Papermaking fiber classification intelligent analysis model based on computer vision

Abstract

The invention discloses a papermaking fiber classification intelligent analysis model based on computer vision, and aims at solving the problems that an existing model is low in precision and poor in impurity interference resistance. The model takes improved ResNeXt50 as a backbone, and is internally provided with an ECA attention submodule, a CSIM cross-scale feature interaction submodule, an NASM noise suppression submodule, a matching hierarchical category mapping module, an intelligent data processing module, a training evaluation module and a time monitoring module. Data balance is guaranteed through stratified sampling and differential preprocessing, training is completed through label smooth cross entropy loss and an Adam optimizer, and hierarchical mapping from 17 fiber subclasses to three major classes is achieved. The large class classification accuracy of the model exceeds 99%, the reasoning efficiency is more than 10 times that of traditional manual detection, the noise robustness is high, the expansion capacity is good, the fine classification requirement of the papermaking industry can be met, and the model is suitable for industrial real-time detection scenes.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning, and paper industry inspection technology, and in particular to an intelligent analysis model for paper fiber classification based on an improved ResNeXt50 network. It is applicable to the accurate classification and identification of three major categories of paper fibers: single virgin pulp fibers, secondary fibers, and mixed virgin pulp fibers, and also has the scalability to perform sub-category classification. Background Technology

[0002] In the papermaking and paper recycling industry, the determination of the major categories of papermaking fibers directly affects the setting of paper production process parameters, quality control of recycled paper, and traceability of raw materials. Currently, there are many pain points in papermaking fiber classification and detection: traditional manual microscopic identification methods are inefficient, highly subjective, and easily influenced by human experience; conventional deep learning classification models (such as basic ResNeXt and ResNet) have not been customized to the characteristics of papermaking fiber images, and cannot effectively capture impurities and noise in secondary fibers, or the cross-scale feature differences between different fibers, resulting in insufficient accuracy in major category classification; most models lack a systematic training time monitoring and multi-dimensional evaluation system, which cannot guarantee the stability and traceability of the models in industrial applications; at the same time, although there are theoretical solutions for minor category classification in existing technologies, in actual implementation, there are problems such as uneven sample distribution and excessive model complexity leading to insufficient inference efficiency, making it difficult to adapt to industrial real-time detection scenarios.

[0003] Therefore, there is an urgent need for an intelligent analysis model that can adapt to the characteristics of paper fiber images, support accurate classification of major categories, have efficient training and accurate evaluation capabilities, and can be expanded to include minor category classification functions, in order to address the shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a computer vision-based intelligent analysis model for paper fiber classification. Through customized network modules, hierarchical category mapping, intelligent data processing, and end-to-end monitoring, efficient and accurate classification of paper fibers is achieved.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A computer vision-based intelligent analysis model for paper fiber classification includes:

[0007] The feature enhancement network module, based on an improved pre-trained ResNeXt50-32x4d backbone network, includes an ECA attention submodule, a cross-scale feature interaction submodule (CSIM), a noise adaptive suppression submodule (NASM), and a dual-pooling classification head. The ECA attention submodule performs adaptive feature calibration along the channel dimension, strengthening the weights of key fiber texture features. The CSIM submodule performs cross-scale interaction of local details, global contours, and statistical features of the paper fiber image through multi-scale feature extraction. The NASM submodule suppresses noise and preserves effective fiber features. The dual-pooling classification head outputs the fiber classification results.

[0008] The hierarchical category mapping module stores the mapping relationship between subcategories and major categories of paper fibers, enabling automatic conversion from subcategory labels to major category indexes and providing label support for major category classification. This module has a built-in subcategory-major category mapping table that supports incremental entry of new fiber subcategories. The subcategory index-major category index mapping dictionary can be automatically synchronized and adjusted according to updates to the mapping table. It also reserves an interface for extending subcategory classification labels, allowing for expansion of subcategory classification functionality by increasing the number of classification head neurons.

[0009] The intelligent data processing module includes a data preprocessing submodule and a stratified sampling module. The data preprocessing submodule is used for data standardization and data augmentation; the stratified sampling module is used to ensure a balanced distribution of samples at each level.

[0010] The model training and evaluation module includes a training submodule and an evaluation submodule. The training submodule covers the complete process of data loading, model building, training, validation, testing, and model saving, and selects the optimal model weights through validation set accuracy. The evaluation submodule includes a major category classification report, a confusion matrix, and t-SNE visualization to achieve full-dimensional quantification of classification performance.

