Special paper fiber identification method based on deep learning

Abstract

The invention discloses a deep learning-based special paper fiber identification method. The method comprises the following steps of shooting special paper through a microscopic camera to obtain paper surface fiber image data; marking a rotating target frame, and establishing a special paper fiber data set; operation such as brightness adjustment, contrast adjustment, rotation angle adjustment, vertical and horizontal overturning, translation and shearing is carried out on a training set, so that the generalization ability of the model is improved, and overfitting is reduced; a two-stage rotating target detection model is constructed based on a Faster-RCNN network, the two-stage rotating target detection model is structurally adjusted and optimized, and an RPN part is improved through a midpoint offset representation method; training parameters are set, and a network model with the best fiber target detection effect is obtained through a large amount of training and verification; calling the network model, checking the test data, and verifying the model performance; and based on a detection result, statistically analyzing fiber distribution and characteristics.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a special paper fiber recognition method based on deep learning. Background Art

[0002] With the development of my country's economy and society and the advancement of industrialization, the use of paper in production and life is increasing. The scope of use covers many fields such as culture, education, science and technology, national economy, industrial and agricultural production, and specialty paper is one of the most important categories. At present, specialty paper has become an indispensable product in modern production and life. During the production process, it usually needs to undergo a large number of quality tests to ensure that the produced paper can meet the expected requirements. The quality problems of specialty paper will directly affect the qualification rate of the product, and specialty paper is also widely used in various production industries, which will have a great impact on the quality of the final product. The quality inspection of specialty paper by naked eyes is not only subjective, but also inefficient; and the efficiency of destructive testing by traditional mechanical devices is also low.

[0003] In the study of the characteristics of specialty paper, the fiber characteristics of paper were found to be an important branch of specialty paper quality monitoring. The morphology and overall distribution of paper fibers determine the strength and other characteristics of the produced specialty paper. The quality of specialty paper can be judged by detecting and identifying the fibers of specialty paper.

[0004] Gunnar et al. summarized the application of ultrasonic measurement technology on paper fibers (Application of ultrasonic measurement technology in paper and paperboard production). However, this method has the disadvantages of being too costly and not being able to intuitively reflect the fiber details. There are also studies on paper fiber detection in the field of images. Ma Lei (Research on the measurement of pulp fiber morphological parameters based on machine vision) measured the length and width of fibers through refinement algorithms, morphological processing and other methods. However, this type of traditional algorithm relies on data preprocessing and consistency, and performs poorly in generalization performance.

[0005] Therefore, it is necessary to develop a special paper fiber identification method with low cost, good applicability and high recognition accuracy to address the shortcomings of the above methods. Summary of the invention

[0006] In view of the above-mentioned deficiencies, the present invention proposes a special paper fiber identification method based on deep learning, which can realize the intelligent identification of special paper fibers through the image of the special paper surface during the quality inspection after the production of special paper, and then perform statistical analysis on the various characteristics of the fibers. It has the advantages of high accuracy and high efficiency, and expands the application of deep learning technology in fiber detection scenarios.

[0007] The present invention is achieved by at least one of the following technical solutions.

[0008] A special paper fiber identification method based on deep learning includes the following steps:

[0009] The image of the surface fiber of the special paper is taken through a microscope and a camera, the original image data is sampled, and the image is cut to obtain the original fiber image data;

[0010] In the original fiber image data, a rotating directed bounding box is used to mark the shape, position and category information of the fiber objects in the image;

[0011] Based on the original fiber image data and the annotation information of the rotated target frame, a special paper fiber target detection dataset is constructed and divided into a training set, a validation set, and a test set.

[0012] Perform data augmentation on images in the training set of the special paper fiber target detection dataset;

[0013] Construct a network model for detecting special paper fiber targets, and adjust and optimize the model structure based on the rotation characteristics of the target;

[0014] Set training parameters and configurations, use training sets and test sets to train the model, obtain the best model parameters after multiple rounds of training, call the network model, and test the test image data;

[0015] The model test results are analyzed to obtain the overall distribution and characteristic data of special paper fibers.

