Multi-disciplinary test paper intelligent identification method based on end-cloud collaboration

Abstract

The invention discloses a multidisciplinary test paper intelligent identification method based on end-cloud collaboration, which is applied to the field of paper document classification and identification and aims at solving the problems of low paper document classification accuracy and low efficiency in the prior art. The method comprises the following steps: firstly, carrying out image acquisition on a paper document, then carrying out preprocessing on the acquired image at end side equipment, including calculating an integrogram of an input image, calculating a pixel sum of any rectangular region within constant time based on the integrogram, and dynamically calculating a binarization threshold value of each pixel point according to a local neighborhood gray feature; and then the binarized image is uploaded to a cloud for classification and identification, and finally an identification result is returned to an end side device.

Description

[0001] TECHNICAL FIELD The present application belongs to the field of paper document classification and recognition, and particularly relates to a multi-disciplinary test paper classification and recognition technology. BACKGROUND

[0002] There are mainly two types of mainstream technical solutions in the current test paper management field: one is the traditional physical folder storage solution, which manually classifies test papers according to disciplines, time, etc. and puts them into folders for manual labeling and sorting; the other is the electronic device scanning storage solution, which uses mobile phones, scanners and other devices to convert test papers into electronic pictures or documents, and stores them in local devices or the cloud. Some solutions support basic keyword retrieval. In addition, there are a small number of automatic document classification devices in the industrial field, which use fixed keyword matching or simple image recognition technology, mainly applied to office file archiving. However, these solutions have significant defects in the algorithm level, and are not optimized for the multi-disciplinary test paper scenario, making it difficult to meet the efficient and accurate classification and recognition needs.

[0003] The electronic device scanning storage solution, although solves some problems of physical storage, has significant deficiencies in the algorithm level:

[0004] (1) The adaptability of the OCR recognition module to specific content of the test paper (such as printed fonts, handwritten annotations);

[0005] (2) The processing flow does not integrate effective image preprocessing algorithms, and poor image quality directly affects the recognition effect;

[0006] The automatic classification devices in the industrial field have the following problems in the algorithm level:

[0007] (1) Using fixed keyword matching or simple image feature matching, the accuracy of discipline classification is less than 70%;

[0008] (2) The recognition model is not optimized for the content features of multi-disciplinary test papers;

[0009] (3) The classification algorithm cannot understand professional terms, formulas, question types and other discipline characteristics;

[0010] In summary, the current intelligent test paper recognition field mainly has the following technical bottlenecks, and these algorithm-level defects seriously limit the accuracy and efficiency of multi-disciplinary test paper classification:

[0011] 1. Adaptability defects of traditional OCR technology

[0012] Traditional optical character recognition (OCR) technology has insufficient recognition capability for complex layout content in test papers, especially for mathematical formulas, chemical symbols, handwritten annotations and other professional content, with an extraction accuracy generally less than 70%. The main problem is:

[0013] (1) The layout analysis algorithm based on fixed rules cannot effectively handle complex formats such as multi-column mixed layout and interlaced text and graphics;

[0014] (2) The character segmentation algorithm has poor processing effect on connected characters (such as handwritten connected characters);

[0015] (3) Lack of targeted optimization of discipline-specific terminology and special symbols.

[0016] 2. Limitations of image binarization algorithm

[0017] The prior art mostly uses global thresholding method for image binarization, which performs poorly under uneven lighting conditions:

[0018] (1) A single threshold value cannot adapt to the lighting changes of the entire image, resulting in a text extraction completeness rate of only 70-80%;

[0019] (2) Interference factors such as shadows and reflections can cause text breakage or background noise amplification;

[0020] (3) Lack of local adaptive mechanism, poor robustness to actual situations such as test paper wrinkles and stains.

[0021] 3. Deficiencies of single modal classification model

[0022] Traditional methods rely on single text modal or image modal for classification, which has obvious limitations:

[0023] (1) Pure text method cannot utilize the layout features of the test paper (such as formula position, chart layout);

[0024] (2) Pure image method is difficult to understand semantic content, and lacks depth of understanding of text meaning;

[0025] (3) Lack of multi-modal feature fusion mechanism, unable to simultaneously utilize text content and visual features. SUMMARY

[0026] In view of the defects of the prior art, the present application constructs a full-process solution of "end-side preprocessing + cloud-side deep learning recognition + mechanical automation execution + mobile terminal interaction", and the core improvements are around four dimensions of intelligent recognition algorithm, end-cloud collaborative architecture, modular mechanical structure and human-computer interaction system, realizing full-automatic management of test papers from physical original to electronic data.

