Circular knitting machine cloth cover defect online detection device and detection method based on main / auxiliary double camera shooting mode

Abstract

The invention discloses a circular knitting machine cloth cover defect online detection device based on a main / auxiliary double-camera mode, which comprises a machine table, an industrial personal computer is arranged in one machine leg of the machine table, a system console is arranged at the upper part of the machine leg, an industrial camera module is arranged at the upward position of the inner side of the machine leg, and a rotatable cloth rolling frame is fixed on the machine table; the cloth rolling frame is connected to the circular needle cylinder on the upper portion through two cloth rolling frame linkage rods, a hanging rod is arranged in the center of the top of the machine table, a backlight lamp is installed below the hanging rod, cameras and edge computing platforms which are connected with each other are installed on the cloth rolling frame linkage rods on the left side and the right side, and an annular cloth supporting ring is installed on the lower side of the top of the machine table. The invention further discloses a detection method of the circular knitting machine cloth cover defect online detection device based on the main / auxiliary double camera shooting mode. According to the invention, the problems of missing detection and lagging in online detection of cloth surface defects of a circular weft knitting machine in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of knitting textile machinery, and specifically relates to an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode. The present invention also relates to a detection method for the online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode. Background Art

[0002] Circular knitting machines may produce defects due to various reasons during the weaving process. There are two main methods for detecting defects. One is offline detection. After weaving, the cloth is inspected manually or by a cloth inspection machine. This cannot be detected in real time, is time-consuming and labor-intensive, and will cause a lot of waste once a defect is detected. The other is online detection, but the internal space of the circular knitting machine is narrow, and the rotary weaving of the needle cylinder makes the woven grey cloth also a rotating sealed cylindrical structure, which is not suitable for placing the detection device; most of the current online detection methods are placed on the outside of the rotating cylindrical part, that is, a camera is installed on the bottom of the circular machine to collect the image data of the cloth in real time, and judge whether there are defects based on the image data. This avoids the waste of offline detection and saves time, but this method has a problem. The camera cannot detect the cloth surface directly behind the cloth roll linkage rod on the cloth roll, and there is a certain problem of missed detection.

[0003] In addition, the patent application with the publication number CN115639143A discloses a cloth detection device and method suitable for an open-width cloth winding machine, which sets the transmission spacing of the cloth by installing the first rubber roller group and the second rubber roller group in parallel on the frame, and arranges at least two camera modules along the length direction of the mounting seat to monitor the production status of the cloth in real time, thereby improving the accuracy and comprehensiveness of the monitoring. Although this method realizes the real-time and comprehensive monitoring of the production status of the cloth, it is only applicable to open-width cloth winding machines, and is not applicable to most knitting circular weft machines. In addition, this method requires about 2 meters of cloth to be lowered from the appearance of defects to their discovery, and there is also a certain lag.

[0004] Therefore, how to conduct online inspection of circular knitting machine fabrics more conveniently, comprehensively and timely to avoid missed inspections and use more circular knitting machines has become an urgent problem to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode, which solves the problems of missed detection and lag in the online detection of circular knitting machine fabric surface defects in the prior art.

[0006] Another object of the present invention is to provide a detection method for an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera method.

[0007] The first technical solution adopted by the present invention is an online detection device for fabric defects of a circular knitting machine based on a main / auxiliary dual camera method, including a machine platform, an industrial computer is placed inside one of the machine feet of the machine platform, the upper part of the machine foot is a system console, an industrial camera module is installed on the upper inner side of the machine foot, a rotatable cloth winding rack is fixed on the machine platform, the cloth winding rack is connected to the upper circular needle cylinder through two cloth winding rack linkage rods, a hanging rod is arranged at the center position of the top of the machine platform, a backlight is installed on the machine platform below the hanging rod, interconnected cameras and edge computing platforms are installed on the cloth winding rack linkage rods on the left and right sides, and an annular cloth supporting ring is installed on the lower side of the top of the machine platform.

[0008] The first technical solution of the present invention is also characterized in that: The camera is located at the vertical center of the upper end of the cloth rolling rack linkage rod. The center of the camera shooting is located between the two rings on the lower side of the cloth supporting ring, and is used to collect the cloth surface image directly behind that is blocked by the cloth rolling rack linkage rod.

[0009] The specific structure of the industrial camera module is: it includes an industrial camera, and the lens of the industrial camera passes through the circular hole in the middle of the light source.

[0010] The backlight is located on the inner side of the cloth support ring close to the cloth surface, and illuminates the cloth from the reverse side, so that the defects on the reverse side of the fabric can be presented on the front side through the illumination of the backlight, making it convenient for the industrial camera module (4) and the camera to capture an ideal fabric image.

[0011] The second technical solution adopted by the present invention is a detection method of an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode, which is specifically implemented according to the following steps: Step 1: Collect image data and mark it; Step 2: Use the YOLOv5 model to perform initial training on the labeled image data; Step 3: Optimize and improve the YOLOv5 model; Step 4: Retrain according to the improved YOLOv5 model; Step 5, obtaining the woven cloth image data in real time; Step 6: The edge computing platform and the industrial computer use the deployed weight file to perform AI detection and classification on the obtained image. After the detection is completed, it will display whether the current fabric has defects. If so, it will display the current defect type; Step 7. The edge computing platform and the internal program of the industrial computer judge the defect type obtained from the detection results. If the defect type is a broken needle, bad needle, wrong yarn, broken yarn, or other defects that seriously affect product quality, it is sent to the system console through Wifi or serial port communication to execute a shutdown operation; if the defect type of the detection result is not one of the above serious defect types, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to stop and continues to run.

[0012] The second technical solution of the present invention is also characterized in that: Step 1 is as follows: Collect existing cloth image data with and without defects, and perform preprocessing of the collected image data by rotation, cropping, scaling, adding noise, and color dithering. After the preprocessing is completed, the images with defects are divided into 11 types of defects commonly produced by circular knitting machines, including broken needles, bad needles, wrong yarns, broken yarns, holes, flower needles, hanging needles, burst yarns, sparse needles, cross-roads, and others. The defects in the images are annotated using the labelimg image annotation tool in the Anaconda environment, and the annotations are in txt format; Step 2 is as follows: Set the number of model training epochs to 300, that is, the model completely traverses the data set 300 times, the image size imgsz is 640, that is, the size of the input image is 640*640, the number of worker threads is 4, which is responsible for loading data in parallel during training, and the batch size batch-size defines the number of images processed in each batch, which is set to 8. After the training is completed, a pt format weight file is obtained, which contains the weights and biases of the convolutional layer, model architecture information, learning rate, number of rounds, accuracy of each round, loss value, optimizer status, momentum, and learning rate scheduler during model training.

