An ancient book non-destructive page turning scanning and repairing integrated method and system

Abstract

This application discloses an integrated method and system for non-destructive page-turning scanning and restoration of ancient books, including: collecting the physical characteristics and environmental parameters of the ancient books, and outputting benchmark operating parameters for page turning, scanning, and correction; dynamically adjusting the adsorption negative pressure to complete the non-destructive separation of pages using a negative pressure-vibration-airflow-electrostatic neutralization coupling mechanism; synchronously and adaptively completing high-fidelity scanning of fragile area data, correcting spine distortion according to the distortion value, and feeding back the correction error to optimize page-turning parameters; using a CRNN model to recognize text, combining an authoritative ancient book corpus with contextual semantics to complete missing text, and adaptively calibrating the recognition results according to the layout; and updating the CRNN model and page-turning parameters through incremental learning to achieve closed-loop optimization of the CRNN model. This application achieves intelligent obstacle avoidance in fragile areas through "tactile perception" of pressure sensing, and combines dynamic negative pressure with humidity and fiber strength as dual factors with a multi-field real-time coupling mechanism to achieve "gentle page turning at the level of a cultural relic restorer".

Description

Technical Field

[0001] This application belongs to the field of ancient book digitization preservation technology, specifically involving an integrated method and system for non-destructive page-turning scanning and restoration of ancient books. Background Technology

[0002] With the deepening of the Chinese Ancient Books Preservation Project, the digitization of ancient books has become a core means to achieve the permanent preservation and revitalization of China's outstanding traditional cultural heritage. Domestic institutions have completed the digitization of millions of ancient books, and related technical solutions and equipment have been continuously iterated. At present, the mainstream ancient book digitization solutions mostly focus on the isolated optimization of single modules such as page turning, scanning, correction, and restoration, or simple function superposition. They have not yet formed a dynamic collaborative system covering the entire chain of "perception-page turning-scanning-restoration-optimization". In the actual digitization of precious ancient books in the collection (especially those from the Song and Yuan dynasties that are severely brittle), there are still many core technical defects.

[0003] 1. Insufficient page-turning safety. Existing page-turning devices mostly use adhesive or mechanical separation methods. For example, patent CN208857592U discloses a page-turning device suitable for deacidification and reinforcement of ancient books, which uses a combination of vacuum suction cups and adhesive suction cups for page turning. However, the adhesive parts are in direct contact with the pages, which poses a risk of damaging the paper fibers of the brittle ancient books. Patent CN106042696A discloses a page-turning device that also uses an adhesive method and does not have an adaptive adjustment mechanism for the adsorption force of different paper characteristics.

[0004] 2. Data is severely fragmented across different work modules. Existing technologies generally suffer from the problem of independent "page turning-scanning-repairing" modules. For example, the image distortion correction methods disclosed in patents CN119359605A and CN117333374B only focus on the correction of scanned images, and the ancient text recognition method disclosed in patent CN118038467A only targets the OCR repair stage. Data is not shared between modules, and control is not synchronized.

[0005] 3. The algorithms have high computational requirements and are difficult to adapt to edge embedded platforms. Some existing studies have attempted to use machine learning for the digitization of ancient books, but they have not been optimized for lightweight edge devices and are difficult to adapt to embedded platforms such as STM32 and ARM.

[0006] 4. Insufficient robustness of OCR repair. For example, the method disclosed in patent CN118038467A can identify missing characters and generate a candidate set, but it is mainly for repairing single-character missing characters. It lacks a mechanism to verify the semantic coherence of the entire sentence context and does not introduce authoritative ancient text corpora for verification.

[0007] 5. The system lacks continuous adaptive evolution capabilities. Once the existing ancient book digitization equipment is deployed, the core operational parameters are basically fixed, making it impossible to continuously optimize the model and strategies through the accumulation of data from daily operations.

[0008] Therefore, there is an urgent need for a robust integrated system for digitizing ancient books that can dynamically sense the state of the books, intelligently adjust acquisition and restoration strategies, and achieve end-to-end collaborative optimization. This invention proposes a systematic innovation centered on multi-field collaborative page turning and graded distortion correction, combined with a lightweight parameter dynamic matching model, to solve the aforementioned problems. Summary of the Invention

[0009] This application provides an integrated method and system for non-destructive page-turning scanning and restoration of ancient books, in order to solve the technical problem that existing mainstream solutions focus on optimizing single modules or simply adding functions without forming a dynamic collaborative system across the entire chain.