[0027] The technical scheme adopted by the present application is: a multi-disciplinary test paper intelligent recognition method based on end-cloud collaboration, comprising:

[0028] S1, image acquisition: acquiring original test paper images through a camera;

[0029] S2, end-side device preprocessing:

[0030] Calculate the integral image of each pixel, calculate the local gray-scale mean based on the integral images of each pixel in the neighborhood of the current pixel, calculate the adaptive threshold of the current pixel based on the local gray-scale mean, and perform binarization processing on each pixel based on the adaptive threshold corresponding to each pixel, finally obtaining the binarized image;

[0031] S3. Data Upload: Upload the preprocessed binarized image to the cloud server;

[0032] S4, Cloud-based Recognition: Performs multi-level classification and recognition on a cloud server;

[0033] S5. Result Return: Return the classification results to the end device.

[0034] The beneficial effects of this invention are as follows: This invention provides a multi-disciplinary intelligent test paper recognition method based on edge-cloud collaboration. It optimizes image quality through edge-side local adaptive threshold preprocessing. Specifically, the edge-side local adaptive threshold algorithm dynamically adjusts the threshold based on the local neighborhood characteristics of each pixel. Under uneven lighting conditions, compared to traditional methods where the text extraction completeness rate is only 70-80%, the dynamic adaptive threshold of this invention achieves a text extraction completeness rate of over 94% under the same lighting conditions, providing high-quality binarized images for cloud processing. The cloud adopts a three-level multimodal fusion classification architecture to achieve high-precision subject recognition. After edge-side preprocessing, the text recognition accuracy of cloud OCR increases from 85% to 98%. The method of this invention has the following advantages:

[0035] 1) End-side local adaptive threshold preprocessing is used to solve the problem of uneven illumination;

[0036] 2) Construct a multimodal fusion classification architecture that utilizes both textual and visual features;

[0037] 3) Design a three-level classification pipeline to achieve subject identification from coarse to fine;

[0038] 4) Feature fusion based on attention mechanism enhances the model's understanding ability. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention.

[0040] Figure 2 A cloud-based identification flowchart provided for an embodiment of the present invention. Detailed Implementation

[0041] like Figure 1 As shown, the method of the present invention includes the following steps:

[0042] S1. Image Acquisition: The original image of the test paper is acquired through the OV5640 camera, with an image resolution of 2592×1944 pixels;

[0043] S2. End-side preprocessing: Perform image preprocessing based on local adaptive thresholding on the Renesas RA8D1 development board, where the local window size is 15×15 pixels, the sensitivity coefficient is set to 0.8, and the constant offset is 10.

[0044] S3. Data Upload: The pre-processed images are uploaded to the cloud server via the RW007 WiFi module using the HTTP protocol, with the transmission time for a single image controlled within 200ms.

[0045] S4. Cloud-based recognition: Performs multi-level classification and recognition on Alibaba Cloud servers, including OCR text extraction, JIEBA word segmentation, and DeepSeek API classification;

[0046] S5. Result Return: Return the classification results to the end device with an end-to-end delay of no more than 1.2 seconds.

[0047] In step S1, the end-side device uses the Renesas RA8D1 development board as the main control unit to replace the expensive industrial controller. It integrates an OV5640 camera (image acquisition), an RW007 WiFi module (data transmission), and an infrared sensor (position detection) to build a lightweight end-side processing unit.

[0048] Users place the test paper into the device's input terminal, complete the binding process by entering the device's unique number through the mini-program, and send classification instructions.

[0049] After receiving the command, the Renesas RA8D1 development board controls the LED spotlight to turn on and the OV5640 camera to capture the image of the exam paper.

[0050] The preprocessing steps in step S2 specifically include:

[0051] The local grayscale mean is calculated using the integral image principle. The integral image allows the calculation of the pixel sum of any rectangular region in constant time: the pixel sum Σ of a 15×15 neighborhood centered on the current pixel (x,y) can be calculated using the integral image values ​​of the four vertices (Σ = I(x+7,y+7) - I(x-8,y+7) - I(x+7,y-8) + I(x-8,y-8)). Dividing by the region area (15×15) instead of the number of vertices (4) is because the integral image calculates the sum of all pixel values ​​within the region, and the mean needs to be normalized based on the total number of pixels to ensure accurate grayscale estimation.