[0013] Step 3 is as follows: Step 3.1, use K-means clustering algorithm instead of Euclidean distance in YOLOv5 to accurately measure the similarity between sample markers, replace the distance metric between samples with 1-IoU, so that the initial anchor frame is closer to the actual size of various marker frames, and then combine genetic algorithm GA to perform 3000 rounds of anchor frame optimization, and finally determine 9 prior anchor frame sizes, and the best anchor frame achieves a recall rate of more than 99.9%; Step 3.2: Introduce the CA attention mechanism into the C3 module of the YOLOv5 model to form the C3CA module. The original C3 modules at the 13th and 20th layers of the model are replaced by the C3CA module. After the feature map X is sent to CA, the size of the feature map X is C×H×W. The module first performs horizontal and vertical average pooling on each channel through two sets of pooling kernels (H, 1) and (1, W) respectively to obtain two feature maps. The output of the Cth channel at height h and width w is as shown in formulas (1) and (2): Then f = δ(F1([Z h ,Z W])) These two feature maps are processed by a shared 1×1 convolution kernel F1 to obtain an intermediate feature map with a size of C / r×1×(H+W), where δ represents the nonlinear activation function, f represents the intermediate feature map encoding horizontal and vertical spatial information, and r is the downsampling rate; then f is split into f along the spatial dimension h and f w Two tensors, f h The dimensions are C / r×H×1, f w The size is C / r×W×1; finally, two 1×1 convolutions F h and F w f h and f w Using g h =σ(F h (f h )) and g w =σ(F w (f w )) is transformed back to the same channel size as the input X, σ is the sigmoid activation function, and finally, the CA attention mechanism is Output.

[0014] Step 3.3, combined with the network layer structure of separation convolution, GSConv with fusion SC, DSC and Shuffle hybrid strategy is used to replace the SC module of the Neck layer of the YOLOv5 module. After the input channels enter the network, a regular convolution layer is used to reduce the number of channels to Half of the feature map is then processed by separate convolution, and finally the two feature maps are combined to form The feature map of the channel; the data is reorganized by Shuffle. The time complexity of GSConv processing the feature map is about half of that of the standard convolution. The specific time complexity formulas are shown in (3) and (4): (3) (4) Among them, W and H represent the width and height of the feature map respectively. and represents the size of the convolution kernel, is the number of channels of the input feature map, is the number of channels of the output feature map; Step 3.4: Enhance the network performance by using a bidirectional feature pyramid network BiFPN with efficient bidirectional cross-scale connections and weighted features, as shown in formula (5): (5) The Relu activation function is used to ensure that each feature weight Non-negative, learning rate is 0.0001. By improving BiFPN, the 6th layer in BackBone of the YOLOv5 model and the 19th layer in Neck are connected by Add to strengthen the feature fusion of intermediate scale defects. The fusion formula at the 19th layer is shown in formula (6): (6) In the formula, For the improved network The output features of the layer, Conv is the convolution operation.

[0015] Step 4 is as follows: The obtained pt weight file is converted into onnx format, and then converted into rknn format, which is the format required for edge computing platform deployment, and deployed to the edge computing platform. At the same time, the pt weight file is directly deployed on the industrial computer, so that the edge computing platform and the industrial computer can perform defect detection on the image; Step 5 is as follows: The camera installed on the linkage rod of the circular knitting machine's cloth winding frame and the industrial camera module installed on the upward side of the machine acquire the image data of the woven cloth in real time. At this time, the backlight illuminates from the back of the cloth, so that the defects on the back of the fabric can be presented on the front through the backlight, which is convenient for the industrial camera module and the camera to capture ideal fabric images. The light source illuminates the part of the cloth that the camera collects, providing lighting conditions for the camera. The light hood covers the light source and the camera to prevent the external environment light source from affecting the camera's photography, and gathers the light emitted by the light source to facilitate the camera to better capture the cloth image. The cloth support ring supports the cloth to form an ideal ring, so that the appearance of the cloth can be fully presented. At the same time, it is convenient for the industrial camera module to maintain the same focal length when shooting and obtain clear images.

[0016] The beneficial effect of the present invention is that the online detection device for circular knitting machine fabric defects based on the main / secondary dual camera mode collects the fabric image of the fabric behind the fabric roll frame linkage rod by installing cameras and edge computing platforms on the left and right sides of the knitting circular knitting machine fabric roll frame linkage rod device, and installs an industrial camera on the inner side of the machine facing upward to shoot other fabrics, so that the fabric image woven by the circular knitting machine is fully and completely collected to avoid missed detection. The collected image data is detected and classified in real time on the edge device and industrial computer using the YOLOv5 target detection algorithm, and the detected information is transmitted to the control console, completing the online real-time, comprehensive and efficient detection of fabric defects on the circular knitting machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a three-dimensional schematic diagram of the present invention; Figure 2 It is a schematic diagram of the installation of the industrial camera of the present invention; Figure 3 It is a schematic diagram of the camera structure of the present invention; Figure 4 is a schematic diagram of a camera and an edge device of the present invention; Figure 5 It is a schematic diagram of the backlight source installation and the cloth support ring of the present invention; Figure 6 It is a flow chart of the implementation of the present invention; Figure 7 This is the image processing flow of the edge device and industrial computer of the present invention.

