Cashmere fiber length measurement method and system based on convolutional neural network, and medium

Abstract

The invention discloses a cashmere fiber length measurement method and system based on a convolutional neural network, and a medium, and belongs to the technical field of cashmere fiber detection, and the cashmere fiber length measurement system based on the convolutional neural network comprises an obtaining and processing module which obtains and processes an actual to-be-detected cashmere fiber sample image; the calculation module is used for calculating the average fiber length of the actual cashmere fiber sample to be detected by using the fiber fracture repair model and the fiber cross detection model; the display module is used for displaying, outputting and calculating the average fiber length of the actual to-be-detected cashmere fiber sample. The method can detect the average length of cashmere fibers.

Description

Technical Field

[0001] This invention application belongs to the technical field of cashmere fiber detection, and specifically relates to a method, system, and medium for measuring the length of cashmere fibers based on a convolutional neural network. Background Art

[0002] The length of cashmere fibers is an important indicator for evaluating the quality of cashmere, directly affecting the performance and price of cashmere products. Traditional measurement methods mainly use the hand-laying method, which relies on manual operation and has problems such as low efficiency and being easily affected by subjective factors. Although there are already some automated measurement instruments, they are expensive and consume a large amount of test samples. The measurement method based on image processing has received extensive attention due to its high efficiency and low cost, but there are still deficiencies in the process of fiber recognition and segmentation. Especially after the preprocessing of fiber images, due to the certain curl of cashmere fibers themselves, the fibers reflect light differently during the shooting process, resulting in different brightnesses of fibers in the image. After preprocessing, it appears as fiber breaks in the image, and the processing accuracy of the break and intersection areas is relatively low. Therefore, a new method for measuring the length of cashmere fibers based on a convolutional neural network is needed to overcome the deficiencies of the existing technology and achieve fast and accurate measurement. Summary of the Invention

[0003] This invention application provides a method, system, and medium for measuring the length of cashmere fibers based on a convolutional neural network, aiming to partially or fully solve the problems in the existing technology such as the low efficiency of traditional manual measurement methods and being easily affected by subjective factors, and overcome the technical problems of the existing automated measurement instruments being expensive and consuming a large amount of test samples. This invention application can use deep learning technology to automatically analyze fiber images, can effectively identify fiber break and intersection areas, restore fiber integrity, and accurately calculate fiber length. To achieve the above objectives, this invention application adopts the following technical solutions:

[0004] In the first aspect, a method for measuring the length of cashmere fibers based on a convolutional neural network includes:

[0005] Step S100: Obtain multiple groups of fiber image sample data sets, and each group of fiber image sample data sets includes: fiber sample images, fiber sample binary images corresponding to the fiber sample images, and fiber sample repair images corresponding to the fiber sample binary images;

[0006] Step S200: Based on the U-net model, train and construct a fiber break repair model according to multiple groups of fiber image sample data sets;

[0007] Step S300: Obtain multiple groups of fiber sample cross-detection image sets, and each group of fiber image sample data sets includes: fiber sample repair images, fiber sample cross images corresponding to the fiber sample repair images, and construct a fiber cross-detection model based on the ResNet-50 convolutional neural network;

[0008] Step S400: Obtain an image of the actual cashmere fiber sample to be detected, process the image of the actual cashmere fiber sample to be detected, and use the fiber breakage repair model and the fiber crossing detection model to calculate the average fiber length of the actual cashmere fiber sample to be detected.

[0009] Optionally, step S100 includes:

[0010] Step S101: Make a cashmere fiber sample and obtain multiple groups of fiber sample images;

[0011] Step S102: Perform image processing on each group of fiber sample images to obtain the binary image of each group of fiber samples, perform processing on the binary image of each group of fiber samples to obtain the repaired image of each group of fiber samples, and then obtain multiple groups of binary images of fiber samples and multiple groups of repaired images of fiber samples;

[0012] Step S103: Correlate each group of fiber sample images, the binary image of each group of fiber samples, and the repaired image of each group of fiber samples to form a data set of fiber image samples for each group, and then obtain multiple data sets of fiber image samples.

[0013] Optionally, step S101 includes:

[0014] Making a cashmere fiber sample: First, prepare two pieces of black cardboard that are 10 cm long and 1 cm wide. Stick black double-sided tape on one of them. Take about 0.3 g of cashmere sample, pinch one end by hand, and comb the other end with a steel comb until the fibers are straight. Paste the straightened part on the double-sided tape; repeat the above steps until the double-sided tape on the cardboard is covered with fibers; then, cover the other piece of cardboard on the double-sided tape and press it firmly to fix the fibers between the two pieces of cardboard; finally, comb the sample neatly on the velvet board, ensure that there is no obvious entanglement at both ends, and then use a camera to take pictures of the cashmere fiber sample multiple times to obtain fiber sample images, and then obtain multiple groups of fiber sample images

[0015] Optionally, step S102 includes:

[0016] Step S1021: Grayscale each group of fiber sample images, that is, convert the pixel points originally represented by three values of red, green, and blue into a single grayscale value, and then perform binary processing. Convert the pixels with a grayscale value higher than the set threshold into white pixels, representing the fiber body, and convert the pixels lower than the threshold into black pixels, representing the background, so as to obtain the binary image of each group of fiber samples;

[0017] Step S1022: Crop each group of fiber sample images and each group of binary images of fiber samples. After cropping, the size of both is 256×256. The insufficient parts are filled with black pixels. Using a painting software, place each group of fiber sample images at the bottom layer and each group of binary images of fiber samples at the upper layer. Use a white pixel pen to depict and connect all the fiber parts that are disconnected due to the binarization process to obtain fiber sample repaired images.