[0011] The time monitoring module is used for full-process training time statistics, including recording and visualization of total training time, time per epoch, and average time spent in the training / validation phases, and outputting a training time summary report to provide data support for efficiency evaluation of model industrial deployment.

[0012] Preferably, in the feature enhancement network module, the convolution kernel size configured in the ECA attention submodule includes, but is not limited to, 3, 5, and 7. The kernel size is designed to adapt to the feature scale of different network levels through hierarchical convolution kernel size design. The kernel size is designed according to the texture complexity of the paper fiber image and the resolution of the microscopic imaging.

[0013] Preferably, in the hierarchical category mapping module, the stored sub-category mapping table includes, but is not limited to, the mapping relationship between sub-categories of virgin pulp fibers made from various common plant fiber raw materials and recycled secondary fibers, and the three major categories of single virgin pulp fibers, secondary fibers, and mixed virgin pulp fibers. The mapping table supports incremental entry of new fiber sub-categories / majors, and the sub-category index-major category index mapping dictionary can be automatically synchronized and adjusted according to the update of the mapping table.

[0014] Preferably, in the intelligent data processing module, the data augmentation strategies adopted by the data preprocessing submodule include, but are not limited to, random cropping (size range 224×224-256×256), horizontal / vertical flipping, ±15° rotation, 0.1-0.3 standard deviation Gaussian blur, and brightness / contrast ±0.2 color dithering. The activation ratio and parameter threshold of the augmentation strategies are dynamically adjusted according to the overfitting degree of the training set samples.

[0015] Preferably, in the model training and evaluation module, the smoothing coefficient of the labeled smoothed cross-entropy loss function includes, but is not limited to, 0.05, 0.1, and 0.15. Label smoothing reduces the model's overconfidence in hard labels, and the smoothing coefficient is adapted according to the proportion of misclassified samples in the classification confusion matrix of the validation set. The quantitative indicators output by the evaluation submodule include, but are not limited to, the precision, recall, F1 score, confusion matrix, and Top-k accuracy of the major categories.

[0016] Preferably, in the time monitoring module, the training time statistics include, but are not limited to, total training time, single epoch time, single batch iteration time, and average training / validation phase time. The training efficiency can be monitored in all dimensions through time statistics. The output format of the statistical results includes, but is not limited to, text reports and line charts. The visualization format is customized according to the data analysis needs of industrial deployment.

[0017] Preferably, the analysis model uses an intelligent analysis method for paper fiber classification based on an improved ResNeXt, comprising the following steps:

[0018] S1. Model initialization, configuring the device environment (prioritizing GPU), setting random seeds to ensure reproducibility of experiments, loading the subclass-major class mapping table and completing the association mapping between subclass indexes and major classes;

[0019] S2. Dataset Construction and Preprocessing: Scan fiber image dataset catalog, construct a flat dataset containing image paths and subclass labels; complete stratified sampling and differential preprocessing of training / validation / test sets through intelligent data processing module, and generate data loader;

[0020] S3. Feature enhancement network construction: Initialize the improved ResNeXt50 network, insert ECA, CSIM, and NASM sub-modules, freeze the shallow parameters of the backbone network, and deploy the network to the target computing device.

[0021] S4. Model training: The model is trained using a labeled smooth cross-entropy loss function, Adam optimizer, and cosine annealing learning rate scheduler. The time consumption of each stage is recorded through a time monitoring module, and the optimal model weights are saved based on the accuracy of the validation set.

[0022] S5. Model Evaluation: Use the optimal model to reason on the test set. The model training and evaluation module completes the dual-dimensional classification effect evaluation of major and minor categories, and generates a classification report and confusion matrix.

[0023] S6. Model saving and deployment: Saves information such as model weights and classification mapping relationships, training history, and evaluation metrics, and supports modular deployment and industrial system integration.

[0024] The present invention has the following beneficial effects:

[0025] 1. Significantly Improved Classification Accuracy: Through the synergistic effect of three customized modules—ECA, CSIM, and NASM—the present invention effectively solves the challenge of feature extraction from fiber images. Experiments show that on the same test set, the classification accuracy of this invention reaches 99.39%.

[0026] 2. Strong scalability: It reserves a sub-category classification interface, which can achieve a two-level classification of "major category - minor category" through simple model structure adjustment, without the need to reconstruct the model, thus reducing the cost of technology upgrade.