[0052] An adaptive threshold is generated by linearly transforming the local grayscale mean using a sensitivity coefficient of 0.8 and a constant offset of 10. Binarization is then performed pixel-by-pixel based on this adaptive threshold. Based on extensive experimental data, this embodiment demonstrates that a sensitivity coefficient of 0.8 and a constant offset of 10 result in the highest text extraction completeness and lowest noise interference under most lighting conditions. In practical applications, the values ​​of the sensitivity coefficient and constant offset can be adjusted according to the specific lighting conditions of the application scenario.

[0053] The adaptive threshold calculation process for pixel (x, y) is as follows:

[0054] First, the sum of pixels in the neighborhood is quickly calculated using an integral image: Σ = I(x+7,y+7) - I(x-8,y+7) - I(x+7,y-8) + I(x-8,y-8); where I(x+7,y+7), I(x-8,y+7), I(x+7,y-8), and I(x-8,y-8) represent the integral image values ​​of the pixels at the four vertices (x+7,y+7), (x-8,y+7), (x+7,y-8), and (x-8,y-8) in a 15×15 neighborhood centered on the current pixel (x, y). Traditional methods require calculating every pixel in the neighborhood, with a computational complexity of O(n²), while this invention only requires four addition and subtraction operations to complete the calculation, with a complexity of O(1). Compared to traditional calculation methods, this invention has higher processing efficiency, increasing the calculation speed by approximately 50 times.

[0055] Those skilled in the art should note that this invention only requires calculating the sum of the integral images of the pixels at the four vertices of a 15×15 neighborhood. For image boundary pixels, a mirror filling method is used to ensure that a complete processing window can be formed. Specifically, with the current pixel as the center, the missing neighboring pixels constituting the processing window are filled using the integral images of other neighboring pixels symmetrical to the current pixel. Furthermore, the processing window size of the boundary region pixels is dynamically adjusted to ensure that at least a 5×5 effective neighborhood is included. The window size adjustment formula is: window_size = max(5, min(15, available neighborhood size)).

[0056] Then, based on the previous step, the pixels in the neighborhood and Σ are calculated, and the local mean is calculated as Σ / (15×15).

[0057] Finally, based on the sensitivity coefficient and constant offset, the adaptive threshold is calculated as: Local Mean × 0.8 + 10.

[0058] To address the problem of character segmentation failure caused by cursive handwriting in existing technologies, this invention performs morphological closing operations on the binarized results using 3×3 circular structural elements to fill in the discontinuities of character strokes.

[0059] Step S2 also includes image compression processing: the binarized image is compressed using optimized run-length encoding, with 2 bytes used to store the run-length and 1 byte to store the pixel value; the compressed image data size is reduced by 25% compared to the standard 2-byte length and 2-byte pixel value scheme, and a single test paper image is compressed from the original approximately 15MB to approximately 300KB, and the transmission time is shortened from 3 seconds to 200ms; the compression algorithm processing speed reaches 30 frames of 1920×1080 resolution images per second.

[0060] like Figure 2 As shown, the specific processing procedure of step S4 is as follows:

[0061] The preprocessed image is uploaded to the Alibaba Cloud server via the RW007 WiFi module for multimodal feature fusion. The multimodal features include text features and image features. The text feature extraction process is as follows: OCR text extraction, JIEBA word segmentation, and DeepSeek API classification are performed sequentially in the cloud to generate subject classification results (≤1s). The image feature extraction process is as follows: the preprocessed image is input into a CNN network to extract the image features; then, an attention mechanism is used to fuse the text features and image features.

[0062] The cloud-based identification step adopts a three-level classification architecture:

[0063] The first level of classification uses the DeepSeek API for preliminary subject category classification. The specific process is as follows: First, all text content is extracted from the entire test paper image using OCR text extraction; then, the regional positions of the title and the first paragraph are located using a layout analysis algorithm; next, all extracted text content is processed by JIEBA word segmentation, and the title and the first paragraph are used as key information and input together with the JIEBA word segmentation results into the DeepSeek API for preliminary subject category classification.

[0064] The second-level classification uses the BERT-based model for fine-grained semantic analysis; it achieves subject category recognition with an accuracy of 99.2% and a processing time of 200ms; specifically:

[0065] 1. Calculate term density:

[0066] 11. Obtaining the number of occurrences of terms:

[0067] During the JIEBA word segmentation process, the actual number of times specific subject terms appear in the test paper text is counted by using a professional dictionary in the field of education (containing 50,000+ subject-specific words); for example, if "derivative" appears 5 times and "integral" appears 3 times in a math test paper, then the number of times the mathematical terms appear is 8.