[0018] Optionally, step S300 includes:

[0019] Step S301: Annotate each group of fiber sample repaired images. Use an annotation software to mark the fiber crossing parts in the fiber sample repaired images with polygons to obtain each group of fiber sample crossing images, and then obtain multiple groups of fiber sample crossing images;

[0020] Step S302: Correlate each group of fiber sample repaired images and each group of fiber sample crossing images correspondingly to form each group of fiber sample crossing detection image sets, and then obtain multiple groups of fiber sample crossing detection image sets.

[0021] Step S303: Based on the ResNet-50 convolutional neural network, construct a fiber crossing detection model.

[0022] Optionally, step S400 includes:

[0023] Step S401: Make actual cashmere fiber samples and obtain images of actual cashmere fiber samples to be detected;

[0024] Step S402: Perform image processing on the images of actual cashmere fiber samples to be detected to obtain binary images of actual cashmere fiber samples to be detected;

[0025] Step S403: According to the images of actual cashmere fiber samples to be detected and the binary images of actual cashmere fiber samples to be detected, use the fiber breakage repair model to obtain repaired images of actual cashmere fiber samples to be detected. Calculate the total number of fiber pixels FP and the total number of fiber roots F in the repaired images of actual cashmere fiber samples to be detected; According to the repaired images of actual cashmere fiber samples to be detected, use the fiber crossing detection model to obtain fiber crossing part images, and calculate the number of fiber pixels CP in the fiber crossing parts. The average fiber length FL is:

[0026]

[0027] Where d is the fiber width value, FP is the total number of fiber pixels in the repaired images of actual cashmere fiber samples to be detected, CP is the number of fiber pixels in the fiber crossing parts, F is the total number of fiber roots in the repaired images of actual cashmere fiber samples to be detected, w is the width of the cardboard, and μ is the ratio of the number of pixels on the velvet board to the actual length of the velvet board, unit: number of pixels / mm.

[0028] In a second aspect, a cashmere fiber length measurement system based on a convolutional neural network is used to implement any one of the above-mentioned cashmere fiber length measurement methods based on a convolutional neural network, and includes:

[0029] An acquisition and processing module that acquires and processes an actual cashmere fiber sample image to be detected;

[0030] A calculation module that uses a fiber break repair model and a fiber crossing detection model to calculate the average fiber length of the actual cashmere fiber sample to be detected;

[0031] A display module that displays and outputs the average fiber length of the actual cashmere fiber sample to be detected calculated.

[0032] Optionally, the acquisition and processing module includes: an actual image acquisition unit and an actual image processing unit. The actual image acquisition unit acquires an actual cashmere fiber sample image to be detected, and the actual image processing unit performs image processing on the actual cashmere fiber sample image to obtain a binary image of the actual cashmere fiber sample to be detected.

[0033] Optionally, the calculation module includes: a fiber break repair model unit, a fiber crossing detection model unit, and a length calculation unit;

[0034] The fiber break repair model unit obtains a repaired image of the actual cashmere fiber sample to be detected according to the actual cashmere fiber sample image to be detected and the binary image of the actual cashmere fiber sample to be detected;

[0035] The fiber crossing detection model unit obtains a fiber crossing part image according to the repaired image of the actual cashmere fiber sample to be detected;

[0036] The length calculation unit calculates the total number of fiber pixels FP and the total number of fiber roots F in the repaired image of the actual cashmere fiber sample to be detected, and the number of fiber pixels CP in the fiber crossing part in the fiber crossing part image, and calculates the average fiber length FL as:

[0037]

[0038] where d is the fiber width value, FP is the total number of fiber pixels in the repaired image of the actual cashmere fiber sample to be detected, CP is the number of fiber pixels in the fiber crossing part, F is the total number of fiber roots in the repaired image of the actual cashmere fiber sample to be detected, w is the cardboard width, and μ is the ratio of the number of pixels of the velvet board to the actual length of the velvet board, unit: number of pixels / mm.

[0039] In a third aspect, a computer-readable storage medium has instructions stored thereon, which, when run on a computer, execute any one of the methods for measuring the length of cashmere fibers based on a convolutional neural network in the first aspect above.

[0040] In a fourth aspect, a device for measuring the length of cashmere fibers based on a convolutional neural network includes a memory and a processor connected by communication. The memory is used to store a computer program, and the processor is used to read the computer program and execute any one of the methods for measuring the length of cashmere fibers based on a convolutional neural network in the first aspect.

[0041] (1) In the present invention application, by collecting high-resolution cashmere fiber images, it is ensured that the morphology and detailed features of the fibers can be accurately captured, and diverse cashmere fiber image data is collected. Image processing operations help improve the image quality, enabling the fiber break repair model to more effectively extract fiber features, learn the features of different types of cashmere fibers, and improve the generalization ability of the fiber break repair model. By repairing all possible breakage situations through breakage repair, it is ensured that the breakage repair model can identify and repair all broken fibers.

[0042] (2) In the present invention application, through a fiber break repair model based on U-Net and a fiber crossing detection model based on ResNet-50, the feature extraction ability of deep learning is fully utilized. Through training and optimization, the fiber break repair model and the fiber crossing detection model can learn the features of the fibers in the image and verify their performance on the test set, improving the accuracy and reliability of the fiber break repair model, being able to more effectively process complex fiber image data, improving the detection accuracy of the fiber break repair model, reducing the workload of manually measuring the length of cashmere fibers, reducing labor costs, and at the same time improving the objectivity and consistency of cashmere fiber length detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0044] Figure 1 It is a schematic flowchart of the method for measuring the length of cashmere fibers based on a convolutional neural network of the present invention application;

[0045] Figure 2 It is a schematic diagram of the production of cashmere fiber samples of the present invention application.

[0046] Figure 3Schematic diagram constructed for training the fiber fracture repair model of this invention application;

[0047] Figure 4 Schematic diagram constructed for training the fiber crossing detection model of this invention application;

[0048] Figure 5 Schematic diagram of the principle for obtaining the repaired image of the fiber sample in this invention application;

[0049] Figure 6 Schematic diagram of the image of the fiber crossing part in this invention application;

[0050] Figure 7 Schematic diagram of the composition of the cashmere fiber length measurement system based on convolutional neural network in this invention application.