[0027] 3. Strong noise robustness: The NASM module effectively suppresses impurity noise in secondary fibers through adaptive threshold soft threshold filtering, solving the problem of traditional models being sensitive to noise and improving the industrial applicability of the model.

[0028] 4. Controllable and traceable training process: The time monitoring module provides complete training time statistics, and the evaluation module provides multi-dimensional performance indicators, making the model training process transparent and optimizable, and providing reliable data support for industrial deployment.

[0029] 5. The stratified sampling mechanism ensures a balanced distribution of samples across all categories, and the diversification of data enhances the model's generalization ability, solving the problem of model overfitting caused by small samples and uneven sample distribution. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0031] A computer vision-based intelligent analysis model for paper fiber classification includes:

[0032] The feature enhancement network module is built upon a pre-trained ResNeXt50-32x4d backbone network and includes an ECA attention submodule, a cross-scale feature interaction submodule (CSIM), a noise adaptive suppression submodule (NASM), and a dual-pooling classification head. The ECA attention submodule is configured with convolutional kernel sizes including, but not limited to, 3, 5, and 7. Specifically, a 3×1 kernel is configured for layer 2, a 5×1 kernel for layer 3, and a 7×1 kernel for layer 4. This hierarchical kernel size design adapts to the feature scales of different network layers, taking into account the texture complexity of the paper fiber image and the resolution of the microscopic imaging. The CSIM submodule sets up three parallel branches: a local branch (3×3 convolution), a global branch (5×5 convolution), and a global statistical branch (adaptive average pooling + 1×1 convolution). The output features of the branches are concatenated and then fused by 1×1 convolution to achieve residual connection, thus completing the cross-scale interaction between local details and global features of papermaking fibers. The NASM submodule first halves the number of input channels through 1×1 convolution, then achieves secondary fiber impurity noise suppression based on soft thresholding with adaptive threshold, and finally retains effective fiber features through feature enhancement and residual connection. The dual-pooling classification head fuses global average pooling and global max pooling features, and outputs the major classification results through multi-level fully connected layers and dropout regularization.

[0033] The hierarchical category mapping module stores the hierarchical mapping relationship between papermaking fiber subcategories and major categories, supporting automatic conversion from subcategory indexes to major category indexes. The subcategory-major category mapping table stored in the module includes five single-species virgin pulp fiber subcategories: Eucalyptus globulus, Pinus simonii, Eucalyptus urophylla, Pinus yunnanensis, and Phyllostachys edulis. It also includes mixed office waste paper deinked pulp fiber, packaging waste paper non-deinked pulp long fiber, packaging waste paper non-deinked pulp medium fiber, packaging waste paper non-deinked pulp short fiber, and packaging waste paper non-deinked pulp mixed fiber. The mapping table establishes a correspondence between five secondary fiber subcategories, seven mixed virgin pulp fiber subcategories (including softwood pulp with beating degrees of 30°SR, 40°SR, and 50°SR, hardwood pulp with beating degrees of 40°SR and 70°SR, and bamboo pulp with beating degrees of 40°SR and 50°SR) and the three major categories. The mapping table supports incremental entry of newly added fiber subcategories / majors, and the subcategory index-major category index mapping dictionary can be automatically synchronized and adjusted according to the mapping table update to adapt to the testing requirements of new fiber categories.

[0034] The intelligent data processing module includes a data preprocessing submodule and a hierarchical sampling module. The data preprocessing submodule configures enhancement strategies for the training set, such as random cropping (224×224, scaling range 0.8-1.0), random horizontal flipping (probability 0.5), random vertical flipping (probability 0.3), ±15° rotation, 0.1-2.0 standard deviation Gaussian blur, and color dithering with brightness / contrast / saturation ±0.2 and hue ±0.1. For the validation and test sets, it configures a 256-size Re... Standardized preprocessing with size+224 center cropping was performed, and all images were normalized to the ImageNet standard mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225). The stratified sampling module set the target sample threshold, with 80 images for the Vida subclass and 200 images for the other subclasses. Stratified random sampling was used to achieve an 8:1:1 training / validation / test set partition. When the samples were insufficient, they were supplemented by repeated sampling to ensure a balanced distribution of samples in each level of the category.