[0068] 12. Calculation of total text length:

[0069] Total text length refers to the total number of words in the test paper (based on the result after word segmentation); for example, if a test paper has 1500 words after word segmentation, then the total text length is 1500.

[0070] 13. Formula for calculating term density:

[0071] Term density = Number of terms in a specific subject / Total text length; Example: If mathematical terms appear 80 times and the total text length is 2000 words, then mathematical term density = 80 / 2000 = 0.04.

[0072] The term density in this invention is used to eliminate the influence of text length differences and make features more comparable.

[0073] 2. Terminology Density Vector: The terminology density vector is a multi-dimensional feature vector composed of terminology density values ​​from multiple disciplines. For example, the terminology density is calculated for each preset subject category (mathematics, physics, chemistry, Chinese, English, biology, etc.); resulting in a terminology density vector = [mathematics terminology density, physics terminology density, chemistry terminology density, ..., biology terminology density]. For example, the terminology density vector of a test paper might be [0.04, 0.02, 0.01, 0.03, 0.02, 0.005]. In this embodiment, the dimension of the terminology vector is consistent with the number of preset subject categories.

[0074] The term density vector is input as a numerical feature into the deep learning model; by analyzing the distribution pattern of term density in each discipline, the model is assisted in making discipline classification decisions; for example, when the density of mathematical terms is significantly higher than that of other disciplines, it strongly indicates that it is a mathematics exam paper.

[0075] This multidimensional feature representation method enables the model to capture the subject-specific tendencies of the test papers, thereby improving classification accuracy.

[0076] This embodiment specifically uses the improved Sentence-BERT model, which is an improved version of the BERT-base model. The term density vector and the output of the DeepSeek API are input into the improved Sentence-BERT model for fine-grained semantic analysis. That is, the standardized text data after the initial subject classification by the DeepSeek API is mapped to a 768-dimensional semantic space. The term density vector is first transformed through a fully connected layer, and then concatenated and fused with the 768-dimensional text features output by DeepSeek. The concatenated feature dimension is: 768 dimensions (text features) + n dimensions (dimensionality after term density feature transformation).

[0077] Feature fusion: The contribution weights of text features and term density features are balanced through an attention mechanism; through feature fusion, the numerical characteristics of term density are preserved, while the semantic information of the text is combined.

[0078] Cosine similarity is used for semantic matching, with a similarity threshold set to 0.75; the final output is a semantic feature vector, i.e., text features.

[0079] The third-level classification combines TF-IDF weighting and SVM classifiers for knowledge module identification, with a feature dimension of 5000. The result is a fine-grained knowledge module classification, such as:

[0080] (1) Characteristics of the mathematics test paper: including features of algebra, geometry, probability, etc.;

[0081] (2) Characteristics of the physics exam paper: including features of mechanics, electromagnetism, optics, etc.;

[0082] (3) Characteristics of the chemistry exam paper: including chemical reaction equations, chemical formulas, experimental apparatus, etc.;

[0083] (4) Characteristics of Chinese language test papers: including the structure of classical Chinese poetry and prose, rhetorical devices, and essay genres;

[0084] (5) Characteristics of English test papers: including grammatical structure, reading comprehension question types, writing patterns, etc.;

[0085] (6) Features of the biology test paper: including biological classification, physiological processes, experimental diagrams, etc.

[0086] The third level of classification also includes identifying the structural features of question types such as multiple choice, fill-in-the-blank, and problem-solving questions through layout analysis algorithms.

[0087] Traditional layout analysis algorithms include projection-based methods, connected component-based methods, and machine learning-based methods; for example, the projection and watershed algorithm proposed by Xiao et al. in "Document Image Layout Analysis" (2009); and the deep learning-based instance segmentation method proposed by He et al. in "Mask R-CNN" (2017).

[0088] This invention optimizes general layout analysis algorithms for the characteristics of exam documents, focusing on special layout elements such as formula areas, table areas, and handwritten annotation areas; it combines multi-scale feature fusion and attention mechanisms to improve adaptability to complex layouts; specifically, it uses an improved CNN network (such as ResNet-50 backbone + FPN feature pyramid) for end-to-end layout analysis.