[0016] Furthermore, the surface fiber image data of special paper was sampled by the National Papermaking Laboratory using a low-power microscope and a professional camera to shoot the produced special paper samples. Paper fibers have tiny characteristics, so a small number of special paper image samples have a large number of fiber target objects to be detected, and the adequacy of the data is guaranteed. In addition, since the resolution of the acquired image data is too high, before constructing the special paper fiber detection data set, it is necessary to divide the large image into small images in the data set by cutting and scaling. When cutting, a certain overlapping area is retained to enhance the data and retain the integrity of the target, thereby improving the accuracy of the trained model.

[0017] Furthermore, the original image data is sampled, including:

[0018] After acquiring a high-resolution image, it is cropped to obtain an image with a resolution suitable for the model. First, the size of the cropping window is determined, and then the size of the overlapping area to be retained when the cropping window moves is determined based on this. Finally, the cropping window is cyclically slid on the image using a specified step size.

[0019] Furthermore, after obtaining appropriate image data, it is necessary to annotate all fiber objects in the image, that is, mark the location and type of the fiber. In order to more accurately represent the fiber objects, a directed target box is used for marking during image annotation. The fiber objects in the image are marked with shape, location information and category information using a rotating directed bounding box, including:

[0020] OBB={x c ,y c ,w,h,a}

[0021] where x c ,y c are the horizontal and vertical coordinates of the center point of the target frame, w and h are the width and height of the target frame, and a is the rotation angle of the frame relative to the horizontal axis. A directed target frame accurately represents the actual boundary of the target by describing the shape and direction of the object.

[0022] Furthermore, the cleaned and labeled data set is divided into training set, validation set, and test set in a ratio of 8:1:1. The simplest random division method is used in the division process, which is suitable for the current situation where the data is uniform.

[0023] Furthermore, the training set is enhanced by:

[0024] Data enhancement is performed on each image in the training set, including adjusting brightness, adding Gaussian noise, adjusting contrast, rotating angle, vertical and horizontal flipping, translation and shearing, etc., to artificially increase the diversity of training data, thereby improving the generalization ability of the training model, reducing overfitting, and improving the performance of the model on unseen data.

[0025] Furthermore, the special paper fiber target detection network model is a two-stage target detection model based on Faster-RCNN, on this basis, angle regression parameters are introduced into both stages, so that the model can detect rotated angled target frames; the backbone network of the network model adopts the residual network ResNet50 structure, and the obtained feature map adopts the feature pyramid network FPN for feature fusion.

[0026] The special paper fiber target detection network model includes:

[0027] Furthermore, the one-stage RPN network realizes the improvement of the representation of the rotated target and the network structure. When regressing the rotated target frame for the first time, the angle is not directly regressed but the rotated target frame represented by the midpoint offset representation containing six parameters:

[0028] δx=(xx a ) / x a ,δy=(yy a ) / ya

[0029] δw=log(w / w a ),δh=log(h / h a )

[0030] δα=(x'-x) / w, δβ=(y'-y) / h

[0031] Where δ is the mapping that the network needs to learn, x, y, w, h are the horizontal and vertical coordinates of the center point of the rotated external frame output by the network, w and h are the width and height of the target frame respectively, and x a ,y a ,w a ,h a is the horizontal and vertical coordinates and width and height of the corresponding anchor box, α and β are the midpoint offsets of the target box relative to its external box, which are obtained by x', y', that is, the horizontal and vertical coordinates of the vertices of the rotation box; this method of indirectly representing the rotated target box can speed up training convergence and improve accuracy, corresponding to the six-parameter midpoint offset method, and at the same time make certain improvements and optimizations to the regression structure of the detection model.

[0032] The midpoint offset method constructs an external bounding box of a directed target box, and uses the offset of the external bounding box midpoint and the directed box vertex to encode as the target of neural network training. The network prediction value is decoded in the same encoding method to obtain a directed bounding box. In this way, a loose target box is regressed, so that the model can obtain a higher degree of matching when training the target, thereby enhancing the training effect and improving the model accuracy.

[0033] Furthermore, in the special paper fiber target detection network model, the loss functions used in the two stages are cross entropy loss and smooth L1 loss respectively:

[0034]

[0035] in is the cross entropy loss, To smooth the L1 loss, is the predicted vector, Y is the actual vector, y k , is the kth predicted value and actual value, K is the number of predicted values.