[0010] To solve the above-mentioned technical problems, this application adopts the following technical solution: a method for integrated non-destructive page-turning scanning and restoration of ancient books, comprising:

[0011] S1. Collect the physical characteristics and environmental parameters of ancient books, and output the baseline operation parameters for page turning, scanning and correction through a lightweight XGBoost model;

[0012] S2. Based on the benchmark operating parameters, the suction cup-integrated MEMS pressure sensor array is used to avoid vulnerable areas and dynamically adjust the adsorption negative pressure. The negative pressure-vibration-airflow-electrostatic neutralization coupling mechanism is used to complete the non-destructive separation of the pages and perform page-turning penetration detection.

[0013] S3. Synchronize vulnerable area data to adaptively complete high-fidelity scanning, correct spine distortion according to the distortion value, and feed back the correction error to optimize page turning parameters;

[0014] S4. The CRNN model is used to recognize text, and the missing text is completed by combining an authoritative ancient book corpus and contextual semantics. The recognition results are adaptively calibrated according to the format.

[0015] S5. Collect full-process operation data, and update the CRNN model and page turning parameters through incremental learning to achieve closed-loop optimization of the CRNN model.

[0016] Furthermore, in step S1, the lightweight XGBoost model takes the ancient book format, fiber strength, aging degree, temperature and humidity as inputs, and trains 6 lightweight XGBoost sub-models according to the era and ancient book format, and outputs the baseline parameters only once during material loading.

[0017] Furthermore, the first threshold for the size of ancient books is 260mm × 185mm; if the size of an ancient book is smaller than the first threshold, it is a small-format book; if the size of an ancient book is greater than or equal to the first threshold, it is a large-format book.

[0018] Furthermore, the suction cup integrates a 4×4 MEMS pressure sensor array, which automatically adjusts by 1-3mm when the force exceeds the safety threshold, thus avoiding vulnerable areas.

[0019] Furthermore, in step S2,

[0020] Based on formula (1), dynamic negative pressure adjustment is performed; where formula (1) is:

[0021] (1);

[0022] in, To absorb negative pressure in real time; For safety factor; The tensile strength of the paper fibers being tested; This refers to the humidity adjustment coefficient. Real-time ambient relative humidity, The standard reference humidity is 50%RH.

[0023] Furthermore, in step S2, the method for realizing the negative pressure-vibration-airflow-electrostatic neutralization coupling mechanism includes:

[0024] Based on real-time force feedback, paper separation state, and ambient humidity as input variables, a multi-parameter coupled linkage model is established by real-time closed-loop adjustment of coupled control parameters including negative pressure, vibration, airflow, and electrostatic neutralization parameters.

[0025] Based on formulas (2)-(3), the micro-vibration parameters and airflow parameters are obtained; where formulas (2)-(3) are:

[0026] (2);

[0027] (3);

[0028] in, For micro-vibration displacement function, For vibration amplitude and vibration frequency The frequency is dynamically adjusted within the 50-100Hz range based on the paper fiber strength and adsorption negative pressure. Time, in seconds; air density; This is the air pressure difference; This represents the airflow velocity.

[0029] Furthermore, the method for implementing graded distortion collaborative correction in step S3 is as follows:

[0030] S31. Extract semantic features of the spine based on the MobileNetV2 network, segment the spine distortion region and calculate the distortion variation value;

[0031] S32. Based on distorted values Implement graded correction, the formula is:

[0032] (4);

[0033] in, Let be the affine transformation matrix. It is a 3D surface reconstruction algorithm, corresponding to severe distortion correction; the pixel error after correction is controlled within 1 pixel.

[0034] Furthermore, the method in step S4 includes:

[0035] S41. It adopts a lightweight architecture that combines CRNN model text recognition, ViT feature encoding, lightweight BiLSTM semantic encoding and RAG retrieval, and enhancement engine to replace the large parameter generative language model, thereby reducing computing power requirements and inference latency.