[0089] Compared to existing technologies, tests on public datasets (such as PubLayNet) show that the method of this invention improves the accuracy of test paper format analysis by more than 15%.

[0090] Although the layout analysis algorithm in this invention is based on existing technology, it achieves better practical results through optimization and innovation tailored to the characteristics of test papers.

[0091] like Figure 2 As shown, the input to the CNN network is a binarized image that has undergone edge preprocessing; the CNN network extracts visual features of the image, such as formula symbols and chart structures, thereby obtaining image features; this solves the problem that traditional regular layout analysis cannot handle multi-column mixed layouts.

[0092] After concatenating the formula and question type features as independent feature vectors with the text features output by the improved Sentence-BERT model, and combining TF-IDF weighting and SVM classifiers, fine-grained knowledge module classification is obtained.

[0093] By fusing text and image features at various levels through an attention mechanism, the number of attention heads is 12; multimodal fusion enables the model to understand both text content and visual layout simultaneously; compared with unimodal methods, the classification accuracy is improved by 8-10%.

[0094] Based on the identification method of this invention, semantic vectorization can also be realized based on the Sentence-Transformers framework and m3e-base model to realize the search function; it supports multi-dimensional combined retrieval, including four dimensions: subject, time, question type and difficulty; the retrieval accuracy reaches 98% and the average response time is no more than 3 seconds.

[0095] The search function can be further improved by following these steps:

[0096] The BM25 algorithm was used to reorder the preliminary search results, with parameters k1=1.5 and b=0.75.

[0097] Personalized recommendations based on users' historical behavior data improve recommendation accuracy by 15%.

[0098] It supports semantic association retrieval, with a retrieval recall rate of 97.5%.

[0099] During the training process, this embodiment collected at least 5,000 multi-disciplinary paper test paper samples, covering major subject areas such as Chinese, mathematics, English, physics, and chemistry. The collected paper test papers were sequentially image acquired and preprocessed on the edge. Known annotation information, including subject classification, knowledge point tags, and question type information, was added to each binarized image obtained from the edge preprocessing. The dataset was divided into training set, validation set, and test set in a 14:3:3 ratio.

[0100] The specific training process is divided into three stages:

[0101] The first stage involved pre-training. The OCR module was pre-trained on over 5000 test paper images, achieving a text recognition accuracy of 98.5%. The JIEBA word segmentation module was optimized for the education field, adding over 50,000 subject-specific terms to address the lack of optimized terminology in existing technologies. The DeepSeek model underwent transfer learning in subject classification tasks. Through the establishment of a multi-disciplinary terminology database, the system can accurately identify the characteristics of specialized vocabulary from different disciplines.

[0102] The second phase is end-to-end joint training, where all modules are cascaded for end-to-end training, optimizing processes such as... Figure 2 The overall recognition process is shown below. FP32 is used to represent the binarized image to ensure training stability. The loss function during training is Focal Loss + cross-entropy loss. The AdamW optimizer is used with a learning rate of 3e-5 and weight decay of 0.01. Early stopping and validation set monitoring are employed; training stops if the validation set loss shows no improvement after 10 consecutive epochs. Gradient checkpointing is used, reducing memory usage during training by 60%. The loss function is Focal Loss + cross-entropy loss.

[0103] The third stage involves quantization deployment optimization, converting the trained FP32 precision model to INT8 format for recognition processing of the binarized images of the test papers to be processed. This balances accuracy and efficiency, reducing the model size by 75% and increasing inference speed by 3 times. Specifically, the third stage involves randomly selecting 500 images from the training set, performing symmetrical quantization, and calibrating each layer individually; the accuracy loss after quantization is less than 0.5%.

[0104] A dynamic loading mechanism was implemented, reducing model loading time from 5 seconds to 1 second.

[0105] As shown in Table 1, the contribution of each technical component was verified through ablation experiments, and it can be seen that the classification accuracy of the complete technical solution of the present invention reached 98.1%.

[0106] Table 1. Contribution of each technical component of the present invention

[0107]

[0108] On 5000 test images, the local adaptive thresholding method of this invention improved the text extraction integrity rate by 17.4% compared to the global thresholding method's 76.8%, achieving a text extraction integrity rate of 94.2%. Furthermore, the processing speed was improved from 187ms / image to 450ms / image.