[0036] Furthermore, the training parameters and configurations are set before training, including setting the epoch, batch size, learning rate and other training parameters, and the Adam optimizer is used to perform a large number of multiple rounds of training and test verification to obtain the network model with the best fiber detection effect, the model's parameter file and the model's inference results on the verification set, and the validity of the model is judged based on the verification results.

[0037] Furthermore, the network model is called to test the test set, and the average precision AP and recall Recall are used as the model evaluation criteria to verify the model performance. Build the detection model and load the parameter file, conduct detection experiments on the test data, and use the average precision (Average Precision) and recall (Recall) as the evaluation criteria:

[0038]

[0039] TP is a correctly identified positive sample, FP is a wrongly identified positive sample, and FN is a non-identified positive sample. The average accuracy is calculated by averaging the accuracy at different recall levels, which comprehensively evaluates the accuracy of the target detection model. The average accuracy and recall are calculated through the test results to effectively verify the model performance.

[0040] Furthermore, after obtaining the fiber detection results, the number of detected fibers is counted and the density of the fibers in the paper is calculated; the angles of the detected fibers are counted to obtain the overall fiber orientation of the paper; the detected fiber images are analyzed to obtain the length and thickness of the fibers, and finally the quality of the specialty paper is analyzed based on this.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. The present invention proposes a method for detecting special paper fibers through a target detection network model based on deep learning, and uses a convolutional neural network and a target detection algorithm for paper fiber detection. Compared with traditional image algorithms, the method has strong applicability, high accuracy, extremely strong reliability and stability, simple operation, and is easy to use, thereby improving production efficiency.

[0043] 2. The present invention proposes to use oriented target detection for fiber target detection. For fibers with arbitrary orientation and slender characteristics, a rotating target frame is used to represent the fiber object, and the corresponding network model is modified to enable it to learn the angle parameters of the target, so that the model can obtain more accurate detection results.

[0044] 3. The present invention uses the network regressor representation method of the midpoint offset method in the RPN network, and indirectly represents the angle parameter as the midpoint offset, thereby obtaining a better training effect and improving the model accuracy.

[0045] 4. After obtaining the test results, the model of the present invention automatically analyzes the fiber characteristics of the paper, realizes end-to-end special paper fiber detection and identification, and provides great convenience in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the following briefly introduces the drawings required for use in the specific implementation or the description of the prior art.

[0047] Figure 1 The figure is a flow chart of a method for identifying special paper fibers based on deep learning in an embodiment;

[0048] Figure 2 This is a diagram for defining the angle of the rotating target frame in the embodiment;

[0049] Figure 3 is a structural diagram of a fiber target detection model constructed in the embodiment;

[0050] Figure 4 This is a structural diagram of the improved OrientedRPN in the embodiment;

[0051] Figure 5 This is a result diagram of the target fiber detected in the embodiment;

[0052] Figure 6 This is the target fiber angle distribution diagram obtained by analysis in the example. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical method and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific examples. These examples are merely illustrative and not limiting of the present invention.

[0054] A method for identifying special paper fibers based on deep learning in this embodiment first uses microscopic imaging to collect high-definition image data of special paper surface fibers, and then cuts them to a suitable data set size. Then, each fiber object in the processed image set is marked with a target frame and assigned a classification label to construct a fiber detection data set. The target frame uses a rotated directed frame to lay the foundation for model training. After that, the data set is divided, and the training set is enhanced with data such as rotation and flipping to improve the adaptability of the model. Then, a deep learning fiber target detection network model is constructed, combining CNN, Faster-RCNN and rotation target detection, and optimizing the model structure to improve detection accuracy. After setting the training parameters, the model is trained using the training set and the validation set to obtain the optimal parameters. After the training is completed, the model is used to detect and identify fibers on the test set. Finally, the test results are analyzed, and the fiber distribution and characteristic data are statistically analyzed to provide a basis for paper quality control.

