality, and optimizing production management. In addition, the process compliance detection system provided by this method has the characteristics of real-time detection, high accuracy, and adaptability to complex scenarios, so that it can provide a reliable compliance detection solution in an actual production environment.

[0345] Embodiment 2:

[0346] like Figure 1 As shown, this embodiment provides a production management method for the clothing manufacturing industry, and the method includes the following steps:

[0347] According to the clothing manufacturing process compliance detection method based on deep learning described in Example 1, the process of the clothing manufacturing industry is subjected to compliance detection to obtain the detection result;

[0348] Performing statistical query on the detection results to obtain statistical query results;

[0349] According to the statistical query results, work time allocation is performed to obtain a work time allocation result;

[0350] According to the working time allocation result, maintenance and servicing reminders are given to the production equipment of the clothing manufacturing industry to realize the production management of the clothing manufacturing industry.

[0351] Embodiment 3:

[0352] like Figure 3 As shown, this embodiment provides a device for detecting compliance of clothing manufacturing process based on deep learning, and the device includes:

[0353] The acquisition unit 10 is used to acquire frame images in the garment manufacturing process;

[0354] An input unit 20, connected to the acquisition unit 10, is used to input the frame image into a target detection model based on deep learning to obtain target detection information for each frame image;

[0355] The determination unit 30 is connected to the input unit 20 and is used to determine whether the garment manufacturing process is compliant based on the target detection information.

[0356] In specific implementation, the device includes a construction unit, which is connected to the input unit and is responsible for building the YoLov7 target model. The construction unit specifically includes several modules: the acquisition module is used to collect relevant training data to ensure that the model can learn enough features; the construction module is used to create the initial model of YoLov7 to provide a basic structure for subsequent training; the setting module is used to set a total loss value minimization target to guide the optimization direction of model training. The training module is connected to the acquisition module, the construction module and the setting module respectively, and is responsible for training the YoLov7 initial model according to the collected training data and the set loss value target, thereby generating the final YoLov7 target model so that it can effectively perform target detection tasks.

[0357] The device in this embodiment can execute the method in embodiment 1.

[0358] Embodiment 4:

[0359] This embodiment provides a production management system for the clothing manufacturing industry, the system comprising:

[0360] A detection device, wherein the detection device is the clothing manufacturing process compliance detection device based on deep learning in Example 3, and the detection device is used to perform compliance detection on the process of the clothing manufacturing industry to obtain a detection result;

[0361] A statistical query device, connected to the detection device, for performing a statistical query on the detection result to obtain a statistical query result;

[0362] A working time configuration device, connected to the statistical query device, is used to perform working time configuration according to the statistical query result to obtain a working time configuration result;

[0363] The equipment management device is connected to the working time configuration device and is used to provide maintenance and servicing prompts for the production equipment of the clothing manufacturing industry according to the working time configuration results, so as to realize the production management of the clothing manufacturing industry.

[0364] This embodiment provides a production management system for the clothing manufacturing industry, which includes multiple key components: first, a detection device, which is based on deep learning technology and is derived from Example 3, and is specifically used to perform compliance detection on each process of the clothing manufacturing industry, thereby generating corresponding detection results; second, a statistical query device, which is connected to the detection device and is responsible for performing in-depth statistical analysis on the detection results to obtain comprehensive statistical query results; next, a working time configuration device, which is connected to the statistical query device and performs reasonable configuration of working hours based on the statistical query results, thereby obtaining effective working time configuration results; finally, an equipment management device, which is connected to the working time configuration device and provides maintenance and maintenance tips for production equipment in the clothing manufacturing industry based on the working time configuration results. The collaborative work of this series of devices is aimed at achieving efficient production management and improving the overall operational efficiency and compliance of the clothing manufacturing industry.

[0365] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for compliance detection of garment manufacturing process based on deep learning, characterized in that: The method comprises the following steps: Step S1: Acquire frame images in the garment manufacturing process; Step S2: inputting the frame image into a target detection model based on deep learning to obtain target detection information for each frame image; Step S3: judging whether the garment manufacturing process is compliant based on the target detection information.

2. The method for compliance detection of garment manufacturing process based on deep learning according to claim 1 is characterized in that: The step S1 specifically includes the following steps: Monitor the garment manufacturing process through surveillance cameras; Get the video stream of the surveillance camera; Frames are extracted from the video stream according to time to obtain frame images in the clothing manufacturing process.

3. The method for compliance detection of garment manufacturing process based on deep learning according to claim 1 is characterized in that: The step S3 specifically includes the following steps: According to the process information in the target detection information, detect whether the process steps and process materials are compliant to obtain a first detection result; and According to the clothing information in the target detection information, detect whether the clothing quality and the clothing process are compliant to obtain a second detection result; and According to the worker's hand in the target detection information, the compliance of the operation is judged in combination with the hand key point regression model to obtain a third detection result; Whether the garment manufacturing process is compliant is determined based on the first test result, the second test result and the third test result.