[0109] This embodiment compares the method of the present invention with traditional keyword matching methods and single-modal deep learning methods. The test subjects are 100 mixed subject test papers without added annotation information, including: 30 mathematics papers, 30 physics papers, 20 chemistry papers, and 20 Chinese papers. The test results are shown in Table 2. The recognition accuracy of the method of the present invention is significantly better than that of traditional methods, while maintaining high processing efficiency, supporting 100+ concurrent requests, and keeping the CPU utilization rate below 70%.

[0110] Table 2. Comparison of the effects of the method of the present invention and the prior art.

[0111]

[0112] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A multi-disciplinary test paper intelligent identification method based on end-cloud cooperation, characterized in that, Comprise: S1, image acquisition: collect the original image of the test paper through the camera; S2, end side device pretreatment: Calculate the integral image of each pixel point in the original image of the test paper, calculate the local gray mean value according to the integral image of each pixel point in the neighborhood of the current pixel point, calculate the adaptive threshold value of the current pixel point based on the local gray mean value, and perform binaryzation processing on each pixel point based on the adaptive threshold value corresponding to each pixel point, and finally obtain the binaryzation image; S3, data upload: upload the binaryzation image after pretreatment to the cloud server; S4, cloud recognition: perform multi-level classification recognition on the cloud server; S5, result return: return the classification result to the end side device. 2.The multi-discipline test paper intelligent recognition method based on end-cloud cooperation according to claim 1, characterized in that, The integral image of each pixel point in step S2 is the sum of the pixel values from the first pixel point to the current pixel point in the collected original image of the test paper. 3.The multi-discipline test paper intelligent identification method based on end-cloud collaboration according to claim 2, characterized in that, In step S2, the local gray mean value is calculated according to the integral image of each pixel point in the neighborhood of the current pixel point, which is specifically to sum the integral images of the four vertex position pixel points of the local rectangular region centered on the current pixel point, and then divide by the total number of pixel points in the local rectangular region to obtain the average, thereby obtaining the local gray mean value.

4. The multi-disciplinary test paper intelligent identification method based on end-cloud collaboration according to claim 3, characterized in that, The calculation process of the adaptive threshold value in step S2 is to multiply the local gray mean value corresponding to the current pixel point by the product of the sensitivity coefficient and add the constant offset, and the result obtained is the adaptive threshold value corresponding to the current pixel point.

5. The multi-disciplinary test paper intelligent identification method based on end-cloud collaboration according to claim 4, characterized in that, The sensitivity coefficient is 0.8, and the constant offset is 10.

6. The multi-disciplinary test paper intelligent identification method based on end-cloud collaboration according to claim 5, characterized in that, Step S2 also includes performing morphological closing operation of 3*3 circular structure element on the binaryzation image to fill the text stroke discontinuity.

7. The multi-disciplinary test paper intelligent recognition method based on end-cloud collaboration according to claim 4 or 6, characterized in that, It also includes using optimized run-length encoding to compress the binaryzation image, specifically using 2 bytes to store the run-length and 1 byte to store the pixel value.

8. The multi-disciplinary test paper intelligent identification method based on end-cloud collaboration according to claim 7, characterized in that, Step S4 includes three levels of classification: The first level classification process is: first, extract all text content from the entire test paper image through OCR text extraction; Then locate the area position of the title and the first paragraph of text through the layout analysis algorithm; Then, the JIEBA word segmentation processing is performed on all the extracted text content, and the title and the first paragraph of text are input into the DeepSeek API together with the JIEBA word segmentation result for preliminary classification of the subject category; The second level classification process is: in the JIEBA word segmentation process, the actual number of specific subject terms appearing in the test paper text is counted through the professional dictionary in the education field; Based on the number of subject terms appearing in the test paper text and the length of the test paper text, the term density vector is composed of the term density values of multiple subjects; Map the normalized text data after preliminary subject category classification by DeepSeek API to a 768-dimensional semantic space to obtain a 768-dimensional text feature; The term density vector is first transformed by a fully connected layer, and then spliced and fused with the 768-dimensional text feature output by DeepSeek; The spliced and fused result is input into the improved Sentence-BERT model to obtain the final text feature; The third-level classification process is: identifying the question type structure characteristics through the layout analysis algorithm, splicing the question type structure characteristics as an independent feature vector with the text features output by the improved Sentence-BERT model, combining TF-IDF weighting and a SVM classifier, and obtaining the fine-grained knowledge module classification result.

9. The multi-disciplinary test paper intelligent identification method based on end-cloud collaboration according to claim 8, characterized in that, The length of the test paper text is determined based on the number of words obtained after the test paper is segmented by JIEBA.

Images