[0051] The attached drawings are used to provide further understanding of this invention application, and constitute a part of the specification. Together with the embodiments of this invention application, they are used to explain this invention application, and do not constitute a limitation to this invention application. Specific embodiments

[0052] Next, the technical solutions in the embodiments of this invention application will be clearly and completely described in conjunction with the attached drawings in the embodiments of this invention application. Obviously, the described embodiments are only a part of the embodiments of this invention application, rather than all the embodiments; based on the embodiments in this invention application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this invention application.

[0053] In the description of this invention application, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or position relationship are based on the orientation or position relationship shown in the attached drawings, and are only for the convenience of describing this invention application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to this invention application. To make the purpose, technical solutions and advantages of this invention application clearer, the embodiments of this invention application will be further described in detail below in conjunction with the attached drawings.

[0054] Method for Measuring Cashmere Fiber Length Based on Convolutional Neural Network

[0055] As Figure 1 shown, in the first aspect, a cashmere fiber length measurement method based on convolutional neural network includes:

[0056] Step S100: Obtain multiple groups of fiber image sample datasets. Each group of fiber image sample datasets includes: fiber sample images, fiber sample binary images corresponding to the fiber sample images, and fiber sample repaired images corresponding to the fiber sample binary images.

[0057] Specifically, step S100 includes:

[0058] Step S101: Make cashmere fiber samples and obtain multiple groups of fiber sample images.

[0059] Specifically, step S101 includes: Making cashmere fiber samples: First, prepare two black cardboards that are 10 cm long and 1 cm wide. Stick black double-sided tape on one of them. Take about 0.3 g of cashmere samples. Pinch one end by hand and comb the other end with a steel comb until the fibers are straight. Paste the straightened part on the double-sided tape. Repeat the above steps until the double-sided tape on the cardboard is covered with fibers. Then, cover the other cardboard on the double-sided tape and press it tightly to fix the fibers between the two cardboards. Finally, comb the samples neatly on the velvet board. After ensuring that there are no obvious tangles at both ends, use a camera to take pictures of the cashmere fiber samples to obtain fiber sample images.

[0060] Step S102: Perform image processing on each group of fiber sample images to obtain each group of fiber sample binary images, and perform processing on each group of fiber sample binary images to obtain each group of fiber sample repaired images, thereby obtaining multiple groups of fiber sample binary images and multiple groups of fiber sample repaired images.

[0061] Specifically, step S102 includes:

[0062] Step S1021: Grayscale each group of fiber sample images, that is, convert the pixel points originally represented by three values of red, green, and blue into a single grayscale value, and then perform binary processing. Convert the pixels with grayscale values higher than the set threshold into white pixels, representing the fiber body, and convert the pixels lower than the threshold into black pixels, representing the background, thereby obtaining each group of fiber sample binary images.

[0063] Step S1022: Crop each group of fiber sample images and each group of fiber sample binary images. After cropping, the size of both is 256×256. Fill the insufficient parts with black pixels. Using painting software, place each group of fiber sample images at the bottom layer and each group of fiber sample binary images at the upper layer. Use a white pixel pen to depict and connect all the fiber parts that are disconnected due to binary processing to obtain fiber sample repaired images.

[0064] Step S103: Correlate each group of fiber sample images, each group of fiber sample binary images, and each group of fiber sample repaired images correspondingly to form each group of fiber image sample datasets, thereby obtaining multiple groups of fiber image sample datasets.

[0065] In some embodiments, a high-resolution camera can be used to capture images of cashmere fiber samples. By grayscaling and binarization, noise is removed, the fiber region is extracted, and the image is segmented into small pieces suitable for processing by a deep learning model.

[0066] Step S200: Based on the U-net model, a fiber fracture repair model is trained and constructed according to multiple groups of fiber image sample datasets.

[0067] In some embodiments, the U-Net model is a deep learning model designed specifically for image segmentation tasks. Its structure is in the shape of a "U", including an encoder (downsampling path), a decoder (upsampling path), and an output layer. Skip connections are made between the encoder and the decoder. This architecture achieves high-precision pixel-level segmentation through multi-level feature extraction and fusion. The U-Net model is also an end-to-end model that can directly generate segmentation results from the original image input without additional manual feature extraction or preprocessing steps, greatly simplifying the implementation process of image segmentation tasks and performing well in fields such as medical image analysis and remote sensing image processing.

[0068] In some embodiments, the encoder is the downsampling path of the U-Net model, which is used to extract high-level features of the input image and gradually reduce the spatial resolution of the feature map. The encoder consists of multiple convolutional layers and max pooling layers alternating. Each group of convolutional layers contains two convolutional operations, followed by a ReLU activation function for non-linear transformation. Subsequently, downsampling is performed through the max pooling layer to reduce the spatial size of the feature map. As the network deepens, the spatial size of the feature map gradually decreases, and the number of channels gradually increases, thereby capturing more semantic information.

[0069] In some embodiments, the decoder is the upsampling path of the U-Net model, which is used to restore the high-level features extracted by the encoder to the resolution of the original image and generate a segmentation result of the same size as the input image. The decoder consists of upsampling layers (such as transposed convolution or bilinear interpolation) and convolutional layers. The upsampling layer is used to increase the spatial size of the feature map, and the convolutional layer further refines the features. As the network becomes shallower, the spatial size of the feature map gradually increases, and the number of channels gradually decreases, finally restoring to the resolution of the input image.

[0070] In some embodiments, skip connections are one of the core innovations of the U-Net model, which are used to fuse low-level features in the encoder with high-level features in the decoder to avoid the loss of detailed information caused by downsampling. Through skip connections, the decoder can utilize the feature maps at the corresponding positions in the encoder, retain more local detailed information, and thus improve the segmentation accuracy. At each layer of the decoder, the feature map of the corresponding layer in the encoder is directly concatenated to the feature map of the current layer (usually through a concatenation operation), and then convolutional operations are performed to fuse the features.