[0035] The model training and evaluation module is used to complete the transfer training of the model and evaluate the classification performance across all dimensions. In this module, a labeled smoothed cross-entropy loss function (smoothing coefficient 0.1) is used for loss calculation. An Adam optimizer (learning rate 5e-5, weight decay 5e-4) and a cosine annealing learning rate scheduler (T_max = 140, eta_min = 1e-6) are configured. During training, the parameters of the backbone network conv1, bn1, and layer1 are frozen to achieve transfer learning. The evaluation submodule can output a major category classification report, confusion matrix, weighted precision / recall / F1 score, and simultaneously calculate the major category classification accuracy for each subclass. Customized indicators such as single-class fiber false negative rate and false positive rate can also be added according to industrial quality control needs.

[0036] The time monitoring module is used to realize the time consumption statistics and visualization of the entire model training process. The time consumption statistics granularity of the time monitoring module includes total training time, single epoch time, average time of training / validation phase, and single batch iteration time. The statistical results can be output in the form of text reports, epoch time line graphs, phase time heatmaps, etc. The visualization form can be customized according to the data analysis needs of industrial deployment, providing data support for optimizing model training efficiency.

[0037] The analysis model uses an intelligent analysis method for paper fiber classification based on an improved ResNeXt, which includes the following steps:

[0038] S1. Initialize the model and dataset, configure GPU / CPU computing devices, set random seeds to ensure the reproducibility of experiments, load the subclass-major class mapping table and generate the subclass index-major class index mapping dictionary; scan the fiber image dataset directory, construct a flat dataset containing image paths and subclass labels, and complete the basic information verification of the dataset.

[0039] S2. Data stratified sampling and preprocessing: The stratified sampling module divides the training / validation / test sets according to the target sample threshold. Then, the data preprocessing submodule performs differentiated preprocessing on different datasets to generate a batch data loader that adapts to the model input.

[0040] S3. Feature enhancement network construction and training: Initialize the improved ResNeXt50 network and insert each customized sub-module. After freezing the shallow parameters of the backbone network, deploy it to the target computing device. Use labeled smooth cross-entropy loss, Adam optimizer and cosine annealing learning rate scheduler to carry out model training. Simultaneously record the time consumption of each stage through the time monitoring module. Optimal model weights are selected and saved based on the accuracy of the validation set.

[0041] S4. Multi-dimensional model evaluation: Use the optimal model to reason on the test set. Through the model training and evaluation module, complete the quantitative statistics of major category classification indicators and minor category classification accuracy, generate classification reports and confusion matrices, and identify the model's classification shortcomings.

[0042] S5. Model saving and iterative optimization: Package and save information such as model weights, class mapping relationships, training history, and evaluation metrics, supporting modular deployment; at the same time, newly collected fiber samples and misclassified cases can be fed back to the hierarchical class mapping module and intelligent data processing module to realize incremental updates of the mapping table and expansion of the training set, completing iterative optimization of the model.

[0043] Before the initialization of the S1 model and dataset, there are preliminary steps such as detecting the computing power of the computing device and verifying the legality of the dataset path. After the S5 model is saved and iteratively optimized, there are steps such as adjusting the network module parameters (e.g., ECA convolution kernel size, CSIM branch weights) or data augmentation strategies based on the model evaluation results, and then retraining the model. The optimized model performance data is synchronized to the statistical system of the time monitoring module to form a closed loop of "training-evaluation-optimization".

[0044] Example 1

[0045] Accuracy testing of virgin pulp fibers

[0046] I. Experimental Preparation

[0047] Virgin pulp fibers produced by a paper mill were selected as test samples. A total of 50 microscopic images of this batch of virgin pulp fibers were collected, with a uniform image resolution of 3480×2612 pixels to ensure that key information such as fiber freeness characteristics, morphology, and texture were clearly presented in the images, with no obvious acquisition noise. At the same time, a manually calibrated fiber category report for the same batch was prepared as a reference standard.

[0048] II. Experimental Procedure

[0049] Fifty images of mixed virgin pulp fibers were sequentially input into the intelligent analysis model for paper fiber classification based on the improved ResNeXt of this invention. First, image preprocessing was completed through a standardized preprocessing procedure of the intelligent data processing module (resize to 256×256, center-trimmed to 224×224, and ImageNet mean and standard deviation normalization). Then, the feature enhancement network module extracted fiber features, with the CSIM submodule capturing the freeness-related texture features and the ECA attention submodule enhancing the unique morphological features of the mixed virgin pulp fibers. Finally, the hierarchical category mapping module completed the classification from minor to major categories, outputting the classification results. The accuracy of single-sample judgment and the overall batch accuracy were statistically analyzed.