[0018] In the figure, 1. circular knitting machine, 2. cloth roll frame, 3. cloth roll frame linkage rod, 4. industrial camera module, 41. industrial camera, 42. lens, 43. light source, 44. light shield, 5. camera, 6. edge device, 7. circular needle cylinder, 8. backlight, 9. cloth support ring, 10. suspension rod, 11. system console. DETAILED DESCRIPTION

[0019] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The invention discloses an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode, the structure of which is as follows: Figure 1 , Figure 2 As shown, it includes a machine 1, an industrial computer is placed inside one of the machine feet of the machine 1, and the upper part of the machine foot is a system console 11. The system console 11 controls the circular knitting machine, starts the detection device, triggers the industrial camera and the camera to take pictures, receives signals from the industrial computer and the edge device 6, and makes corresponding action instructions to the circular knitting machine according to the signals. An industrial camera module 4 is installed on the upper side of the machine foot. A rotatable cloth roll 2 is fixed on the machine 1. The cloth roll 2 is connected to the upper circular needle cylinder 7 through two cloth roll linkage rods 3. A suspension rod 10 is set at the center of the top of the machine 1, combined with Figure 4 , Figure 5 , a backlight 8 is installed under the boom, and interconnected cameras 5 and edge computing platforms 6 are installed on the left and right cloth roll rack linkage rods 3. The cameras and edge computing platforms 6 are installed on the left and right cloth roll rack linkage rods 3 and combined with the industrial camera 4 to collect and detect images in dual-camera mode, record and keep the detection results, and then send signals to the system console to control the weaving process of the circular knitting machine. An annular cloth support ring 9 is installed on the lower side of the top of the machine 1. The annular cloth support ring 9 props up the cloth surface to form an ideal ring, which is convenient for the full presentation of the cloth surface appearance, and at the same time, it is convenient for the industrial camera module 4 to maintain the same focal length when shooting and obtain clear shooting images.

[0021] The edge computing platform 6 can deploy and train the YOLO model and is responsible for processing the data collected by the camera. The images collected by the camera 5 are detected using the YOLOv5 model in the edge computing platform 6. If a defect is detected, it is necessary to further determine the defect type and choose whether to stop the machine based on the defect type, and transmit the information to the system console.

[0022] The camera 5 is located at the vertical center of the upper end of the cloth rolling rack linkage rod 3. The center of the camera 5 is located between the two rings on the lower side of the cloth supporting ring 9, and is used to collect the cloth surface image directly behind that is blocked by the cloth rolling rack linkage rod.

[0023] The deployment and training of the YOLOv5 model, the YOLOv5 model is a target detection algorithm, which improves the processing speed by simplifying the target detection task into a single neural network prediction, and has a high accuracy. It is necessary to collect a large number of fabric defects in advance, classify them and do annotation training. After continuous optimization and improvement of the model and enrichment of the data set, finally a best weight file with high accuracy and low missed detection rate is obtained and deployed on the edge computing platform 6 to detect the fabric image obtained by the camera 5 in real time.

[0024] like Figure 3 As shown, the specific structure of the industrial camera module 4 is as follows: it includes an industrial camera 41, and the lens 42 of the industrial camera 41 passes through the central circular hole of the light source 43. The industrial camera module 4 is used to collect images of all the annular cloth surfaces except those blocked by the two cloth roll rack linkage rods 3, transmit the collected image data to the industrial computer, and use the YOLOv5 algorithm to detect and classify the collected images, record and save the defect information, and for serious defects, the user can choose whether to stop the machine. The YOLOv5 model used in the industrial computer detection can use the same data set and weight file as the model deployed on the edge computing platform 6, and detect the images collected by the industrial camera module 4 on the industrial computer.

[0025] The backlight 8 is located on the inner side of the cloth support ring 9 near the cloth surface, and illuminates the cloth from the reverse side, so that the defects on the reverse side of the fabric can be presented on the front side through the illumination of the backlight, making it convenient for the industrial camera module 4 and the camera 5 to capture an ideal fabric image.

[0026] Combination Figure 6 , Figure 7 The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode is specifically implemented according to the following steps: Step 1: Collect image data and mark it; Step 1 is as follows: Collect existing cloth image data with and without defects, rotate, crop, scale, add noise, and perform color jitter preprocessing on the collected image data to enrich the data set and improve the generalization and robustness of the model. After the preprocessing is completed, the images with defects are divided into broken needles, bad needles, wrong yarns, broken yarns, holes, flower needles, hanging needles, burst yarns, sparse needles, cross-roads, and other 11 types of defects commonly produced by circular knitting machines according to the defect type. The defects in the images are annotated using the labelimg image annotation tool in the Anaconda environment, and the annotations are in txt format; Step 2: Use the YOLOv5 model to perform initial training on the labeled image data; Step 2 is as follows: Set the number of model training epochs to 300, that is, the model completely traverses the data set 300 times, the image size imgsz is 640, that is, the size of the input image is 640*640, the number of worker threads is 4, which is responsible for loading data in parallel during training, and the batch size batch-size defines the number of images processed in each batch, which is set to 8. After the training is completed, a pt format weight file is obtained, which contains the weights and biases of the convolutional layer, model architecture information, learning rate, number of rounds, accuracy of each round, loss value, optimizer status, momentum, and learning rate scheduler during model training.

[0027] Step 3: Based on the weight file obtained from model training, when the model performs well on the training set but poorly on the validation set, it is necessary to reduce the learning rate or the number of training rounds to prevent overfitting. Reducing the learning rate can make the model more stable in the later stages of training and avoid overfitting caused by excessive update steps. When the model performs poorly on both the training set and the validation set, it is necessary to increase the learning rate or adjust the model architecture, such as increasing the number of model layers or neurons, to improve the learning ability of the model. When the model training is unstable or converges slowly, it is necessary to adjust the momentum parameter of the optimizer or replace the optimizer. After these adjustments, a weight file with good stability, fast speed, and high accuracy based on the original YOLOv5 model training is obtained.