[0071] In some embodiments, the output layer is the last part of the U-Net model and is used to generate the final segmentation result. The output layer typically includes a 1x1 convolutional layer for adjusting the number of channels to the number of target classes, followed by a softmax or sigmoid activation function for classification. The output layer generates the probability distribution of the class to which each pixel belongs, completing the pixel-level classification task.

[0072] In some embodiments, the multi-group fiber image sample dataset is divided into a training fiber image sample dataset, a validation fiber image sample dataset, and a test fiber image sample dataset, and the ratio of the training fiber image sample dataset, the validation fiber image sample dataset, and the test fiber image sample dataset is 8:1:1. At the same time, the initial fiber fracture repair model is trained using the training fiber image sample dataset and the validation fiber image sample dataset. After training, the trained fiber fracture repair model is obtained. Then, the test fiber image sample dataset is input into the trained fiber fracture repair model for testing, and the trained fiber fracture repair model is optimized according to the test results to obtain the final fiber fracture repair model.

[0073] In some embodiments, by using the U-Net model, a fiber fracture repair model dedicated to fiber break repair can be constructed. Each group of fiber sample images and each group of fiber sample binary images are used as inputs at the same time. The grayscale image of each group of fiber sample images can provide the morphology and background information of the fiber, and each group of fiber sample binary images can highlight the structural features of the fiber, enhancing the ability to locate fiber fracture points.

[0074] In some embodiments, the fiber fracture repair model includes an encoder, a decoder, and an output layer. First, the encoder is alternately composed of multiple convolutional layers and max-pooling layers. Each group of convolutional layers contains two convolutional operations. The encoder uses a 3-pixel × 3-pixel convolutional kernel, which can capture the local detail features of the image. The number of convolutional channels is gradually increased according to the network depth to enhance the diversity of fiber feature extraction. Then, the encoder and the decoder are connected by skip connections to directly transfer the high-resolution features in the encoder to the corresponding positions in the decoder, effectively restoring the details of the fiber edge and thus reducing the problem of fiber feature loss. Finally, a Sigmoid activation function is used in the output layer to map the feature map to the [0,1] interval to generate a fiber fracture repair probability map. The fiber fracture repair probability map represents the probability that each pixel belongs to the fiber or the background. In subsequent processing, through threshold operation, the fiber fracture repair probability Figure 2 is thresholded to obtain the fiber sample repair image, that is, the fiber fracture repair model can output the fiber sample repair image.

[0075] Specifically, step S200 includes:

[0076] Step S201, construct an initial fiber fracture repair model and set the parameters of the initial fiber fracture repair model: the number of training iterations is α, the batch size is β, the optimizer is the Adam optimizer, and the initial learning rate is γ;

[0077] Step S202, introduce an early stopping strategy. When the loss value of the validation fiber image sample dataset does not decrease in consecutive δ iterations, terminate the training of the fiber fracture repair model in advance to prevent overfitting of the fiber fracture repair model. After stopping the training, output the loss value change graph, pixel accuracy, precision, and average training time of the training fiber image sample dataset, the loss value change graph, pixel accuracy, precision, and average training time of the validation fiber image sample dataset, determine the final parameters, and obtain the trained fiber fracture repair model; input the test fiber image sample dataset into the trained fiber fracture repair model for testing, optimize the trained fiber fracture repair model according to the test results, and then obtain the final fiber fracture repair model;

[0078] In step S202, the loss value of the training fiber image sample dataset is calculated using the training loss function. The training loss function Lt is:

[0079] Lt = -θ·(1 - p t ) η ·[y·logp i +(1 - y)·log(1 - p i )]

[0080] where p i is the predicted value of the i-th pixel of the fiber repair image. If p i is greater than or equal to 0.5, then this pixel is considered a fiber; if p i is less than 0.5, then this pixel is considered a background, and the value range is 0 ≤ p i ≤ 1; y is a parameter, y takes 0 or 1. When y = 1, p t = p i , when y = 0, p t = 1 - p i ; θ is a balance factor; η is a weight.

[0081] It should be understood that the loss value of the validation fiber image sample dataset is not directly calculated by the loss function, but is obtained by comparing the result predicted by the model with the real fiber sample repair image. The loss values of the training fiber image sample dataset and the validation fiber image sample dataset are important indicators in the model training process, which can more accurately evaluate the performance of the model and guide the optimization and adjustment of the model.

[0082] Step S300: Obtain multiple groups of fiber sample cross-detection image sets. Each group of fiber image sample data sets includes: fiber sample repair images and the corresponding fiber sample cross images of the fiber sample repair images; construct a fiber cross-detection model based on the ResNet-50 encoder.

[0083] Specifically, step S300 includes:

[0084] Step S301: Annotate each group of fiber sample repair images, and use annotation software to mark the fiber cross parts in the fiber sample repair images with polygons to obtain each group of fiber sample cross images, and then obtain multiple groups of fiber sample cross images.

[0085] Step S302: Correlate each group of fiber sample repair images and each group of fiber sample cross images correspondingly to form each group of fiber sample cross-detection image sets, and then obtain multiple groups of fiber sample cross-detection image sets.

[0086] In some embodiments, the multiple groups of fiber sample cross-detection image sets are divided into a training fiber sample cross-detection image set, a validation fiber sample cross-detection image set, and a test fiber sample cross-detection image set, and the ratio of the training fiber sample cross-detection image set, the validation fiber sample cross-detection image set, and the test fiber sample cross-detection image set is 8:1:1; at the same time, use the training fiber sample cross-detection image set and the validation fiber sample cross-detection image set to train the initial fiber cross-detection model; after training, obtain the trained fiber cross-detection model, and then input the test fiber sample cross-detection image set into the trained fiber cross-detection model for testing, and optimize the trained fiber cross-detection model according to the test results to obtain the final fiber cross-detection model.

