Reed wire detection equipment and detection method

Abstract

The invention discloses reed wire detection equipment. The reed wire detection equipment comprises a rack, a driving component and an image acquisition component, the rack comprises a workbench making contact with a reed, linear guide rails are installed at the two ends of the workbench, and the driving component can drive the detection component to move along the linear guide rails. The image acquisition part is used for acquiring reed surface images of the reed; the invention further discloses a reed wire detection method which comprises the following steps: acquiring an original reed wire picture through the image acquisition part, and establishing a reed wire data set; labeling the reed wire data set, and inputting the reed wire data set into a Yolody model for training after data enhancement; collecting reed surface images in real time, performing image splicing on the reed surface real-time images through an SIFT algorithm to obtain a reed surface complete image, and then inputting the complete reed image into the trained Yologue model to obtain the position of each reed wire; and counting the number of the reed wires. According to the invention, the detection speed and precision of the reed wire can be effectively improved, and the requirements of the reed on-line detection scene are met.

Description

Technical Field

[0001] The present invention belongs to the technical field of textile machinery detection devices, relates to a reed and reed blade detection device, and also relates to a reed and reed blade detection method based on deep learning. Background Art

[0002] As an important part of textile machinery, the reed is mainly made of steel wires of the same specification that are wound and connected with reed blades of the same length and thickness. The number of reed blades determines the number of warp threads that can pass through a unit length, which directly affects the tightness, width, air permeability, wear resistance and other properties of the fabric. Therefore, in weaving production, it is necessary to select the appropriate number and arrangement of reed blades according to specific needs and loom types.

[0003] With the continuous development and technological progress of the textile machinery industry, the amplitude of the steel reed has been continuously lengthened, resulting in an increasing number of reed blades. The detector of the reed braiding machine can determine the number of reed blades. However, due to factors such as the accumulation of reed chips and shutdown due to failures, the detection of reed blades is easily inaccurate. In addition, during the production process of the steel reed, the problematic reed blades need to be replaced, and missed detection is inevitable. At present, the number of steel reed blades unloaded from the reed braiding machine is mainly counted manually, which depends on the experience and concentration of the operator. It is easily affected by factors such as fatigue and distraction, resulting in large detection errors and low work efficiency. Problems such as, it is difficult to meet the dual requirements of speed and accuracy of modern production. Therefore, how to efficiently and accurately detect steel reed blades has become an urgent problem to be solved in the field of textile machinery detection. Summary of the invention

[0004] The purpose of the present invention is to provide a steel reed blade detection device, which can realize the rapid detection of the steel reed blade and effectively improve the detection speed and accuracy of the steel reed blade.

[0005] The present invention also aims to provide a method for detecting reed blades.

[0006] The first technical solution adopted by the present invention is a reed blade detection device, which includes a frame, a driving component, and an image acquisition component. The frame includes a workbench in contact with the reed, and the workbench is fixedly connected to the bottom bracket; the driving component can drive the image acquisition component to move along the X, Y, and Z directions;

[0007] The image acquisition component includes a CCD camera, which is used to acquire the surface image of the reed; the image acquisition component is connected to the driving component through a synchronous belt clamp;

[0008] It also includes an LED light strip as a light source, which is arranged on the workbench and along the length direction of the reed.

[0009] The present invention is also characterized in that:

[0010] The frame includes a pair of Y-direction linear guide rails fixedly connected to both sides of the horizontal plane of the workbench, the Y-direction linear guide rails are connected to the I-type slider, the I-type slider is connected to the top bracket through the I-type slider connecting plate, and the top bracket is fixedly connected to the side plate;

[0011] The driving component includes a stepper motor fixedly connected to the side plate, the main shaft of the stepper motor is connected to the driving wheel, the driving wheel is connected to the driven wheel through a synchronous belt, and the driven wheel and the driving wheel are located at the same height;

[0012] The synchronous belt clamp is connected to the type II slider connecting plate, the type II slider connecting plate is connected to the type II slider, the type II slider is installed on the X-direction linear guide rail, and the X-direction linear guide rail is fixedly connected to the vertical surface of the side plate;

[0013] The type II slider connecting plate is connected to the Z-axis linear guide rail, the Z-axis linear guide rail is connected to the type III slider, and the type III slider is connected to the camera adjustment frame through screws; the image acquisition component includes a camera adjustment block connected to the type III slider, the camera adjustment block is fixedly connected to the CCD camera, and the CCD camera is connected to the lens.