[0028] After completion, the obtained weight file was tested and it was found that the model was still lacking in detection accuracy and speed for online detection of circular knitting machine defects, so the YOLOv5 model was optimized and improved; Step 3 is as follows: Step 3.1: Use K-means clustering algorithm instead of Euclidean distance in YOLOv5 to accurately measure the similarity between sample markers. Replace the distance metric between samples with 1-IoU (intersection over union of marker and initial anchor) to make the initial anchor closer to the actual size of various markers. Then, use genetic algorithm GA to optimize anchors for 3000 rounds. Finally, 9 prior anchor sizes are determined, and the best anchor achieves a recall rate of more than 99.9%. Step 3.2: Introduce the CA attention mechanism into the C3 module of the YOLOv5 model to form the C3CA module. The original C3 modules at the 13th and 20th layers of the model are replaced by the C3CA module. After the feature map X is sent to CA, the size of the feature map X is C×H×W. The module first performs horizontal and vertical average pooling on each channel through two sets of pooling kernels (H, 1) and (1, W) respectively to obtain two feature maps. The output of the Cth channel at height h and width w is as shown in formulas (1) and (2): (1) (2) Subsequently by The two feature maps are processed by a shared 1×1 convolution kernel F1 to obtain an intermediate feature map with a size of ,in, represents a nonlinear activation function, represents the intermediate feature map encoding horizontal and vertical spatial information, is the downsampling rate; then Split along the spatial dimension into and Two tensors, Size , Size ; Finally, through two 1×1 convolutions and Will and use and Transform back to the same channel size as the input X, is the sigmoid activation function. Finally, the CA attention mechanism is implemented by Output. It effectively improves the features that are beneficial to the model weight and suppresses redundant features.

[0029] Step 3.3, combined with the network layer structure of separation convolution, GSConv with fusion SC, DSC and Shuffle hybrid strategy is used to replace the SC module of the Neck layer of the YOLOv5 module. After the input channels enter the network, a regular convolution layer is used to reduce the number of channels to Half of the feature map is then processed by separate convolution, and finally the two feature maps are combined to form The feature map of the channel; Shuffle is used to reorganize the data to make the data more random, thereby further improving the generalization of the network. The time complexity of GSConv processing feature maps is about half of that of standard convolution. The specific time complexity formulas are shown in (3) and (4): (3) (4) Among them, W and H represent the width and height of the feature map respectively. and represents the size of the convolution kernel, is the number of channels of the input feature map, In order to output the number of channels of the feature map, GSConv can be used to extract the defect image of the target fabric more quickly and accurately, and reduce the number of parameters during training.

[0030] Step 3.4: Enhance network performance through a bidirectional feature pyramid network (BiFPN) with efficient bidirectional cross-scale connections and weighted features. Based on the traditional feature pyramid network (FPN), BiFPN is optimized to remove nodes with only a single input edge to reduce unnecessary calculations. In addition, when the fusion node is at the same level as the input node, BiFPN adds additional connections to fuse more feature information without increasing the computational cost. When aggregating nodes at different levels, BiFPN uses a conversion function , adjust the size of the feature map by upsampling and downsampling to achieve feature fusion at different resolutions. Unlike the traditional FPN that treats input features equally, BiFPN introduces a weight mechanism to fuse the features of different nodes through a fast normalization fusion method, as shown in formula (5): (5) The Relu activation function is used to ensure that each feature weight Non-negative, learning rate is 0.0001, which improves the running efficiency of the model on the GPU. By improving BiFPN, the 6th layer in BackBone of the YOLOv5 model and the 19th layer in Neck are connected by Add to strengthen the feature fusion of intermediate scale defects. The fusion formula at the 19th layer is shown in formula (6): (6) In the formula, For the improved network The output features of the layer, Conv is the convolution operation; After the above improvements, compared with the original YOLOv5 model, the average single image defect detection speed is 0.9ms faster and the accuracy is 2.8 percentage points higher. Step 4: Retrain according to the improved YOLOv5 model; Step 4 is as follows: The obtained pt weight file is converted into onnx format, and then converted into rknn format, which is the format required for edge computing platform deployment, and deployed to the edge computing platform. At the same time, the pt weight file is directly deployed on the industrial computer, so that the edge computing platform and the industrial computer can perform defect detection on the image; Step 5, obtaining the woven cloth image data in real time; Step 5 is as follows: The camera 5 installed on the connecting rod 3 of the circular knitting machine's cloth winding rack and the industrial camera module 4 installed on the upward side of the machine obtain the woven cloth image data in real time through the edge computing platform and the industrial computer driver. At this time, the backlight 8 is illuminated from the back of the cloth, so that the defects on the back of the fabric can be presented on the front through the backlight, which is convenient for the industrial camera module 4 and the camera 5 to shoot ideal fabric images, and the light source 43 illuminates the part of the cloth collected by the camera to provide lighting conditions for the camera. The light shield 44 covers the light source and the camera to prevent the external environment light source from affecting the camera's photography, and gathers the light emitted by the light source to facilitate the camera to better capture the cloth image. The cloth support ring 9 supports the cloth to form an ideal ring, so that the appearance of the cloth is fully presented, and at the same time, it is convenient for the industrial camera module 4 to maintain the same focal length when shooting to obtain clear shooting images.

[0031] Step 6. The edge computing platform and the industrial computer use the deployed weight file to perform AI detection and classification on the obtained image. After the detection is completed, it will display whether the current fabric has defects. If so, the current defect type will be displayed; when one or more needle loops fall off, small holes or uncircled patterns appear on the fabric, usually manifested as vertical lines and holes on the cloth surface, which is called broken needles (long missed needles); bad needles are caused by damaged knitting needles or damaged needle hooks, resulting in abnormal loop formation during weaving, and vertical lines appear on the cloth surface; due to the use of incorrect yarns, horizontal stripes on the cloth surface are called wrong yarns; broken yarn refers to the spandex core yarn or outer covering yarn, which is broken or entangled in other yarn groups during the weaving process, which is usually manifested as horizontal stripes on the cloth surface.

[0032] Step 7. The edge computing platform and the internal program of the industrial computer judge the defect type obtained from the detection results. If the defect type is a broken needle, bad needle, wrong yarn, broken yarn, or other defects that seriously affect product quality, it is sent to the system console through Wifi or serial port communication to execute a shutdown operation; if the defect type of the detection result is not one of the above serious defect types, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to stop and continues to run.