[0087] Step S303: Construct a fiber cross-detection model based on the ResNet-50 convolutional neural network.

[0088] In some embodiments, ResNet-50 (Residual Network-50) is a deep convolutional neural network architecture. The ResNet-50 convolutional neural network includes: an input layer, a convolutional layer, a max pooling layer, a residual module, and a fully connected layer.

[0089] In some embodiments, the input layer usually receives image data of a fixed size, such as a common RGB image of 256×256 pixels, and is used to preprocess the original image data, such as normalization operations, to make the data distribution within a suitable range to accelerate the convergence of the model.

[0090] In some embodiments, the convolutional layer includes: a first convolutional layer (7x7 convolution with a stride of 2), which uses a convolutional kernel of size 7×7 and a stride of 2 to perform a convolution operation on the input image. The convolutional layer extracts the preliminary features of the image through the convolutional kernel. At the same time, since the stride is 2, the image size is halved at this layer, reducing the subsequent computational amount.

[0091] In some embodiments, a max pooling layer (3x3 max pooling with a stride of 2) follows the first convolutional layer immediately. It performs a max pooling operation with a window of size 3×3 and a stride of 2. Max pooling is used to further reduce the image size and computational amount while retaining the main feature information in the image. After this step, the image size is halved again.

[0092] In some embodiments, the residual module is the core component of the ResNet-50 convolutional neural network. The ResNet-50 convolutional neural network contains multiple residual modules, and these modules are organized according to different stages. Specifically, ResNet-50 contains 4 different stages of residual modules, and the number of residual modules in each stage is 3, 4, 6, and 3 respectively. The structures of the residual modules within each stage are similar, but parameters such as the number of convolutional kernels and the stride will be adjusted. Each residual module contains multiple convolutional layers (usually 2 - 3), and feature extraction operations are performed in these convolutional layers. At the same time, through a shortcut connection, the input is directly added to the output of the convolutional layer. This structure makes the network easier to optimize during training, can effectively alleviate the vanishing gradient problem, and thus a deeper network can be constructed.

[0093] In some embodiments, after passing through multiple convolutional layers and residual modules, the fully connected layers perform a global average pooling operation on the feature map output by the last stage, converting it into a feature vector of a fixed length. Then, this feature vector is sequentially connected to multiple fully connected layers. The role of the fully connected layers is to comprehensively analyze and classify the extracted features, and finally output a probability vector, where the length of the vector is equal to the number of categories in the classification task.

[0094] Through the convolution operations in the convolutional layer and the residual module, ResNet-50 can automatically learn features at different levels and abstraction degrees in the image. After a series of feature extractions, the fully connected layer uses the extracted features for classification decision-making. According to the degree of association between the feature vector and different categories, it calculates the probability that the input image belongs to each category, and the fiber crossing detection model can output the image of the fiber crossing part.

[0095] Specifically, step S303 includes:

[0096] Step S3031, construct an initial fiber crossing detection model and set the parameters of the initial fiber crossing detection model: the number of training iterations is α', the batch size is β', the optimizer is the Adam optimizer, and the initial learning rate is γ';

[0097] Step S3032, introduce an early stopping strategy. When the loss value of the validation fiber sample crossing detection image set does not decrease in consecutive δ' iterations, terminate the training of the fiber crossing detection model in advance to prevent overfitting of the fiber crossing detection model. After stopping the training, output the loss value change graph, pixel accuracy, precision, and average training time of the training fiber sample crossing detection image set, the loss value change graph, pixel accuracy, precision, and average training time of the validation fiber sample crossing detection image set, determine the final parameters, and obtain the trained fiber crossing detection model; input the test fiber sample crossing detection image set into the trained fiber crossing detection model for testing, and then obtain the final fiber crossing detection model.

[0098] In step S3032, the loss value of the training fiber sample crossing detection image set is calculated using a mixed loss function, and the mixed loss function is:

[0099] L m = ω D ·L D + ω F ·L F

[0100]

[0101] L F = -θ'·(1 - p s ) σ′ ·logp s

[0102] where L m is the mixed loss function, ω D is the weight of the first loss function, ω F is the weight of the second loss function, L D is the first loss function, L F is the second loss function; p j is the predicted value of the j-th pixel of the fiber crossing part image. If p j is greater than or equal to 0.5, it is considered that this pixel is a fiber crossing point; if p j is less than 0.5, it is considered that this pixel is not a fiber crossing point, and the value range is 0 ≤ p j ≤ 1; g j is the true label of the j-th pixel, g jis 0 or 1, where 1 represents that the j-th pixel is indeed a fiber crossing point, and 0 represents that the j-th pixel is not a fiber crossing point; N is the number of all pixels in the image, and ∈ is a smoothing parameter; when g j = 1, p s = p j , when g j = 0, p s = 1 - p j ; θ' is a balance coefficient, and σ' is a weight coefficient.

[0103] It should also be understood that the loss value of the verification fiber sample cross-detection image set is not directly calculated by the loss function, but is obtained by comparing the results predicted by the fiber cross-detection model with the real fiber sample repair image; the loss values of the training fiber sample cross-detection image set and the verification fiber sample cross-detection image set are important indicators in the training process of the fiber cross-detection model, which can more accurately evaluate the performance of the fiber cross-detection model and guide the optimization and adjustment of the model.

[0104] In the present invention application, since the fiber crossing part (which can also be called the fiber crossing point) usually accounts for a very low proportion of the whole picture, class weights can be added to increase the weight value of the fiber crossing part and decrease the weight value of the non-crossing area, thereby enhancing the sensitivity of the fiber cross-detection model to fiber crossing points and alleviating the class imbalance problem; at the same time, an improved ResNet-50 convolutional neural network architecture can also be used to detect fiber crossing points, combined with a hybrid loss function, to improve the accuracy of the model in small area segmentation and effectively alleviate the class imbalance problem.