[0036] S42. A pre-built authoritative corpus of historical classics is constructed, including a full-text database of ancient books, a high-frequency word database, a variant character-orthodox character mapping database, a taboo character database, and a dedicated corpus of annotations for ancient books;

[0037] S43. Based on formulas (5)-(6), complete the missing text; where formulas (5)-(6) are:

[0038] (5);

[0039] (6);

[0040] in, For contextual semantic sequence, Here, K represents the feature vector of the incomplete text, and K represents the matching result retrieved from the classical text corpus. It is a lightweight semantic coding model. To provide the optimal text completion;

[0041] S44. Based on the completed incomplete text, set differentiated semantic thresholds according to the main text format, annotation format, and chart / figure annotation format, and obtain adaptive calibration recognition results.

[0042] Furthermore, in step S5, valid samples are defined as those with a job success rate ≥ 95% and a correction error ≤ 1 pixel. The iCaRL incremental learning algorithm is used to update the lightweight XGBoost sub-model, coupling control parameters, and semantic threshold.

[0043] One technical solution adopted in this application is: an integrated system for non-destructive page-turning scanning and restoration of ancient books, comprising:

[0044] The multi-dimensional state perception module is used to collect physical properties, environmental parameters, distribution of damaged areas, and real-time stress data of ancient books;

[0045] A multi-field collaborative lossless page-turning module is used to achieve contactless separation of pages and intelligent avoidance of vulnerable areas;

[0046] High-fidelity scanning and distortion correction module for page imaging and distortion correction;

[0047] A dual-library driven OCR intelligent repair module is used for text recognition, incomplete character recognition, and semantic calibration.

[0048] The edge computing and adaptive optimization module is used for parameter matching, algorithm iteration and module optimization; each module realizes bidirectional data interaction through the industrial bus, forming a closed-loop collaborative system of "perception-decision-execution-feedback-optimization".

[0049] The beneficial effects of this application are as follows: This application achieves intelligent obstacle avoidance in vulnerable areas through the "tactile perception" of MEMS pressure sensing, and combines the dynamic negative pressure of humidity and fiber strength with a multi-field real-time coupling mechanism to achieve "gentle page turning at the level of a cultural relic restorer". Through the lightweight XGBoost sub-model and the RAG lightweight OCR architecture, computational power consumption is reduced; at the same time, the dynamic parameter adaptive capability can seamlessly adapt to various binding formats such as thread binding, wrapped binding, and accordion binding, as well as ancient books of different ages and degrees of aging, without the need for manual adjustment; through the dual-drive architecture of "authoritative classics database and experience knowledge base", combined with the page layout adaptive dynamic threshold, the accuracy of incomplete text completion in ancient books and the overall text recognition accuracy are improved, and the recognition adaptability of complex formats such as annotations and charts is also greatly improved. Simultaneously, knowledge accumulation and reuse in difficult scenarios are achieved, and the system becomes more accurate with use. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating an embodiment of the integrated method for non-destructive page-turning scanning and restoration of ancient books according to this application;

[0051] Figure 2 This is a schematic diagram of the module structure of an embodiment of the integrated system for non-destructive page turning, scanning and restoration of ancient books in this application;

[0052] Figure 3This is a schematic diagram of an embodiment of the integrated system for non-destructive page turning, scanning and restoration of ancient books in this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0054] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0055] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the integrated method for non-destructive page-turning scanning and restoration of ancient books according to this application. The method includes:

[0056] Step S1. Multi-dimensional state perception and benchmark parameter matching: Based on visual recognition and sensor detection, acquire basic data on ancient book format, binding style, fiber tensile strength, distribution of damaged areas, and environmental temperature and humidity to construct a database of ancient book physical characteristics; through a lightweight XGBoost parameter mapping model, integrate ancient book characteristics and environmental parameters to output benchmark operation parameters for page turning, scanning, and correction.