[0050] III. Experimental Results

[0051] Of the 50 virgin pulp fiber samples, the model classified all of them as virgin pulp fiber. The classification results of all samples were consistent with the manual labeling, and the classification accuracy reached 100%.

[0052] The results show that the model has extremely high accuracy in classifying virgin pulp fibers used in actual factory production, and can accurately complete the quality control judgment of virgin pulp fiber raw materials, meeting the core requirements of screening virgin pulp fiber raw materials entering the factory.

[0053] Example 2

[0054] Secondary fiber accuracy test

[0055] I. Experimental Preparation

[0056] Waste paper pulp fibers recycled from a paper mill were selected as test samples. Due to the complex origin of the recycled raw materials, the fibers contained impurities such as ink particles and fine paper scraps. A total of 50 microscopic images of the waste paper pulp fibers from this batch were collected, with an image resolution of 3480×2612 pixels. The impurity content test report from the factory laboratory and the results of manual classification were also obtained as a reference.

[0057] II. Experimental Procedure

[0058] Fifty images of waste paper pulp fibers containing impurities were input into the model. The model adaptively suppressed noise from impurities such as ink particles through the NASM submodule, then enhanced the fiber bundle morphology features of secondary fibers through the ECA attention submodule, and fused global fiber distribution and local detail features through the CSIM submodule. Finally, the hierarchical category mapping module completed the mapping from minor category labels to major secondary fiber categories, outputting the major category classification results. The model's major category classification results were compared with the manually labeled results to statistically analyze the accuracy of the major category classification.

[0059] III. Experimental Results

[0060] Of the 50 waste paper pulp fiber samples, all were classified as secondary fiber by the model. The classification results of all samples were consistent with the manual labeling, and the classification accuracy reached 100%.

[0061] The results validate the NASM module of this model's ability to effectively suppress impurities in waste paper pulp. Even in waste paper pulp scenarios with high impurity interference, it can still accurately determine the major categories of secondary fibers, making it fully adaptable to the rapid screening and classification confirmation of waste paper pulp raw materials in recycled paper mills.

[0062] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computer vision-based intelligent analysis model for paper fiber classification, characterized in that, include: The feature enhancement network module is built upon a pre-trained ResNeXt50-32x4d backbone network and includes an ECA attention submodule, a cross-scale feature interaction submodule (CSIM), a noise adaptive suppression submodule (NASM), and a dual-pooling classification head. The ECA attention submodule configures differentiated convolutional kernel sizes for different networks. The CSIM submodule extracts and fuses local and global fiber features through multi-branch parallel processing. The NASM submodule implements secondary fiber impurity noise suppression. The dual-pooling classification head outputs the fiber major category classification results. A hierarchical category mapping module stores the hierarchical mapping relationship between papermaking fiber minor categories and major categories, realizing the classification of minor categories. Automatic conversion from index to category index supports incremental mapping updates for newly added fiber categories; the intelligent data processing module includes a data preprocessing submodule and a hierarchical sampling module. The data preprocessing submodule configures multi-dimensional data augmentation strategies for the training set and performs standardized preprocessing on the validation and test sets. The hierarchical sampling module achieves balanced partitioning of samples from each fiber category; the model training and evaluation module uses a labeled smooth cross-entropy loss function, Adam optimizer, and cosine annealing learning rate scheduler to complete model transfer training, supporting quantitative evaluation of classification performance in both major and minor categories; and the time monitoring module is used to realize multi-granular time consumption statistics and visualization throughout the entire model training process.

2. The intelligent analysis model for papermaking fiber classification based on computer vision according to claim 1, characterized in that, In the feature enhancement network module, the convolution kernel size of the ECA attention submodule includes, but is not limited to, 3, 5, and 7. A 3×1 convolution kernel is configured for layer 2, a 5×1 convolution kernel is configured for layer 3, and a 7×1 convolution kernel is configured for layer 4. The convolution kernel size is designed according to the texture complexity of the paper fiber image and the resolution of the microscopic imaging.