[0105] Step S400: Obtain an actual cashmere fiber sample image to be detected, process the actual cashmere fiber sample image to be detected, and use a fiber break repair model and a fiber cross-detection model to calculate the average fiber length of the actual cashmere fiber sample to be detected.

[0106] Specifically, step S400 includes:

[0107] Step S401, make an actual cashmere fiber sample and obtain an actual cashmere fiber sample image to be detected;

[0108] First, prepare two pieces of black cardboard that are 10 cm long and 1 cm wide. Stick black double-sided tape on one of them. Take an actual cashmere sample weighing about 0.3 g. Pinch one end by hand and use a steel comb to comb the other end until the fibers are straight. Paste the straightened part on the double-sided tape. Repeat the above steps until the double-sided tape on the cardboard is covered with fibers. Then, cover the other piece of cardboard on the double-sided tape and press it firmly to fix the fibers between the two pieces of cardboard. Finally, comb the sample neatly on the velvet board. After ensuring that there are no obvious tangles at both ends, take a photo of the actual cashmere fiber sample with a camera to obtain an image of the actual cashmere fiber sample to be detected.

[0109] Step S402: Perform image processing on the image of the actual cashmere fiber sample to be detected to obtain a binary image of the actual cashmere fiber sample to be detected.

[0110] Specifically, step S402 includes: grayscale the image of the actual cashmere fiber sample to be detected, that is, convert the pixel points originally represented by three values of red, green, and blue into a single grayscale value, and then perform binary processing. Convert the pixels with grayscale values higher than the set threshold into white pixels, representing the fiber body of the actual cashmere fiber sample to be detected, and convert the pixels lower than the threshold into black pixels, representing the background, so as to obtain a binary image of the actual cashmere fiber sample to be detected.

[0111] Step S403, according to the image of the actual cashmere fiber sample to be detected and the binary image of the actual cashmere fiber sample to be detected, use the fiber break repair model to obtain a repaired image of the actual cashmere fiber sample to be detected, and calculate the total number of fiber pixels FP and the total number of fiber roots F in the repaired image of the actual cashmere fiber sample to be detected. According to the repaired image of the actual cashmere fiber sample to be detected, use the fiber crossing detection model to obtain an image of the fiber crossing part, and calculate the number of fiber pixels CP in the fiber crossing part. The average fiber length FL is:

[0112]

[0113] Where d is the fiber width value, dimensionless; FP is the total number of fiber pixels in the repaired image of the actual cashmere fiber sample to be detected, CP is the number of fiber pixels in the fiber crossing part, F is the total number of fiber roots in the repaired image of the actual cashmere fiber sample to be detected, dimensionless, w is the cardboard width, and μ is the ratio of the number of velvet board pixels to the actual length of the velvet board, unit: number of pixels / mm.

[0114] In some embodiments, when using a drawing software for the break-filling operation, the brush width used is 3, which is consistent with the width of most fibers in the image. When calculating, the fiber width value d can be defaulted to 3.

[0115] In step S403, a column of pixels is taken from the leftmost end of the restored image of the actual cashmere fiber sample to be detected, and the number of alternations of black and white pixels in this column from top to bottom is counted; when white pixels appear continuously three times, the number of alternations is additionally increased by two times to compensate for possible omissions caused by fiber occlusion; subsequently, with a step size of 10 pixels, the above operation is repeated within the image width range of the restored image of the actual cashmere fiber sample to be detected, the number of alternations of all columns is counted, and finally the maximum value of the number of alternations is divided by 2 to be used as the total number of fiber roots F.

[0116] In step S403, the total number of fiber pixels FP in the restored image of the actual cashmere fiber sample to be detected can be counted by counting all white pixels in the restored image of the actual cashmere fiber sample to be detected. According to the restored image of the actual cashmere fiber sample to be detected, a fiber crossing part image (which can also be understood as or referred to as the fiber crossing part image of the actual cashmere fiber sample to be detected) is obtained by using a fiber crossing detection model, and the white pixels of the fiber crossing part image can be calculated to obtain the number of fiber pixels CP in the fiber crossing part.

[0117] As Figure 2 shown, in the second aspect, a cashmere fiber length measurement system based on a convolutional neural network is used to implement the cashmere fiber length measurement method based on a convolutional neural network described in any one of the above first aspects, and includes:

[0118] It should be noted that the cashmere fiber length measurement system based on a convolutional neural network in this invention application is used to implement the cashmere fiber length measurement method based on a convolutional neural network in any one of the first aspects. Correspondingly, it also includes: all the technical problems, technical solutions, and technical effects recorded in any one of the cashmere fiber length measurement methods based on a convolutional neural network in the first aspect. This invention application will not elaborate here.

[0119] An acquisition and processing module, which acquires and processes an image of an actual cashmere fiber sample to be detected;

[0120] A calculation module, which uses a fiber breakage repair model and a fiber crossing detection model to calculate the average fiber length of the actual cashmere fiber sample to be detected;

[0121] A display module, which displays and outputs the average fiber length of the actual cashmere fiber sample to be detected calculated.

[0122] Optionally, the acquisition and processing module includes: an actual image acquisition unit and an actual image processing unit. The actual image acquisition unit acquires an image of an actual cashmere fiber sample to be detected, and the actual image processing unit performs image processing on the image of the actual cashmere fiber sample to be detected to obtain a binary image of the actual cashmere fiber sample to be detected.