[0057] Specifically, the method for obtaining a lightweight XGBoost model is as follows:

[0058] (1) Simplified input features: Only five core input features are retained: book type, fiber tensile strength, aging degree of ancient books, ambient temperature, and ambient humidity, while redundant dimensions are eliminated;

[0059] (2) Classification sub-model construction: Combining the common format convention of thread-bound books in the domestic ancient book preservation industry, and the adaptation requirements of the page turning operation of this patented equipment, the ancient book finished page size of 260mm×185mm (the standard 16-page format of thread-bound books) is used as the dividing line to divide large and small formats; according to the ancient book era and format type, for the three categories of ancient books of the Song and Yuan period, Ming and Qing period, and modern times, large and small format specifications are matched respectively, and a total of 6 lightweight sub-models are pre-trained to form a total of 6 models, with a single model size ≤2.5MB and an edge device inference time ≤100ms;

[0060] (3) Optimization of reasoning logic: The sub-model reasoning is only performed once in the ancient book loading process to output the baseline operation parameters for the entire process. During the operation, the parameters are fine-tuned through real-time sensor data, without the need for repeated reasoning.

[0061] (4) Parameter matching is achieved through the following formula:

[0062] ;

[0063] in, This is the optimal set of baseline operation parameters for the entire process. For book format, As the ancient books age, The ambient temperature is expressed in degrees Celsius (°C). The relative humidity is expressed as %RH. () represents the lightweight XGBoost sub-model for the corresponding scenario; compared to traditional PID fixed parameter control, it can achieve global adaptive matching of operating parameters under different ancient book characteristics and different environments, without the need for continuous manual debugging.

[0064] The specific method for determining the format of ancient books is as follows: using the physical external dimensions of the finished pages of the ancient book as the measurement object, and combining this with the common format conventions in the thread-bound book industry, a format determination formula is formulated:

[0065] ;

[0066] in, For book format, , These are the length and width of a page from an ancient book, both in mm.

[0067] Step S2. Multi-field real-time coupling non-destructive page turning: Based on reference parameters, the suction cup-integrated MEMS pressure sensor array avoids vulnerable areas of the adsorption point. Combined with the paper fiber strength and ambient humidity, the adsorption negative pressure is dynamically adjusted to establish a real-time coupling and adjustment mechanism of negative pressure-vibration-airflow-static neutralization, completing non-contact separation of the book pages. Simultaneously, the page turning penetration detection is completed by optical flow method.

[0068] Specifically, the implementation method of MEMS pressure sensing array and intelligent avoidance of vulnerable areas is as follows:

[0069] (1) A 4×4 MEMS piezoresistive pressure sensor array is integrated on the surface of a single soft silicone vacuum suction cup, with a force resolution ≤0.01N and a sampling frequency ≥100Hz;

[0070] (2) Before adsorption, the insect-infested and flocculated damaged areas are initially located by visual identification, and the initial adsorption points are planned;

[0071] (3) When the suction cup contacts the paper, the pressure sensor array collects the force distribution data in real time. If the force on any sensor unit exceeds the safe force threshold of the corresponding ancient book (embrittled ancient book ≤ 0.02N), the suction cup position is automatically adjusted by 1-3mm until the adsorption force is evenly distributed in the area of ​​healthy paper without damage, completely avoiding the vulnerable points.

[0072] The dynamic negative pressure adjustment, which combines fiber strength and ambient humidity, is achieved through the following formula:

[0073] (1);

[0074] in, The negative pressure is adsorbed in real time, and the unit is megapascal (MPa). For safety factor; The tensile strength of the paper fibers being tested is expressed in MPa. This is the humidity adjustment coefficient, taken as 0.1 in dry environments and 0.2 in humid environments; Real-time ambient relative humidity, The standard reference humidity is 50% RH.