[0055] like Figure 1 As shown, a special paper fiber identification method based on deep learning in this embodiment specifically includes the following steps:

[0056] S1. Acquire suitable surface fiber image data of specialty paper;

[0057] Through the standard process of the National Papermaking Laboratory, low-power microscopes and professional cameras were used to photograph and sample the produced specialty paper samples. The tiny characteristics of the paper fibers in these samples make even a small number of image samples have a large number of fiber target objects to be detected, thus ensuring the adequacy of the data. Secondly, due to the high resolution of the acquired image data, before constructing the special paper fiber detection data set, it is necessary to divide the large image into small images in the data set by cutting and scaling. Specifically, after the high-resolution image is acquired, it is cropped to obtain an image with a resolution suitable for the model. First, the size of the cropping window is determined, and then the size of the overlapping area to be retained when the cropping window moves is determined accordingly. Finally, the cropping window is cyclically slid on the image using the specified step size on the entire image. Appropriate overlapping areas are deliberately retained during the cutting process, which not only enhances the diversity of the data and the generalization ability of the model, but also ensures the integrity of the fiber target, thereby helping to improve the accuracy of the trained model.

[0058] S2. Data Labeling

[0059] In this step, a visual image annotation software such as roLabelImg is used to create a directed bounding box and adjust the position, width, height and angle of the box to mark the shape, position and category information of the fiber object. The encoding method of the directed box adopts the le90 defined in OpenCV, that is, the long side of the rectangular box is w, the narrow side is h, the angle is the angle between the long side and the x-axis, and the clockwise rotation is defined as a positive angle, such as Figure 2 The marked result is saved as a text file, and the marking information is established in the XML extensible markup language format.

[0060] Based on the characteristic that the fiber target has an arbitrary direction, this embodiment adopts a rotating target frame including a fifth parameter angle to surround the fiber object, which is different from the traditional target inspection:

[0061] OBB={x c ,y c ,w,h,a}

[0062] where x c ,y c are the horizontal and vertical coordinates of the center point of the target frame, w and h are the width and height of the target frame, and a is the rotation angle of the frame relative to the horizontal axis.

[0063] S3. Dataset establishment: Based on the original image data and the annotation information of the rotated target box, a special paper fiber target detection dataset is constructed, in which a certain proportion of training set, validation set, and test set are divided.

[0064] After collecting and annotating the image dataset, it is necessary to establish a special paper fiber dataset for rotation target detection. Therefore, a fiber target detection dataset consisting of oriented target boxes is established based on the DOTA dataset format.

[0065] As an example, the dataset consists of nearly 100 different 1024*1024 pixel images of specialty paper surface fibers, each containing 40-90 fiber objects. Due to the scarcity of image data, the data was expanded and enhanced by changing brightness and contrast, adding Gaussian noise, rotating the image, and mirroring. The training set finally contains 289 images and 13,365 instances, and the validation test set contains 618 instances.

[0066] Data enhancement was performed on the images in the training set of the special paper fiber dataset to improve the generalization ability of the model and reduce overfitting. In order to increase the data volume and adaptability, data enhancement was performed on each image in the training set by adjusting brightness, adding Gaussian noise, adjusting contrast, rotating angle, vertical and horizontal flipping, translation and shearing, etc.

[0067] S4. Constructing a fiber target detection network model

[0068] In this step, a special paper fiber target detection network model is constructed. The model is modified based on Faster-RCNN. Angle regression parameters are introduced for both stages based on Faster-RCNN, so that the model can detect the rotated angled target frame. The backbone network of the network adopts the residual network ResNet50 structure, and the obtained feature map uses the feature pyramid network FPN for feature fusion.

[0069] The constructed model is as follows Figure 3 As shown. In the model, the backbone network is used to analyze image data and extract feature maps of interest. The ResNet50 model structure of the residual network is adopted. The residual learning framework introduced solves the degradation problem encountered in deep network training, and greatly deepens the network depth, improving the network training speed and detection accuracy. The feature pyramid network FPN is used in the neck network part. By using upsampling and lateral connections to extract and fuse the multi-scale features of the original image, the model can process feature maps of different scales at the same time, thereby detecting targets of different sizes. After obtaining the multi-scale feature map, only by inputting it into the head network can the corresponding output detection result be obtained according to the feature. The improved structure of the RPN in one stage is shown below. Figure 4As shown in the figure, the network outputs the feature vector of each feature point in the form of full convolution. Specifically, it is a 3*3 convolution followed by a 1*1 convolution, which realizes spatial feature sharing and information transfer between channels, and outputs a 6-dimensional vector for each point on the feature map:

[0070] X={Δx,Δy,Δw,Δh,Δα,Δβ}

[0071] The six parameters represent the output detection results encoded by the model using the midpoint shift method. The calculation formula of the midpoint shift method is:

[0072] δx=(xx a ) / x a ,δy=(yy a ) / y a

[0073] δw=log(w / w a ),δh=log(h / h a )

[0074] δα=(x'-x) / w, δβ=(y'-y) / h

[0075] Where δ is the mapping that the network needs to learn, x, y are the horizontal and vertical coordinates of the center point of the rotational bounding box output by the network, and x a ,y a ,w a ,h a are the horizontal and vertical coordinates and width and height of the corresponding anchor box, α and β are the midpoint offsets of the target box relative to its external frame, which are obtained by x', y', the horizontal and vertical coordinates of the vertices of the rotation box. The model constructs a directed target box inscribed in the rectangular box through the two midpoint offsets α and β. In addition, the angle parameter is also added to the two-stage RCNN network. The difference is that the network obtains regression and classification results in a fully connected form, and it can access all input features, so it can learn very complex nonlinear mappings, which is more suitable for two-stage refinement detection.

[0076] In addition, the classification loss used in the constructed model adopts the cross entropy loss function, and the regression loss adopts the Smooth L1 loss function:

[0077]

[0078] in is the cross entropy loss, To smooth the L1 loss, is the predicted vector, Y is the actual vector, y k , is the k-th predicted value and the actual value. The cross entropy loss function is a commonly used loss function, especially in classification problems. It measures the difference between the true label and the model prediction. The SmoothL1 loss function is a smoothed version of the L1 loss function, which is a combination of MAE and MSE. It becomes smooth when the gradient of the L1 loss function becomes very large (close to zero or far from zero), thereby reducing the instability of the model in the edge area.

[0079] S5. Set training parameters and train to obtain model parameters

[0080] In order to speed up training and achieve better training results, the backbone network ResNet50 in the constructed model is initialized with the weights pre-trained on the ImageNet dataset, and then fine-tuned on the fiber dataset. Pre-training and fine-tuning are two common training methods in deep learning. Their purpose is to first let the model learn general feature representations on a large-scale dataset, and then further optimize the model on a dataset of a specific task to obtain better performance. In addition, the parameters of model training are set to 60 epochs, 4 batch sizes, and 0.0001 based on the actual data. The periodic fixed decay learning rate strategy is adopted, which can effectively prevent overfitting and usually accelerate the training process. Adam is selected as the optimizer, which uses momentum to accelerate the learning process and can adaptively adjust the learning rate, accelerating the training of the model. After multiple rounds of training, the parameter file of the model is obtained.

[0081] S6. Call the network model to test the test data

[0082] In this step, after the model structure loads the parameter file, it detects the input test image and draws the detection results on the image, such as Figure 5 As shown in the figure, the evaluation indicators are calculated. We get an AP of 71.34 and a Recall of 76.93, which meet the requirements.

[0083] As an example, the average precision and recall are used as evaluation criteria:

[0084]

[0085] TP is a correctly identified positive sample, FP is a wrongly identified positive sample, and FN is a non-identified positive sample. The average accuracy is calculated by averaging the accuracy at different recall levels, which comprehensively evaluates the accuracy of the target detection model. The average accuracy and recall are calculated through the test results to effectively verify the model performance.

[0086] S7. Analyze and collect data on fiber distribution and characteristics of specialty paper

[0087] After obtaining the fiber detection results, the number of detected fibers is counted to calculate the density of the fibers in the paper; the angle of the detected fibers is counted to obtain the fiber orientation of the paper as a whole, such as Figure 6 ; The detected fiber images can be analyzed to determine the length and thickness of the fibers, which can then be used to analyze the quality of the specialty paper.

[0088] The present invention can combine the output results obtained by the recognition method with other quality monitoring of special paper, such as surface defect detection, to provide an expansion means and a more efficient method for fully automatic intelligent quality detection, which can effectively improve efficiency and reduce unnecessary consumption of manpower and material resources, and effectively alleviate the difficulty of detailed inspection when the production volume of special paper is large. The detection method based on convolutional neural network can also effectively expand the actual application scenarios, such as using this method to distinguish multiple fibers.

[0089] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well.