[0014] The I-type slider has a tightening hole on its side, which is used to fix the I-type slider; the III-type slider has a tightening hole on its side, which is used to fix the III-type slider, and an adjustable tightening handle is installed in the tightening hole, and the tightening handle can fix the position of the I-type slider on the X-guide rail;

[0015] The side plate is formed by welding two steel plates, and a driven wheel mounting groove is opened on the vertical surface of the side plate, and the mounting groove is used to fix the driven wheel mounting shaft.

[0016] The main shaft of the stepper motor is connected to the driving synchronous pulley through a key bar and bolts; the driven wheel is installed on the driven wheel mounting shaft, a pair of bearings are arranged between the driven wheel and the mounting shaft, and the driven wheel mounting shaft is installed in the driven wheel mounting groove on the side plate of the frame through screws;

[0017] The tooth profiles of the synchronous belt and the pulley are circular arc teeth;

[0018] The synchronous belt clamp is provided with a slot hole and a bolt hole. The synchronous belt is placed in the slot hole of the synchronous belt clamp. The synchronous belt clamp is connected to the II-type slider connecting plate by bolts. The synchronous belt clamp fixes the synchronous belt by clamping the bolts.

[0019] The image acquisition component can drive the top bracket to move along the Y direction of the workbench through the I-type slider, the camera can move along the X direction of the workbench through the synchronous belt, and the camera can move along the Z direction of the workbench through the Z-direction linear guide.

[0020] The second technical solution adopted by the present invention is a method for detecting the number of reed blades based on deep learning, comprising the following steps:

[0021] S1, obtaining the original image of the reed blade through the image acquisition part and establishing the reed blade data set;

[0022] S2, label the reed blade dataset through Labelimage;

[0023] S3, perform Imgaug data enhancement on the annotated reed blade dataset, increase training samples, and improve the model's adaptability to reed blade images in different environments;

[0024] S4, inputting the data-enhanced reed dataset into the deep learning instance segmentation Yolodent model for training to obtain an optimized Yolodent model;

[0025] S5, collecting the reed surface image in real time, stitching the real-time reed surface image by SIFT algorithm to obtain a complete reed surface image, and then inputting the complete reed image into the optimized Yolodent model to obtain the position of each reed blade;

[0026] S6. Count the number of reed blades and output the number of reed blades through communication.

[0027] In step S3, the Imgaug data enhancement includes reed image preprocessing, and the preprocessing includes five image processing functions: rotation Rotate, horizontal flip FlipLR, vertical flip FlipUD, brightness adjustment Multiply, and adding Gaussian noise Additive Gaussian Noise.

[0028] Step S4 specifically includes:

[0029] S41, loading the reed piece training set data pre-training model weights to achieve transfer learning and reduce the amount of data set training;

[0030] S42, the image data enters the backbone network, and the network layers such as CBS, ELAN, and MP-1 are responsible for extracting the deep features in the image. The convolution kernels of the CBS module are set to 1×1, 3×3, and 3×3, respectively, and the step sizes are set to 1, 1, and 2, respectively. The size of the pooling layer of the MP-1 module is 3×3, generating the feature information of the reed piece;

[0031] S43. The feature map output by the backbone network is sent to the PAFPN network, and multi-scale feature extraction and fusion are completed through the ELAN-W module, UPSample module and SPPCSPC module to enhance the richness and robustness of the reed blade features.

[0032] S44. The feature map after PAFPN fusion is sent to the detection head, and the detection head identifies the type and position of the reed piece.

[0033] S45. Calculate the loss between the predicted result and the true annotation, update the model weight through the back propagation algorithm, and optimize the model performance.

[0034] S46. Use the validation set to evaluate the model’s accuracy, recall and other performance, adjust the hyperparameters, data enhancement strategy or model structure based on the evaluation results, and perform iterative training.

[0035] S47. Save the trained model weights.

[0036] The beneficial effects of the present invention are:

[0037] (1) The steel reed blade detection device can realize non-contact detection of the reed blade, which can effectively improve the detection speed and accuracy of the steel reed blade.

[0038] (2) The present invention realizes the automatic extraction of the reed blade position by adopting the Yolodent model, and counts the number of reed blades, which has the characteristics of high accuracy and high speed.