[0123] In some embodiments, first, prepare two pieces of black cardboard that are 10 cm long and 1 cm wide. Stick black double-sided tape on one of them. Take an actual cashmere sample of about 0.3 g. Pinch one end by hand and comb the other end with a steel comb until the fibers are straight. Paste the straightened part on the double-sided tape. Repeat the above steps until the double-sided tape on the cardboard is covered with fibers. Then, cover the other piece of cardboard on the double-sided tape and press it firmly to fix the fibers between the two pieces of cardboard. Finally, comb the sample neatly on the velvet board. After ensuring that there are no obvious tangles at both ends, the actual cashmere fiber sample to be detected is made.

[0124] In some embodiments, the actual image acquisition unit includes a camera. The camera is used to take pictures of the actual cashmere fiber sample to be detected, and then an image of the actual cashmere fiber sample to be detected is obtained.

[0125] In some embodiments, the actual image processing unit grayscales the image of the actual cashmere fiber sample to be detected, that is, converts the pixel points originally represented by three values of red, green, and blue into a single grayscale value, and then performs binarization processing. The pixels with grayscale values higher than the set threshold are converted into white pixels, representing the fiber body of the actual cashmere fiber sample to be detected, and the pixels lower than the threshold are converted into black pixels, representing the background, thereby obtaining a binarized image of the actual cashmere fiber sample to be detected.

[0126] Optionally, the calculation module includes: a fiber break repair model unit, a fiber crossing detection model unit, and a length calculation unit;

[0127] The fiber break repair model unit obtains a repaired image of the actual cashmere fiber sample to be detected according to the image of the actual cashmere fiber sample to be detected and the binarized image of the actual cashmere fiber sample to be detected;

[0128] The fiber crossing detection model unit obtains an image of the fiber crossing part according to the repaired image of the actual cashmere fiber sample to be detected;

[0129] The length calculation unit calculates the total number of fiber pixels FP, the total number of fiber roots F, and the number of fiber pixels CP in the fiber crossing part in the repaired image of the actual cashmere fiber sample to be detected, and calculates the average fiber length FL as:

[0130]

[0131] Where d is the fiber width value, FP is the total number of fiber pixels in the repaired image of the actual cashmere fiber sample to be detected, CP is the number of fiber pixels in the fiber crossing part, F is the total number of fiber roots in the repaired image of the actual cashmere fiber sample to be detected, w is the cardboard width, and μ is the ratio of the number of velvet board pixels to the actual length of the velvet board, unit: number of pixels / mm.

[0132] In some embodiments, the fiber fracture repair model unit adopts a fiber fracture repair model, and the specific training and construction method of the fiber fracture repair model can be the same as all the technical problems, technical solutions, and technical effects described in step S100 and step S200.

[0133] In some embodiments, the fiber crossing detection model unit adopts a fiber crossing detection model, and the specific training and construction method of the fiber crossing detection model can be the same as all the technical problems, technical solutions, and technical effects described in step S300.

[0134] Computer-readable Storage Medium

[0135] In a third aspect, the present invention provides a computer-readable storage medium, on which instructions are stored. When the instructions run on a computer, they execute the method for measuring the length of cashmere fibers based on a convolutional neural network as described in any one of the first aspects.

[0136] Device for Measuring Cashmere Fiber Length Based on Convolutional Neural Network

[0137] In a fourth aspect, the present invention provides a device for measuring the length of cashmere fibers based on a convolutional neural network, including a memory and a processor connected by communication. The memory is used to store a computer program, and the processor is used to read the computer program and execute the method for measuring the length of cashmere fibers based on a convolutional neural network as described in any one of the first aspects.

[0138] In some embodiments, the device for measuring the length of cashmere fibers based on a convolutional neural network includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in any one of the methods for measuring the length of cashmere fibers based on a convolutional neural network in the first aspect.

[0139] Those skilled in the art can understand that the schematic diagram is only an example of the terminal device, and does not constitute a limitation on the terminal device. It may include more or fewer components than shown, or combine some components, or different components. For example, the terminal device may further include input and output devices, network access devices, buses, etc.

[0140] The so-called processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the terminal device and connects various parts of the entire terminal device through various interfaces and lines.

[0141] The memory can be used to store the computer programs and / or modules. The processor realizes various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash device, or other volatile solid-state storage devices.

[0142] Among them, if the modules / units integrated in the terminal device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention application, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0143] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the drawings of the device embodiments provided in the present invention application, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0144] The above is the preferred embodiment of the present invention application. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention application, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention application.

Claims

1. A cashmere fiber length measurement method based on convolutional neural network, characterized in that: include: Step S100, obtaining multiple groups of fiber image sample data sets, each group of fiber image sample data sets including: a fiber sample image, a fiber sample binarized image corresponding to the fiber sample image, and a fiber sample repaired image corresponding to the fiber sample binarized image; Step S200, based on the U-net model, training and constructing a fiber fracture repair model according to multiple groups of fiber image sample data sets; Step S300, obtaining multiple groups of fiber sample cross detection image sets, each group of fiber image sample data sets includes: a fiber sample repair image, a fiber sample cross image corresponding to the fiber sample repair image, and constructing a fiber cross detection model based on a ResNet-50 convolutional neural network; Step S400: obtaining an actual cashmere fiber sample image to be detected, processing the actual cashmere fiber sample image to be detected, and using a fiber breakage repair model and a fiber crossing detection model to calculate the average fiber length of the actual cashmere fiber sample to be detected.

2. The cashmere fiber length measurement method based on convolutional neural network according to claim 1, characterized in that: Step S100 includes: Step S101: preparing cashmere fiber samples and obtaining multiple groups of fiber sample images; Step S102: performing image processing on each group of fiber sample images to obtain each group of fiber sample binary images, and processing each group of fiber sample binary images to obtain each group of fiber sample repair images, thereby obtaining multiple groups of fiber sample binary images and multiple groups of fiber sample repair images; Step S103: Correspondingly associating each group of fiber sample images, each group of fiber sample binary images, and each group of fiber sample repaired images to form each group of fiber image sample data sets, thereby obtaining multiple groups of fiber image sample data sets.