[0033] Example 1 The invention discloses an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode, the structure of which is as follows: Figure 1 , Figure 2 As shown, it includes a machine 1, an industrial computer is placed inside one of the machine feet of the machine 1, and the upper part of the machine foot is a system console 11. The system console 11 controls the circular knitting machine, starts the detection device, triggers the industrial camera and the camera to take pictures, receives signals from the industrial computer and the edge device 6, and makes corresponding action instructions to the circular knitting machine according to the signals. An industrial camera module 4 is installed on the upper side of the machine foot. A rotatable cloth roll 2 is fixed on the machine 1. The cloth roll 2 is connected to the upper circular needle cylinder 7 through two cloth roll linkage rods 3. A suspension rod 10 is set at the center of the top of the machine 1, combined with Figure 4 , Figure 5 , a backlight 8 is installed under the boom, and interconnected cameras 5 and edge computing platforms 6 are installed on the left and right cloth roll rack linkage rods 3. The cameras and edge computing platforms 6 are installed on the left and right cloth roll rack linkage rods 3 and combined with the industrial camera 4 to collect and detect images in dual-camera mode, record and keep the detection results, and then send signals to the system console to control the weaving process of the circular knitting machine. An annular cloth support ring 9 is installed on the lower side of the top of the machine 1. The annular cloth support ring 9 props up the cloth surface to form an ideal ring, which is convenient for the full presentation of the cloth surface appearance, and at the same time, it is convenient for the industrial camera module 4 to maintain the same focal length when shooting and obtain clear shooting images.

[0034] Example 2 The invention discloses an online detection device for circular knitting machine fabric surface defects based on a main / auxiliary dual camera mode, the structure of which is as follows: Figure 1 , Figure 2 As shown, it includes a machine 1, an industrial computer is placed inside one of the machine feet of the machine 1, and the upper part of the machine foot is a system console 11. The system console 11 controls the circular knitting machine, starts the detection device, triggers the industrial camera and the camera to take pictures, receives signals from the industrial computer and the edge device 6, and makes corresponding action instructions to the circular knitting machine according to the signals. An industrial camera module 4 is installed on the upper side of the machine foot. A rotatable cloth roll 2 is fixed on the machine 1. The cloth roll 2 is connected to the upper circular needle cylinder 7 through two cloth roll linkage rods 3. A suspension rod 10 is set at the center of the top of the machine 1, combined with Figure 4 , Figure 5 , a backlight 8 is installed on the machine 1 below the boom, and interconnected cameras 5 and edge computing platforms 6 are installed on the left and right cloth roll rack linkage rods 3. The cameras and edge computing platforms 6 are installed on the left and right cloth roll rack linkage rods 3 and combined with the industrial camera 4 to collect and detect images in dual-camera mode, record and keep the detection results, and then send signals to the system console to control the weaving process of the circular knitting machine. An annular cloth support ring 9 is installed on the lower side of the top of the machine 1. The annular cloth support ring 9 props up the cloth surface to form an ideal ring, which is convenient for the full presentation of the cloth surface appearance, and at the same time, it is convenient for the industrial camera module 4 to maintain the same focal length when shooting and obtain clear shooting images.

[0035] The edge computing platform 6 can deploy and train the YOLO model and is responsible for processing the data collected by the camera. The images collected by the camera 5 are detected using the YOLOv5 model in the edge computing platform 6. If a defect is detected, it is necessary to further determine the defect type and choose whether to stop the machine based on the defect type, and transmit the information to the system console.

[0036] The camera 5 is located at the vertical center of the upper end of the cloth rolling rack linkage rod 3. The center of the camera 5 is located between the two rings on the lower side of the cloth supporting ring 9, and is used to collect the cloth surface image directly behind that is blocked by the cloth rolling rack linkage rod.

[0037] Example 3 Combination Figure 6 , Figure 7 The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode is specifically implemented according to the following steps: Step 1: Collect image data and mark it; Step 2: Use the YOLOv5 model to perform initial training on the labeled image data; Step 3: Optimize and improve the YOLOv5 model; Step 4: Retrain according to the improved YOLOv5 model; Step 5, obtaining the woven cloth image data in real time; Step 6: The edge computing platform and the industrial computer use the deployed weight file to perform AI detection and classification on the obtained image. After the detection is completed, it will display whether the current fabric has defects. If so, it will display the current defect type; Step 7. The edge computing platform and the internal program of the industrial computer judge the defect type obtained from the detection results. If the defect type is a broken needle, bad needle, wrong yarn, broken yarn, or other defects that seriously affect product quality, it will be sent to the system console via Wifi or serial port communication to immediately shut down the machine. If the defect type of the detection result is not one of the above serious defect types, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to be shut down and continues to run.

[0038] Example 4 Combination Figure 6 , Figure 7 The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode is specifically implemented according to the following steps: Step 1: Collect image data and mark it; Step 1 is as follows: Collect existing cloth image data with and without defects, rotate, crop, scale, add noise, and perform color jitter preprocessing on the collected image data to enrich the data set and improve the generalization and robustness of the model. After the preprocessing is completed, the images with defects are divided into broken needles, bad needles, wrong yarns, broken yarns, holes, flower needles, hanging needles, burst yarns, sparse needles, cross-roads, and other 11 types of defects commonly produced by circular knitting machines according to the defect type. The defects in the images are annotated using the labelimg image annotation tool in the Anaconda environment, and the annotations are in txt format; Step 2: Use the YOLOv5 model to perform initial training on the labeled image data; Step 2 is as follows: Set the number of model training epochs to 300, that is, the model completely traverses the data set 300 times, the image size imgsz is 640, that is, the size of the input image is 640*640, the number of worker threads is 4, which is responsible for loading data in parallel during training, and the batch size batch-size defines the number of images processed in each batch, which is set to 8. After the training is completed, a pt format weight file is obtained, which contains the weights and biases of the convolutional layer, model architecture information, learning rate, number of rounds, accuracy of each round, loss value, optimizer status, momentum, and learning rate scheduler during model training.