[0039] (3) The present invention uses deep learning technology to realize the statistics of the number of steel reed blades, which solves the problems of large errors and low work efficiency in existing reed blade detection, and can meet the dual requirements of speed and accuracy in the detection process of production enterprises.

[0040] (4) The reed blade detection device and detection method can be widely used in the detection of reed defects. Therefore, the disclosure of the present invention has a positive significance for the promotion and application of reed detection technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of a reed blade detection device of the present invention;

[0042] Figure 2 It is a partial schematic diagram of the reed blade detection device of the present invention;

[0043] Figure 3 It is a flow chart of the reed blade detection method based on deep learning of the present invention;

[0044] Figure 4 is a network structure diagram of the Yolodent model in step S4 of the present invention;

[0045] Figure 5 is a SIFT image stitching flow chart of step S5 in the present invention;

[0046] Figure 6 It is the BF feature point matching and splicing effect diagram of step S5 in the present invention;

[0047] Figure 7 This is the effect diagram after the pictures in step S5 of the present invention are stitched together.

[0048] In the figure, 1. workbench, 2. Y-axis linear guide, 3. Type I slider, 4. Type I slider connecting plate, 5. top bracket, 6. driven wheel, 7. driven wheel mounting shaft, 8. synchronous belt, 9. X-axis linear guide, 10. lens, 11. CCD camera, 12. camera adjustment block, 13. camera adjustment frame, 14. Type III slider, 15. Z-axis linear guide, 16. Type II slider connecting plate, 17. Type II slider, 18. synchronous belt clamp, 19. stepper motor, 20. driving wheel, 21. side panel, 22. side panel fixing plate, 23. LED light, 24. bottom bracket. DETAILED DESCRIPTION

[0049] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] Example 1

[0051] The reed blade detection device of the present invention comprises a frame, a driving component, and an image acquisition component. The frame comprises a workbench 1 in contact with the reed, and the workbench 1 is fixedly connected to a bottom bracket 24. The driving component can drive the image acquisition component to move along the X, Y, and Z directions.

[0052] The image acquisition component includes a CCD camera 11, which is used to acquire the surface image of the reed; the image acquisition component is connected to the driving component through a synchronous belt clamp 18;

[0053] It also includes an LED light strip 23 as a light source. The LED light strip 23 is arranged on the workbench 1 and along the length direction of the reed.

[0054] like Figure 1 , Figure 2 As shown, the frame includes a workbench 1, a bottom bracket 24, a Y-axis linear guide rail 2, a type I slider 3, a type I slider connecting plate 4, a top bracket 5 and a side plate 21.

[0055] The frame includes a bottom bracket 24 welded to the bottom surface of the workbench 1, and a pair of Y-axis linear guides 2 are fixedly connected at both ends of the workbench 1 by screws. Two I-type sliders 3 are respectively installed on each Y-axis linear guide 2. The I-type slider 3 is connected to the top bracket 5 through the I-type slider connecting plate 4, and the top bracket 5 is fixedly connected to the side plate 21 through the side plate fixing plate 22.

[0056] The driving component includes a stepper motor 19, which is fixed to the side plate 21 by bolts, and the main shaft of the stepper motor 19 is connected to the driving wheel 20 by a key bar and bolts, and the driving wheel 20 is connected to the driven wheel 6 by a synchronous belt 8, and the driven wheel 6 and the driving wheel 20 are located at the same height; the driven wheel 6 is installed on the driven wheel mounting shaft 7 by bearings and screws, and the driven wheel mounting shaft 7 is fixed in the mounting groove of the vertical surface of the side plate 21 by screws;

[0057] The synchronous belt 8 is fixedly connected to the synchronous belt clamp 18 by bolt clamping, the synchronous belt clamp 18 is connected to the type II slider connecting plate 16 by screws, the type II slider connecting plate 16 is connected to the type II slider 17 by screws, the type II slider 17 is installed on the X-direction linear guide 9, and the X-direction linear guide 9 is fixedly connected to the vertical surface of the side plate 21 by screws;

[0058] The Z-direction linear guide 15 is connected to the II-type slider connecting plate 16 by screws, the III-type slider 14 is installed on the Z-direction linear guide 15, the camera adjustment frame 13 is installed at both ends of the Z-direction linear guide 15 by screws, and the III-type slider 14 is connected to the camera adjustment block 12 by screws.