3. The cashmere fiber length measurement method based on convolutional neural network according to claim 2, characterized in that: Step S101 includes: Preparation of cashmere fiber samples: First, prepare two pieces of black cardboard with a length of 10 cm and a width of 1 cm, and stick black double-sided tape on one of them. Take about 0.3 g of cashmere sample, pinch one end with your hand, comb the other end with a steel comb until the fiber is straight, and stick the combed straight part on the double-sided tape; repeat the above steps until the double-sided tape on the cardboard is covered with fibers; then, cover the double-sided tape with another piece of cardboard and press it tightly to fix the fibers between the two pieces of cardboard; finally, comb the sample neatly on the cashmere board to ensure that there is no obvious entanglement at both ends, and use a camera to take pictures of the cashmere fiber sample multiple times to obtain fiber sample images, and then obtain multiple groups of fiber sample images.

4. The cashmere fiber length measurement method based on convolutional neural network according to claim 3, characterized in that: Step S102 includes: Step S1021, graying each group of fiber sample images, that is, converting the pixels originally represented by the three values ​​of red, green and blue into a single gray value, and then performing binarization processing, converting the pixels with gray values ​​higher than the set threshold into white pixels to represent the fiber body, and converting the pixels with gray values ​​lower than the threshold into black pixels to represent the background, thereby obtaining a binary image of each group of fiber samples; Step S1022, crop each group of fiber sample images and each group of fiber sample binary images, and the size of the two after cropping is 256×256, and the insufficient parts are filled with black pixels. Using drawing software, each group of fiber sample images is placed on the bottom layer, and each group of fiber sample binary images is placed on the upper layer, and a white pixel pen is used to draw and connect all fiber parts that are disconnected due to the binarization process to obtain a fiber sample repair image.

5. The cashmere fiber length measurement method based on convolutional neural network according to claim 3, characterized in that: Step S300 includes: Step S301, annotating each group of fiber sample repair images, using annotation software to use polygon marks on the fiber intersection parts in the fiber sample repair images to obtain each group of fiber sample intersection images, and then obtaining multiple groups of fiber sample intersection images; Step S302, correspondingly associating each group of fiber sample repair images and each group of fiber sample cross images to form each group of fiber sample cross detection image sets, thereby obtaining multiple groups of fiber sample cross detection image sets; Step S303: construct a fiber crossing detection model based on the ResNet-50 convolutional neural network.

6. The method for measuring cashmere fiber length based on convolutional neural network according to claim 5, characterized in that: Step S400 includes: Step S401, preparing an actual cashmere fiber sample and obtaining an actual cashmere fiber sample image to be tested; Step S402: performing image processing on the actual cashmere fiber sample image to be detected to obtain a binary image of the actual cashmere fiber sample to be detected; Step S403, according to the actual cashmere fiber sample image to be detected and the actual cashmere fiber sample binary image to be detected, the fiber fracture repair model is used to obtain the actual cashmere fiber sample repair image to be detected, and the overall fiber pixel number FP and the overall fiber root number F in the actual cashmere fiber sample repair image to be detected are calculated; according to the actual cashmere fiber sample repair image to be detected, the fiber intersection part image is obtained using the fiber intersection detection model, and the fiber pixel number CP of the fiber intersection part is calculated. The average fiber length FL is: Among them, d is the fiber width value, FP is the number of overall fiber pixels in the repaired image of the actual cashmere fiber sample to be tested, CP is the number of fiber pixels in the fiber crossing part, F is the number of overall fiber roots in the repaired image of the actual cashmere fiber sample to be tested, w is the width of the cardboard, μ is the ratio of the number of pixels of the velvet board to the actual length of the velvet board, unit: number of pixels / mm.

7. A cashmere fiber length measurement system based on a convolutional neural network, used to implement the cashmere fiber length measurement method based on a convolutional neural network as described in any one of claims 1 to 6, characterized in that: include: An acquisition and processing module is used to acquire and process the actual cashmere fiber sample image to be detected; The calculation module uses the fiber breakage repair model and the fiber cross detection model to calculate the average fiber length of the actual cashmere fiber sample to be tested; The display module displays and outputs the calculated actual average fiber length of the cashmere fiber sample to be tested.

8. The cashmere fiber length measurement system based on convolutional neural network according to claim 7, characterized in that: The acquisition and processing module includes: an actual image acquisition unit and an actual image processing unit. The actual image acquisition unit acquires the actual cashmere fiber sample image to be detected, and the actual image processing unit performs image processing on the actual cashmere fiber sample image to be detected to obtain a binary image of the actual cashmere fiber sample to be detected.

9. The method for measuring cashmere fiber length based on convolutional neural network according to claim 8, characterized in that: The calculation module includes: a fiber breakage repair model unit, a fiber crossing detection model unit and a length calculation unit; The fiber fracture repair model unit obtains a repair image of the actual cashmere fiber sample to be detected according to the actual cashmere fiber sample image to be detected and the actual cashmere fiber sample binary image to be detected; The fiber crossing detection model unit obtains a fiber crossing part image according to the repair image of the actual cashmere fiber sample to be detected; The length calculation unit calculates the number of overall fiber pixels FP and the number of overall fiber roots F in the repaired image of the actual cashmere fiber sample to be tested, the number of fiber pixels CP in the fiber crossing part of the image of the fiber crossing part, and calculates the average fiber length FL as: Among them, d is the fiber width value, FP is the number of overall fiber pixels in the repaired image of the actual cashmere fiber sample to be tested, CP is the number of fiber pixels in the fiber crossing part, F is the number of overall fiber roots in the repaired image of the actual cashmere fiber sample to be tested, w is the width of the cardboard, μ is the ratio of the number of pixels of the velvet board to the actual length of the velvet board, unit: number of pixels / mm.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the cashmere fiber length measurement method based on a convolutional neural network described in any one of claims 1 to 6 is executed.

Images