[0039] Step 3: Optimize and improve the YOLOv5 model; Step 3 is as follows: Step 3.1: Use K-means clustering algorithm instead of Euclidean distance in YOLOv5 to accurately measure the similarity between sample markers. Replace the distance metric between samples with 1-IoU (intersection over union of marker and initial anchor) to make the initial anchor closer to the actual size of various markers. Then, use genetic algorithm GA to optimize anchors for 3000 rounds. Finally, 9 prior anchor sizes are determined, and the best anchor achieves a recall rate of more than 99.9%. Step 3.2: Introduce the CA attention mechanism into the C3 module of the YOLOv5 model to form the C3CA module. The original C3 modules at the 13th and 20th layers of the model are replaced by the C3CA module. After the feature map X is sent to CA, the size of the feature map X is C×H×W. The module first performs horizontal and vertical average pooling on each channel through two sets of pooling kernels (H, 1) and (1, W) respectively to obtain two feature maps. The output of the Cth channel at height h and width w is as shown in formulas (1) and (2): (1) (2) Subsequently by The two feature maps are processed by a shared 1×1 convolution kernel F1 to obtain an intermediate feature map with a size of ,in, represents a nonlinear activation function, represents the intermediate feature map encoding horizontal and vertical spatial information, is the downsampling rate; then Split along the spatial dimension into and Two tensors, Size , Size ; Finally, through two 1×1 convolutions and Will and use and Transform back to the same channel size as the input X, is the sigmoid activation function. Finally, the CA attention mechanism is implemented by Output. It effectively improves the features that are beneficial to the model weight and suppresses redundant features.

[0040] Step 3.3, combined with the network layer structure of separation convolution, GSConv with fusion SC, DSC and Shuffle hybrid strategy is used to replace the SC module of the Neck layer of the YOLOv5 module. After the input channels enter the network, a regular convolution layer is used to reduce the number of channels to Half of the feature map is then processed by separate convolution, and finally the two feature maps are combined to form The feature map of the channel; Shuffle is used to reorganize the data to make the data more random, thereby further improving the generalization of the network. The time complexity of GSConv processing feature maps is about half of that of standard convolution. The specific time complexity formulas are shown in (3) and (4): (3) (4) Among them, W and H represent the width and height of the feature map respectively. and represents the size of the convolution kernel, is the number of channels of the input feature map, In order to output the number of channels of the feature map, GSConv can be used to extract the defect image of the target fabric more quickly and accurately, and reduce the number of parameters during training.

[0041] Step 3.4: Enhance network performance through a bidirectional feature pyramid network (BiFPN) with efficient bidirectional cross-scale connections and weighted features. Based on the traditional feature pyramid network (FPN), BiFPN is optimized to remove nodes with only a single input edge to reduce unnecessary calculations. In addition, when the fusion node is at the same level as the input node, BiFPN adds additional connections to fuse more feature information without increasing the computational cost. When aggregating nodes at different levels, BiFPN uses a conversion function , adjust the size of the feature map by upsampling and downsampling to achieve feature fusion at different resolutions. Unlike the traditional FPN that treats input features equally, BiFPN introduces a weight mechanism to fuse the features of different nodes through a fast normalization fusion method, as shown in formula (5): (5) The Relu activation function is used to ensure that each feature weight Non-negative, learning rate is 0.0001, which improves the running efficiency of the model on the GPU. By improving BiFPN, the 6th layer in BackBone of the YOLOv5 model and the 19th layer in Neck are connected by Add to strengthen the feature fusion of intermediate scale defects. The fusion formula at the 19th layer is shown in formula (6): (6) In the formula, For the improved network The output features of the layer, Conv is the convolution operation; After the above improvements, compared with the original YOLOv5 model, the average single image defect detection speed is 0.9ms faster and the accuracy is 2.8 percentage points higher. Step 4: Retrain according to the improved YOLOv5 model; Step 5, obtaining the woven cloth image data in real time; Step 6: The edge computing platform and the industrial computer use the deployed weight file to perform AI detection and classification on the obtained image. After the detection is completed, it will display whether the current fabric has defects. If so, it will display the current defect type; Step 7. The edge computing platform and the internal program of the industrial computer judge the defect type obtained from the detection results. If the defect type is a broken needle, bad needle, wrong yarn, broken yarn, or other defects that seriously affect product quality, it will be sent to the system console via Wifi or serial port communication to immediately shut down the machine. If the defect type of the detection result is not one of the above serious defect types, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to be shut down and continues to run.

[0042] Example 5 Combination Figure 6 , Figure 7 The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode is specifically implemented according to the following steps: Step 1: Collect image data and mark it; Step 1 is as follows: Collect existing cloth image data with and without defects, rotate, crop, scale, add noise, and perform color jitter preprocessing on the collected image data to enrich the data set and improve the generalization and robustness of the model. After the preprocessing is completed, the images with defects are divided into broken needles, bad needles, wrong yarns, broken yarns, holes, flower needles, hanging needles, burst yarns, sparse needles, cross-roads, and other 11 types of defects commonly produced by circular knitting machines according to the defect type. The defects in the images are annotated using the labelimg image annotation tool in the Anaconda environment, and the annotations are in txt format; Step 2: Use the YOLOv5 model to perform initial training on the labeled image data; Step 3: Based on the weight file obtained from model training, when the model performs well on the training set but poorly on the validation set, it is necessary to reduce the learning rate or the number of training rounds to prevent overfitting. Reducing the learning rate can make the model more stable in the later stages of training and avoid overfitting caused by excessive update steps. When the model performs poorly on both the training set and the validation set, it is necessary to increase the learning rate or adjust the model architecture, such as increasing the number of model layers or neurons, to improve the learning ability of the model. When the model training is unstable or converges slowly, it is necessary to adjust the momentum parameter of the optimizer or replace the optimizer. After these adjustments, a weight file with good stability, fast speed, and high accuracy based on the original YOLOv5 model training is obtained.

[0043] After completion, the obtained weight file was tested and it was found that the model was still lacking in detection accuracy and speed for online detection of circular knitting machine defects, so the YOLOv5 model was optimized and improved; Step 4: Retrain according to the improved YOLOv5 model; Step 4 is as follows: The obtained pt weight file is converted into onnx format, and then converted into rknn format, which is the format required for edge computing platform deployment, and deployed to the edge computing platform. At the same time, the pt weight file is directly deployed on the industrial computer, so that the edge computing platform and the industrial computer can perform defect detection on the image; Step 5, obtaining the woven cloth image data in real time; Step 5 is as follows: The camera 5 installed on the connecting rod 3 of the circular knitting machine's cloth winding frame and the industrial camera module 4 installed on the upward inner side of the machine foot obtain the woven cloth image data in real time. At this time, the backlight 8 is irradiated from the back of the cloth, so that the defects on the back of the fabric can be presented on the front through the backlight, which is convenient for the industrial camera module 4 and the camera 5 to shoot ideal fabric images. The light source 43 illuminates the part of the cloth collected by the camera to provide lighting conditions for the camera. The light shield 44 shields the light source and the camera to prevent the external environment light source from affecting the camera's photography, and gathers the light emitted by the light source to facilitate the camera to better capture the cloth image. The cloth supporting ring 9 props up the cloth to form an ideal ring, so that the appearance of the cloth can be fully presented. At the same time, it is convenient for the industrial camera module 4 to maintain the same focal length when shooting to obtain clear shooting images.