[0059] The image acquisition component includes a camera adjustment block 12 connected to a type III slider 14, the camera adjustment block 12 is fixedly connected to a CCD camera 11, the CCD camera 11 is connected to a lens 10, the camera adjustment block 12 is connected to the type III slider 14 by bolts, and a camera adjustment frame 13 is installed at both ends of a Z-direction linear guide rail 15 by bolts.

[0060] A tightening hole is provided on the side of the I-type slider 3 , and the tightening hole is used to fix the I-type slider 3 ; a tightening hole is provided on the side of the III-type slider 14 , and the tightening hole is used to fix the III-type slider 14 .

[0061] The side plate 21 is formed by welding two steel plates. A mounting groove is provided on a vertical surface of the side plate 21 . The mounting groove is used to fix the driven wheel mounting shaft 7 . The driven wheel 6 is mounted on the driven wheel mounting shaft 7 .

[0062] The main shaft of the stepping motor 19 is connected to the driving wheel 20 through a key bar and bolts.

[0063] The synchronous belt clamp 18 is provided with a slot and a threaded hole. The synchronous belt 8 is placed in the slot of the synchronous belt clamp 18. The synchronous belt clamp 18 is connected to the II-type slider connecting plate 16 by bolts. The synchronous belt clamp 18 fixes the synchronous belt 8 by clamping the bolts.

[0064] Example 2

[0065] This embodiment provides a method for detecting the number of reed blades based on deep learning, such as Figure 3 As shown, the following steps are included:

[0066] S1, obtaining the original image of the reed blade through the image acquisition component and establishing a reed blade data set;

[0067] The camera model used is CA-HF6400C, the lens model used is CALHW50, and the light source model used is FH-DYDG-01;

[0068] S2, label the reed blade dataset through Labelimage;

[0069] In this embodiment, a total of 1000 reed pieces are selected as the model training data set. The model training set is divided into a data set with a ratio of 7:3, including 700 training sets and 300 verification data sets, and the input image size is 640×640;

[0070] S3, perform Imgaug data enhancement on the annotated reed blade dataset, increase training samples, and improve the model's adaptability to reed blade images in different environments;

[0071] Among them, the Imgaug data enhancement is a professional-level image processing library that supports a variety of image enhancement operations, including geometric transformation (such as rotation, translation, scaling, flipping, etc.), color transformation (such as brightness, contrast, saturation adjustment), blurring and sharpening, noise addition, etc.;

[0072] The Imgaug data enhancement includes reed image preprocessing, and the preprocessing includes five image processing functions: Rotate, FlipLR, FlipUD, Multiply, and AdditiveGaussianNoise.

[0073] Imgaug data augmentation is used to perform data augmentation operations on the reed blade annotated images to expand the reed blade dataset and improve the fitting effect and generalization ability of model training;

[0074] Imgaug data enhancement methods include five image processing effects: random rotation, random horizontal flip, random vertical flip, brightness adjustment, and adding Gaussian noise;

[0075] S4, inputting the data-enhanced reed dataset into the deep learning instance segmentation Yolodent model for training to obtain an optimized Yolodent model;

[0076] S41, loading the reed piece training set data pre-training model weights to achieve transfer learning and reduce the amount of data set training;

[0077] S42, the image data enters the backbone network, and the network layers such as CBS, ELAN, and MP-1 are responsible for extracting the deep features in the image. The convolution kernels of the CBS module are set to 1×1, 3×3, and 3×3, respectively, and the step sizes are set to 1, 1, and 2, respectively. The size of the pooling layer of the MP-1 module is 3×3, generating the feature information of the reed piece;

[0078] S43. The feature map output by the backbone network is sent to the PAFPN network, and multi-scale feature extraction and fusion are completed through the ELAN-W module, UPSample module and SPPCSPC module to enhance the richness and robustness of the reed blade features.

[0079] S44. The feature map after PAFPN fusion is sent to the detection head, and the detection head identifies the type and position of the reed piece.

[0080] S45. Calculate the loss between the predicted result and the actual number of reeds, update the model weights through the back propagation algorithm, and optimize the model performance.

[0081] S46. Use the validation set to evaluate the model’s accuracy, recall and other performance, adjust the hyperparameters, data enhancement strategy or model structure based on the evaluation results, and perform iterative training.