Claims

1. A special paper fiber identification method based on deep learning, characterized in that: The following steps are involved: The image of the surface fiber of the special paper is taken through a microscope and a camera, the original image data is sampled, and the image is cut to obtain the original fiber image data; In the original fiber image data, a rotating directed bounding box is used to mark the shape, position and category information of the fiber objects in the image; Based on the original fiber image data and the annotation information of the rotated target frame, a special paper fiber target detection dataset is constructed and divided into a training set, a validation set, and a test set. Perform data augmentation on images in the training set of the special paper fiber target detection dataset; Construct a network model for detecting special paper fiber targets, and adjust and optimize the model structure based on the rotation characteristics of the target; Set training parameters and configurations, use training sets and test sets to train the model, obtain the best model parameters after multiple rounds of training, call the network model, and test the test image data; The model test results are analyzed to obtain the overall distribution and characteristic data of special paper fibers.

2. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: Sampling the original image data includes: After acquiring a high-resolution image, it is cropped to obtain an image with a resolution suitable for the model. First, the size of the cropping window is determined, and then the size of the overlapping area to be retained when the cropping window moves is determined based on this. Finally, the cropping window is cyclically slid on the image using a specified step size.

3. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: The fiber objects in the image are marked with shape, position and category information using a rotating directed bounding box, including: Based on the characteristic that the fiber target has an arbitrary direction, a rotating target frame with five parameter angles is used to surround the fiber object, which is different from the traditional target inspection: OBB={x c ,y c ,w,h,a} where x c ,y c are the horizontal and vertical coordinates of the center point of the target frame, w and h are the width and height of the target frame, and a is the rotation angle of the frame relative to the horizontal axis.

4. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: Perform data augmentation on the training set, including: Data enhancement is performed on each image in the training set, including adjusting brightness, adding Gaussian noise, adjusting contrast, rotating angle, vertical and horizontal flipping, translation, and shearing.

5. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: The special paper fiber target detection network model is a two-stage target detection model based on Faster-RCNN, on this basis, angle regression parameters are introduced into both stages, so that the model can detect the rotated angled target frame; the backbone network of the network model adopts the residual network ResNet50 structure, and the obtained feature map adopts the feature pyramid network FPN for feature fusion.

6. The method for identifying special paper fibers based on deep learning according to claim 5, characterized in that: The special paper fiber target detection network model includes: The one-stage RPN network realizes the improvement of the representation of the rotated target and the network structure. When regressing the rotated target frame for the first time, the rotated target frame represented by the midpoint offset representation containing six parameters is: δx=(x-x a ) / x a ,δy=(y-y a ) / y a δw=log(w / w a ),δh=log(h / h a ) δα=(x'-x) / w, δβ=(y'-y) / h Where δ is the mapping that the network needs to learn, x, y, w, h are the horizontal and vertical coordinates of the center point of the rotated external frame output by the network, w and h are the width and height of the target frame respectively, and x a ,y a ,w a ,h a are the horizontal and vertical coordinates and width and height of the corresponding anchor box, α and β are the midpoint offsets of the target box relative to its external box, which are obtained by x', y', that is, the horizontal and vertical coordinates of the vertices of the rotation box.

7. The method for identifying special paper fibers based on deep learning according to claim 5, characterized in that: In the special paper fiber target detection network model, the loss functions used in the two stages are cross entropy loss and smooth L1 loss: in is the cross entropy loss, To smooth the L1 loss, is the predicted vector, Y is the actual vector, y k , is the kth predicted value and actual value, K is the number of predicted values.

8. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: Setting training parameters and configurations includes setting the training parameters of epoch, batch size, and learning rate. Using the Adam optimizer, after a lot of training and verification, the network model with the best fiber detection effect is obtained.

9. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: Call the network model, test the test set, use the average accuracy AP and recall Recall as the model evaluation criteria to verify the model performance.

10. The method for identifying special paper fibers based on deep learning according to claim 1, characterized in that: The analysis obtains the overall distribution and property data of specialty paper fibers, including: After obtaining the fiber detection results, the number of detected fibers is counted and the density of the fibers in the paper is calculated; the angles of the detected fibers are counted to obtain the overall fiber orientation of the paper; the detected fiber images are analyzed to obtain the length and thickness of the fibers, and finally the quality of the specialty paper is analyzed based on this.

Images