3. The intelligent analysis model for paper fiber classification based on computer vision according to claim 1, characterized in that, The sub-category mapping table stored in the hierarchical category mapping module specifically includes the following mapping relationships: mappings between single virgin pulp fiber subcategories made from 5 common plant fiber raw materials (including hardwood, softwood, and non-wood types) and their corresponding major categories; mappings between 5 secondary fiber subcategories (including 1 type of mixed office waste paper deinked pulp and 4 types of packaging waste paper non-deinked pulp) and their corresponding major categories; and mappings between 7 mixed virgin pulp fiber subcategories with a beating degree of 30°SR-70°SR (including single or mixed types of softwood, hardwood, and bamboo pulp with different beating degrees) and their corresponding major categories. The three major categories specifically refer to the virgin pulp fiber major category, the secondary fiber major category, and the mixed virgin pulp fiber major category. Furthermore, the sub-category index-major category index mapping dictionary can be automatically synchronized and adjusted according to the updates to the sub-category-major category mapping table.

4. The intelligent analysis model for paper fiber classification based on computer vision according to claim 1, characterized in that, In the intelligent data processing module, the training set enhancement strategies of the data preprocessing submodule include, but are not limited to, random cropping (224×224, scaling range 0.8-1.0), random horizontal / vertical flipping, ±15° rotation, Gaussian blur with 0.1-2.0 standard deviation, and ±0.2 brightness / contrast adjustment. All images are normalized to the ImageNet standard mean and standard deviation. The hierarchical sampling module sets 80 images for the Vida subclass target and 200 images for the other subclasses, and divides the training / validation / test sets in an 8:1:1 ratio. If the samples are insufficient, they are supplemented by repeated sampling.

5. The intelligent analysis model for paper fiber classification based on computer vision according to claim 1, characterized in that, In the model training and evaluation module, the smoothing coefficient of the labeled smoothed cross-entropy loss function is 0.1, the Adam optimizer learning rate is 5e-5 and the weight decay is 5e-4, and the cosine annealing learning rate scheduler T_max is 140 and eta_min is 1e-6. The quantitative indicators output by the evaluation submodule include the major category classification report, confusion matrix, weighted precision / recall / F1 score, and classification accuracy from minor to major category.

6. The intelligent analysis model for papermaking fiber classification based on computer vision according to claim 1, characterized in that, The time monitoring module includes training time statistics, including but not limited to total training time, time per epoch, time per batch iteration, and average time during the training / validation phase. The time statistics enable full-dimensional monitoring of training efficiency. The output of the statistical results includes but is not limited to text reports and line charts. The visualization format is customized according to the data analysis needs of industrial deployment.

7. The intelligent analysis model for paper fiber classification based on computer vision according to claim 1, characterized in that, The construction of the intelligent analysis model includes the following steps: S1. Model initialization, configuring the device environment (prioritizing GPU), setting random seeds to ensure reproducibility of experiments, loading the subclass-major class mapping table and completing the association mapping between subclass indexes and major classes; S2. Dataset Construction and Preprocessing: Scan fiber image dataset catalog, construct a flat dataset containing image paths and subclass labels; complete stratified sampling and differential preprocessing of training / validation / test sets through intelligent data processing module, and generate data loader; S3. Feature enhancement network construction: Initialize the improved ResNeXt50 network, insert ECA, CSIM, and NASM sub-modules, freeze the shallow parameters of the backbone network, and deploy the network to the target computing device. S4. Model training: The model is trained using a labeled smooth cross-entropy loss function, Adam optimizer, and cosine annealing learning rate scheduler. The time consumption of each stage is recorded through a time monitoring module, and the optimal model weights are saved based on the accuracy of the validation set. S5. Model Evaluation: Use the optimal model to reason on the test set. The model training and evaluation module completes the dual-dimensional classification effect evaluation of major and minor categories, and generates a classification report and confusion matrix. S6. Model saving and iterative optimization: Package and save information such as model weights, class mapping relationships, training history, and evaluation metrics, and support modular deployment.

8. The intelligent analysis model for paper fiber classification based on computer vision according to claim 7, characterized in that, In step S1, before the model and dataset are initialized, there are also preliminary steps such as detecting the computing power of the computing device and verifying the legality of the dataset path.

9. The intelligent analysis model for paper fiber classification based on computer vision according to claim 7, characterized in that, In step S5, after model saving and iterative optimization, the method further includes adjusting network module parameters or data augmentation strategies based on model evaluation results, retraining the model, and synchronizing the optimized model performance data to the statistical system of the time monitoring module.