[0082] S47. Save the trained model weights.

[0083] S5, collecting the reed surface image in real time, stitching the real-time reed surface image by SIFT algorithm to obtain a complete reed surface image, and then inputting the complete reed image into the optimized Yolodent model to obtain the position of each reed blade;

[0084] Among them, the specific implementation process of SIFT image stitching is as follows Figure 5 As shown in FIG. 1 , the feature points of the two reed images are extracted, and the feature points of the two images are matched using the BF feature point matching method (as shown in FIG. 1 ). Figure 6 As shown), the homography matrix is ​​calculated by the least square method and other methods. According to the calculated homography matrix, one image is transformed so that it is geometrically aligned with another image, thereby realizing the splicing of the reed image.

[0085] SIFT image stitching Figure 7 As shown, it includes the left image of the reed, the right image of the reed, and the BF feature point matching and stitching results.

[0086] S6. Count the number of reed blades and output the number of reed blades through communication.

[0087] Example 3

[0088] Based on Example 2,

[0089] like Figure 4 As shown, the Yolodent model consists of a backbone network, a feature fusion module and a detection head. The backbone network is used for reed feature extraction and feature output; the feature fusion module adopts a PAFPN network structure to learn and fuse multiple features of the reed; the detection head is used for reed recognition to obtain the category, position and confidence of the reed.

[0090] Among them, the backbone network includes 3 CBS modules, 3 MP-1 modules and 1 ELAN module; the CBS module completes the convolution operation by stacking three times in a row, and gradually extracts the underlying features of the reed image; the MP-1 module completes downsampling by stacking three times in a row to reduce the spatial size of the feature map; the ELAN module enables the network to learn more reed features by controlling the shortest and longest gradient paths.

[0091] The feature fusion module mainly includes ELAN-W module, UPSample module and SPPCSPC module; the ELAN-W module is used to enhance the learning of reed features; the UPSample module is used to implement upsampling operation, improve image resolution and restore image details; the SPPCSPC module completes the extraction and fusion of features of different scales through different pooling layers.

Claims

1. Steel reed blade detection equipment, characterized in that: It comprises a frame, a driving component, and an image acquisition component, wherein the frame comprises a workbench (1) in contact with a reed, and the workbench (1) is fixedly connected to a bottom bracket (24); the driving component can drive the image acquisition component to move along the X, Y, and Z directions; The image acquisition component comprises a CCD camera (11), and the CCD camera (11) is used to acquire a surface image of the reed; the image acquisition component is connected to the driving component via a synchronous belt clamp (18); It also comprises an LED light strip (23) as a light source, wherein the LED light strip (23) is arranged on the workbench (1) along the length direction of the reed.

2. The reed blade detection device according to claim 1, characterized in that: The frame comprises a pair of Y-direction linear guide rails (2) fixedly connected to both sides of the horizontal plane of the workbench (1), the Y-direction linear guide rails (2) are connected to an I-shaped slider (3), the I-shaped slider (3) is connected to a top bracket (5) via an I-shaped slider connecting plate (4), and the top bracket (5) is fixedly connected to a side plate (21); The driving component comprises a stepping motor (19) fixedly connected to a side plate (21), the main shaft of the stepping motor (19) is connected to a driving wheel (20), the driving wheel (20) is connected to a driven wheel (6) via a synchronous belt (8), and the driven wheel (6) and the driving wheel (20) are located at the same height; The synchronous belt clamp (18) is connected to the type II slider connecting plate (16), the type II slider connecting plate (16) is connected to the type II slider (17), the type II slider (17) is installed on the X-direction linear guide rail (9), the type II slider connecting plate (16) is connected to the Z-direction linear guide rail (15), the Z-direction linear guide rail (15) is connected to the type III slider (14), and the type III slider (14) is connected to the camera adjustment frame (13); the camera adjustment frame (13) is installed on both ends of the Z-direction linear guide rail (15) by bolts; The image acquisition component comprises a camera adjustment block (12) connected to a III-type slider (14); the camera adjustment block (12) is fixedly connected to a CCD camera (11); and the CCD camera (11) is connected to a lens (10).

3. The reed blade detection device according to claim 2, characterized in that: The I-type slider (3) has a tightening hole on its side, and the tightening hole is used to fix the I-type slider (3); the III-type slider (14) has a tightening hole on its side, and the tightening hole is used to fix the III-type slider (14).