[0044] Step 6. The edge computing platform and the industrial computer use the deployed weight file to perform AI detection and classification on the obtained image. After the detection is completed, it will display whether the current fabric has defects. If so, the current defect type will be displayed; when one or more needle loops fall off, small holes or uncircled patterns appear on the fabric, usually manifested as vertical lines and holes on the cloth surface, which is called broken needles (long missed needles); bad needles are caused by damaged knitting needles or damaged needle hooks, resulting in abnormal loop formation during weaving, and vertical lines appear on the cloth surface; due to the use of incorrect yarns, horizontal stripes on the cloth surface are called wrong yarns; broken yarn refers to the spandex core yarn or outer covering yarn, which is broken or entangled in other yarn groups during the weaving process, which is usually manifested as horizontal stripes on the cloth surface.

[0045] Step 7. The edge computing platform and the internal program of the industrial computer judge the defect type obtained from the detection results. If the defect type is a broken needle, bad needle, wrong yarn, broken yarn, or other defects that seriously affect product quality, it will be sent to the system console via Wifi or serial port communication to immediately shut down the machine. If the defect type of the detection result is not one of the above serious defect types, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to be shut down and continues to run.

[0046] Example 6 Reference Figure 7 , this embodiment lists the data processing flow of the edge device and the industrial computer, including the following steps: Step 1: Dataset preparation: It is necessary to collect pictures containing various types of fabric defects and various different fabrics, and perform data enhancement such as rotation, cropping, scaling, adding noise, and color jittering on the collected pictures to enrich the dataset and improve the generalization ability and robustness of the model.

[0047] Step 2: Label the dataset. Use the labelimg image labeling tool in the Anaconda environment to label the defects in the image and convert the labeled dataset into the format required for YOLO training.

[0048] Step 3: Model training: Use the YOLOv5 training script to train the model. According to the training results, select a file with the best weight as the weight file for subsequent detection.

[0049] Step 4: Deploy on edge devices or industrial computers. Deploy the selected weight file on the edge computing platform and industrial computer, and prepare to process the collected images on the edge computing platform and industrial computer.

[0050] The weight file reflects the model training results and includes the model's convolutional layer weights, biases, and optimizer status, enabling the model to be deployed in different environments and devices to perform real-time target detection and classification tasks. It is the key for edge computing platforms and industrial computers to detect and classify image data.

[0051] Step 5: The collected data is tested online. The image data collected by the camera and the camera head are sent to the edge computing platform and the industrial computer for online testing.

[0052] Step 6: Logical judgment is performed to determine whether the defect type is serious based on the detection result. If the defect type detected is broken needle, bad needle, wrong yarn, broken yarn, etc., it means that the defect type will seriously affect the product quality, and a signal needs to be sent to the circular knitting machine console to stop the machine.

[0053] If the defect type of the test result is not one of the serious defect types mentioned above, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to be shut down and can continue to run.

[0054] According to the description of the on-line detection device and method of circular knitting machine fabric defects based on the main / auxiliary dual camera mode in the specific implementation, the method can realize on-line detection of circular knitting machine fabric defects, and can effectively avoid the problem of missed detection caused by the occlusion of the cloth rolling rack linkage rod. The use of backlight can better detect the reverse defects of the fabric. The present invention can comprehensively, accurately and quickly perform on-line detection of circular knitting machine fabric defects, effectively solving the defects of the existing detection methods.

Claims

1. A circular knitting machine fabric surface defect online detection device based on main / auxiliary dual camera mode, characterized in that: The machine (1) comprises a machine platform (1), wherein an industrial computer is placed inside one of the machine feet of the machine platform (1), a system control console (11) is located on the upper part of the machine foot, an industrial camera module (4) is installed on the inner side of the machine foot facing upward, a rotatable cloth roll rack (2) is fixed on the machine platform (1), the cloth roll rack (2) is connected to a circular needle cylinder (7) on the upper part through two cloth roll rack linkage rods (3), a suspension rod (10) is arranged at the center of the top of the machine platform (1), a backlight (8) is installed below the suspension rod, mutually connected cameras (5) and edge computing platforms (6) are installed on the cloth roll rack linkage rods (3) on the left and right sides, and an annular cloth support ring (9) is installed on the lower side of the top of the machine platform (1).

2. The on-line detection device for fabric defects of a circular knitting machine based on the main / auxiliary dual camera method according to claim 1 is characterized in that: The camera (5) is located at the vertical center of the upper end of the cloth rolling rack linkage rod (3), and the center of the camera (5) is located between the two rings on the lower side of the cloth supporting ring (9), and is used to capture the image of the cloth surface directly behind that is blocked by the cloth rolling rack linkage rod.

3. The online detection device for circular knitting machine fabric surface defects based on main / auxiliary dual camera mode according to claim 2 is characterized in that: The specific structure of the industrial camera module (4) is as follows: it comprises an industrial camera (41), and the lens (42) of the industrial camera (41) passes through the central circular hole of the light source (43).

4. The on-line detection device for circular knitting machine fabric surface defects based on main / auxiliary dual camera mode according to claim 3 is characterized in that: The backlight (8) is located on the inner side of the cloth support ring (9) close to the cloth surface, and illuminates the cloth from the reverse side, so that defects on the reverse side of the fabric can be presented on the front side through the illumination of the backlight, making it convenient for the industrial camera module (4) and the camera (5) to capture an ideal fabric image.