4. The method for detecting compliance of garment manufacturing process based on deep learning according to claim 3 is characterized in that: The hand key point regression model is obtained by using data enhancement method and wing_loss as regression loss function; The data enhancement method includes random color perturbation, random whitening, random rotation, and random mirroring; The calculation formula of the regression loss function is as follows: C=ω-ωln(1+ω / ∈) (1) In formula (1), y represents the prediction result, Represents the real coordinate label; C is a constant for adjusting the loss value; ω represents the dividing point between the linear and nonlinear regions; ∈ represents the curvature of the control curve.

5. The method for detecting compliance of garment manufacturing process based on deep learning according to any one of claims 1 to 4, characterized in that: The deep learning-based target detection model is a YoLov7 target model; Before the step S2, the method further includes a step S0: constructing a YoLov7 target model; The step S0 specifically includes the following steps: Obtain training data; and build the YoLov7 initial model; Set a minimum target for total loss value; The YoLov7 initial model is trained according to the training data and the total loss value minimum target to obtain the YoLov7 target model.

6. The method for detecting compliance of garment manufacturing process based on deep learning according to claim 5 is characterized in that: The obtaining of training data specifically includes the following steps: Collecting video or image data from historical clothing manufacturing; the video or image data includes process video or image data, clothing material video or image data, and worker hand video or image data; Labeling the process video or image data, the clothing video or image data, and the worker hand video or image data respectively; Performing data enhancement on the labeled process video or image data, the clothing video or image data, and the worker hand video or image data to obtain a video or image data set; The video or image data set is divided into a training set and a test set, thereby obtaining training data consisting of the training set and the test set.

7. The method for compliance detection of garment manufacturing process based on deep learning according to claim 5 is characterized in that: The total loss value is calculated by the total loss function; The total loss function is calculated by the following formula (2): In formula (2), N is the number of detection layers; B i is the number of targets whose labels are assigned to the prior box; S i ×S i is the number of grids into which the current scale of the i-th target is divided; L box is the bounding box regression loss; L obj is the target object loss; L cls is the classification loss; L CloUj is the CloU loss of the jth target; L objj is the target object loss of the jth target; L clsj is the classification loss of the jth target; λ1 is the weight of the bounding box regression loss; λ2 is the weight of target object loss; λ3 is the weight of classification loss; CloU loss L CloU It is calculated by the following formula (3): In formula (3), b, b gt are the prediction box and label box respectively, w gt 、h gt , w, h are the width and height of the label box and the prediction box respectively, ρ represents the distance between the center points of the two boxes, v is the difference between the aspect ratio of the prediction box and the aspect ratio of the label box, α is the weight coefficient, and IoU represents the intersection over union ratio of the prediction box and the label box; Objective loss function L obj It is calculated by the following formula (4): In formula (4), i represents the sample number, n represents the total number of samples, σ represents the sigmoid function, and y i represents the true label of sample i, x i Represents the output of the model.

8. A production management method for the clothing manufacturing industry, characterized in that: The method comprises the following steps: According to the method for compliance detection of clothing manufacturing process based on deep learning according to any one of claims 1 to 7, compliance detection is performed on the process of clothing manufacturing industry to obtain detection results; Performing statistical query on the detection results to obtain statistical query results; According to the statistical query results, work time allocation is performed to obtain a work time allocation result; According to the working time allocation result, maintenance and servicing reminders are given to the production equipment of the clothing manufacturing industry to realize the production management of the clothing manufacturing industry.

9. A device for detecting compliance of garment manufacturing process based on deep learning, characterized in that: include: An acquisition unit, used for acquiring frame images in the garment manufacturing process; An input unit, connected to the acquisition unit, for inputting the frame image into a target detection model based on deep learning to obtain target detection information for each frame image; The determination unit is connected to the input unit and is used to determine whether the garment manufacturing process is compliant based on the target detection information.

10. A production management system for the clothing manufacturing industry, characterized in that: include: A detection device, wherein the detection device is the garment manufacturing process compliance detection device based on deep learning according to claim 9, and the detection device is used to perform compliance detection on the garment manufacturing process to obtain a detection result; A statistical query device, connected to the detection device, for performing a statistical query on the detection result to obtain a statistical query result; A working time configuration device, connected to the statistical query device, is used to perform working time configuration according to the statistical query result to obtain a working time configuration result; The equipment management device is connected to the working time configuration device and is used to provide maintenance and servicing prompts for the production equipment of the clothing manufacturing industry according to the working time configuration results, so as to realize the production management of the clothing manufacturing industry.

Images