4. The reed blade detection device according to claim 2, characterized in that: The side plate (21) is welded from two steel plates. A mounting groove is provided on a vertical surface of the side plate (21). The mounting groove is used to fix a driven wheel mounting shaft (7). The driven wheel (6) is mounted on the driven wheel mounting shaft (7).

5. The reed blade detection device according to claim 2, characterized in that: The main shaft of the stepping motor (19) is connected to the driving wheel (20) through a key bar and bolts.

6. The reed blade detection device according to claim 2, characterized in that: The synchronous belt clamp (18) is provided with a slotted hole and a threaded hole, the slotted hole of the synchronous belt clamp (18) is used to place the synchronous belt (8), the synchronous belt clamp (18) is connected to the II-type slider connecting plate (16) by bolts, and the synchronous belt clamp (18) fixes the synchronous belt (8) by clamping the bolts.

7. A method for detecting a reed blade, using the reed blade detection device according to any one of claims 1 to 6, characterized in that: Follow the steps below to implement it: S1. Place the reed on the workbench, drive the image acquisition component to move through the driving component to obtain the surface image of the reed blade, and establish a reed blade data set; S2, label the reed blade dataset using the image labeling tool Labelimage; S3, perform Imgaug data enhancement on the labeled reed blade dataset to increase training samples; S4, inputting the data-enhanced reed dataset into the deep learning instance segmentation Yolodent model for training to obtain an optimized Yolodent model; S5, collecting the reed surface image in real time, stitching the real-time reed surface image by SIFT algorithm to obtain a complete reed surface image, and then inputting the complete reed image into the optimized Yolodent model to obtain the position of each reed blade; S6. Count the number of reed blades and output the number of reed blades.

8. The reed blade detection method according to claim 7, characterized in that: In step S3, the Imgaug data enhancement includes reed image preprocessing, and the preprocessing method adopts any one or more of the five image processing functions of rotation Rotate, horizontal flip FlipLR, vertical flip FlipUD, brightness adjustment Multiply, and adding Gaussian noise Additive Gaussian Noise.

9. The reed blade detection method according to claim 7, characterized in that: The Yolodent model consists of a backbone network, a feature fusion module and a detection head, wherein the backbone network is used for reed feature extraction and feature output; The feature fusion module adopts a PAFPN network structure to learn and fuse multiple features of the reed blade; the detection head is used to identify the reed blade and obtain the category, position and confidence of the reed blade; The backbone network includes 3 CBS modules, 3 MP-1 modules and 1 ELAN module. The CBS module completes the convolution operation by stacking three times in a row, and gradually extracts the underlying features of the reed blade image. The MP-1 module completes downsampling by stacking three times in a row to reduce the spatial size of the feature map. The ELAN module controls the shortest and longest gradient paths to enable the network to learn more reed blade features. The feature fusion module mainly includes ELAN-W module, UPSample module and SPPCSPC module; the ELAN-W module is used to enhance the learning of reed features; the UPSample module is used to implement upsampling operation, improve image resolution and restore image details; the SPPCSPC module completes the extraction and fusion of features of different scales through different pooling layers.

10. The reed blade detection method according to claim 7, characterized in that: Step S4 specifically includes: S41, loading the reed piece training set data pre-training model weights to achieve transfer learning and reduce the amount of data set training; S42, the image data enters the backbone network, and the CBS, ELAN, and MP-1 network layers are responsible for extracting deep features from the image. The convolution kernels of the CBS module are set to 1×1, 3×3, and 3×3, respectively, and the step sizes are set to 1, 1, and 2, respectively. The size of the MP-1 module pooling layer is 3×3, generating feature information of the reed blade; S43, the feature map output by the backbone network is sent to the PAFPN network, and multi-scale feature extraction and fusion are completed through the ELAN-W module, UPSample module and SPPCSPC module to enhance the richness and robustness of the reed blade features; S44, the feature map after PAFPN fusion is sent to the detection head, and the detection head identifies the type and position of the reed piece; S45, calculating the loss between the predicted result and the actual number of reeds, updating the model weights through the back propagation algorithm, and optimizing the model performance; S46. Use the validation set to evaluate the model accuracy and recall, adjust the hyperparameters, data enhancement strategy or model structure according to the evaluation results, and perform iterative training; S47. Save the trained model weights.

Images