5. A detection method for online detection of fabric defects on a circular knitting machine based on a main / auxiliary dual camera method, characterized in that: The online detection device for circular knitting machine fabric defects based on the main / auxiliary dual camera method according to any one of claims 1 to 4 is implemented in accordance with the following steps: Step 1: Collect image data and mark it; Step 2: Use the YOLOv5 model to perform initial training on the labeled image data; Step 3: Optimize and improve the YOLOv5 model; Step 4: Retrain according to the improved YOLOv5 model; Step 5, obtaining the woven cloth image data in real time; Step 6: The edge computing platform and the industrial computer use the deployed weight file to perform AI detection and classification on the obtained image. After the detection is completed, it will display whether the current fabric has defects. If so, it will display the current defect type; Step 7. The edge computing platform and the internal program of the industrial computer judge the defect type obtained from the detection results. If the defect type is a broken needle, bad needle, wrong yarn, broken yarn, or other defects that seriously affect product quality, it will be sent to the system console via Wifi or serial port communication to immediately shut down the machine. If the defect type of the detection result is not one of the above serious defect types, the defect information needs to be recorded and saved for subsequent processing, and the machine does not need to be shut down and continues to run.

6. The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode according to claim 5 is characterized in that: The step 1 is specifically as follows: The existing cloth image data with and without defects are collected, and the collected image data are preprocessed by rotation, cropping, scaling, adding noise, and color dithering. After the preprocessing is completed, the images with defects are divided into broken needles, bad needles, wrong yarns, broken yarns, holes, flower needles, hanging needles, burst yarns, sparse needles, cross-roads and other 11 types of defects commonly produced by circular knitting machines according to the defect types. The defects in the images are annotated using the labelimg image annotation tool in the Anaconda environment, and the annotations are in txt format.

7. The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode according to claim 6 is characterized in that: The step 2 is specifically as follows: Set the number of model training epochs to 300, that is, the model completely traverses the data set 300 times, the image size imgsz is 640, that is, the size of the input image is 640*640, the number of worker threads is 4, which is responsible for loading data in parallel during training, and the batch size batch-size defines the number of images processed in each batch, which is set to 8. After the training is completed, a pt format weight file is obtained, which contains the weights and biases of the convolutional layer, model architecture information, learning rate, number of rounds, accuracy of each round, loss value, optimizer status, momentum, and learning rate scheduler during model training.

8. The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode according to claim 7 is characterized in that: The step 3 is as follows: Step 3.1: Use the K-means clustering algorithm instead of the Euclidean distance in YOLOv5 to accurately measure the similarity between sample markers. Replace the distance metric between samples with 1-IoU to make the initial anchor frame closer to the actual size of various marker frames. Then combine the genetic algorithm GA to perform 3000 rounds of anchor frame optimization, and finally determine 9 prior anchor frame sizes, and the best anchor frame achieves a recall rate of more than 99.9%; Step 3.2: Introduce the CA attention mechanism into the C3 module of the YOLOv5 model to form the C3CA module. The original C3 modules at the 13th and 20th layers of the model are replaced by the C3CA module. After the feature map X is sent to CA, the size of the feature map X is C×H×W. The module first performs horizontal and vertical average pooling on each channel through two sets of pooling kernels (H, 1) and (1, W) respectively to obtain two feature maps. The output of the Cth channel at height h and width w is as shown in formulas (1) and (2): （1） （2） Subsequently by The two feature maps are processed by a shared 1×1 convolution kernel F1 to obtain an intermediate feature map with a size of ,in, represents a nonlinear activation function, represents the intermediate feature map encoding horizontal and vertical spatial information, is the downsampling rate; then Split along the spatial dimension into and Two tensors, Size , Size ; Finally, through two 1×1 convolutions and Will and use and Transform back to the same channel size as the input X, is the sigmoid activation function. Finally, the CA attention mechanism is implemented by Output; Step 3.3, combined with the network layer structure of separation convolution, GSConv with fusion SC, DSC and Shuffle hybrid strategy is used to replace the SC module of the Neck layer of the YOLOv5 module. After the input channels enter the network, a regular convolution layer is used to reduce the number of channels to Half of the feature map is then processed by separate convolution, and finally the two feature maps are combined to form The feature map of the channel; the data is reorganized by Shuffle. The time complexity of GSConv in processing the feature map is about half of that of the standard convolution. The specific time complexity formulas are shown in (3) and (4): （3） （4） Among them, W and H represent the width and height of the feature map respectively. and represents the size of the convolution kernel, is the number of channels of the input feature map, is the number of channels of the output feature map; Step 3.4: Enhance the network performance by using a bidirectional feature pyramid network BiFPN with efficient bidirectional cross-scale connections and weighted features, as shown in formula (5): （5） The Relu activation function is used to ensure that each feature weight Non-negative, learning rate is 0.0001. By improving BiFPN, the 6th layer in BackBone of the YOLOv5 model and the 19th layer in Neck are connected by Add to strengthen the feature fusion of intermediate scale defects. The fusion formula at the 19th layer is shown in formula (6): （6） In the formula, For the improved network The output features of the layer, Conv is the convolution operation.

9. The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode according to claim 8 is characterized in that: The step 4 is specifically as follows: The obtained pt weight file is converted into onnx format, and then converted into rknn format, which is the format required for edge computing platform deployment and deployed to the edge computing platform. At the same time, the pt format weight file is directly deployed on the industrial computer, so that the edge computing platform and the industrial computer can perform defect detection on the image.

10. The detection method of the circular knitting machine fabric surface defect online detection device based on the main / auxiliary dual camera mode according to claim 9, characterized in that: The step 5 is specifically as follows: The camera (5) installed on the connecting rod (3) of the circular knitting machine's cloth winding frame and the industrial camera module (4) installed on the inner side of the machine upward obtain the woven cloth image data in real time through the edge computing platform and the industrial computer drive. At this time, the backlight (8) irradiates from the back side of the cloth, so that the defects on the back side of the fabric can be presented on the front side through the backlight, which is convenient for the industrial camera module (4) and the camera (5) to capture an ideal fabric image. The light source (43) illuminates the part of the cloth that the camera collects, providing lighting conditions for the camera. The light shield (44) shields the light source and the camera to prevent the external environment light source from affecting the camera's photography, and gathers the light emitted by the light source to facilitate the camera to better capture the cloth image. The cloth support ring (9) supports the cloth to form an ideal ring, so that the appearance of the cloth is fully presented. At the same time, it is convenient for the industrial camera module (4) to maintain the same focal length during shooting and obtain clear shooting images.

Images