[0075] The real-time coupled closed-loop regulation mechanism of negative pressure-vibration-airflow-static neutralization is implemented as follows:

[0076] (1) Establish a multi-parameter coupled linkage model, with real-time force feedback, paper separation state and ambient humidity as input variables, and adjust negative pressure, vibration, airflow and electrostatic neutralization parameters in real time in a closed loop to replace the traditional fixed preset parameters;

[0077] (2) The formula for real-time adjustment of micro-vibration parameters is:

[0078] (2);

[0079] in, For micro-vibration displacement function, For vibration amplitude and vibration frequency The frequency is dynamically adjusted within the 50-100Hz range based on the paper fiber strength and adsorption negative pressure. Time, in seconds;

[0080] (3) The formula for real-time adjustment of airflow auxiliary parameters is:

[0081] (3);

[0082] in, The density of air is taken as 1.205 kg / m³ under standard environmental conditions (temperature 20℃, atmospheric pressure 101.325 kPa). This is the pressure difference, measured in Pa (Pascal). The airflow velocity is expressed in m / s and is dynamically adjusted within the range of 0.3-0.5 m / s depending on the adsorption negative pressure and ambient humidity to ensure precise matching between the separation gas pressure difference and the adsorption negative pressure.

[0083] (4) The static neutralization strength of the ion nozzle is dynamically adjusted according to the ambient humidity and paper material characteristics to eliminate the influence of paper static electricity on the separation effect.

[0084] Step S3. High-fidelity scanning and graded distortion correction: The data of the vulnerable area in the page-turning stage is synchronized to the scanning stage. The pressure of the leveling roller and the scanning exposure parameters are adaptively adjusted based on the proportion of the vulnerable area to complete the high-fidelity scanning of the pages. The spine distortion area is segmented by feature extraction. The graded correction scheme is selected based on the distortion value to complete the distortion correction. At the same time, the correction error data is fed back to the page-turning stage to optimize the subsequent operation parameters.

[0085] Specifically, the method in step S3 includes:

[0086] S31. Extract semantic features of the spine based on the MobileNetV2 network, segment the spine distortion region and calculate the distortion value;

[0087] S32. Based on distorted values Implement graded correction, the formula is:

[0088] (4);

[0089] in, Let be the affine transformation matrix. It is a 3D surface reconstruction algorithm, corresponding to severe distortion correction; the pixel error after correction is controlled within 1 pixel.

[0090] Step S4. Dual-library driven OCR repair and semantic calibration: Text information and position coordinates are extracted through the CRNN algorithm. After feature encoding of the incomplete text blocks, the text is completed by combining the retrieval results of the authoritative corpus of historical classics with the contextual semantics. The semantic similarity threshold is adaptively adjusted based on the ancient book format type to complete the recognition error calibration. At the same time, the processing solutions for variant characters, special ink marks, and difficult formats are stored in the experience knowledge base.

[0091] Specifically, the method in step S4 includes:

[0092] S41. It adopts a lightweight architecture that combines CRNN model text recognition, ViT feature encoding, lightweight BiLSTM semantic encoding and RAG retrieval, and enhancement engine to replace the large parameter generative language model, thereby reducing computing power requirements and inference latency.

[0093] S42. A pre-built authoritative corpus of historical classics is constructed, including a full-text database of ancient books, a high-frequency word database, a variant character-orthodox character mapping database, a taboo character database, and a dedicated corpus of annotations for ancient books;

[0094] S43. Incomplete text completion is achieved using the following formula:

[0095] (5);

[0096] (6);

[0097] in, For contextual semantic sequence, Here, K represents the feature vector of the incomplete text, and K represents the matching result retrieved from the classical text corpus. It is a lightweight semantic coding model. To provide the optimal text completion;

[0098] S44. Based on the completed incomplete text, set differentiated semantic thresholds according to the main text format, annotation format, and chart / figure annotation format, and obtain adaptive calibration recognition results.

[0099] Specifically, the implementation method of adaptive dynamic semantic similarity threshold is as follows:

[0100] (1) Classify the types of ancient book formats by visual recognition, including the main text format, the annotation format, and the chart and annotation format;

[0101] (2) Set different semantic similarity adoption thresholds for different layouts: the threshold for the main text layout is ≥0.8, the threshold for the annotation layout is ≥0.7, and the threshold for the figure and table footnote layout is ≥0.75;

[0102] (3) The formula for calculating semantic similarity is:

[0103] ;

[0104] in, , These are the semantic feature vectors for the completed text and the context, respectively; the completed text is only adopted and output when it meets the similarity threshold of the corresponding layout.

[0105] Step S5. End-to-end adaptive optimization: Collect full-process operation parameters, success rate, and error data, select effective training samples, update the parameter mapping model, control logic, and algorithm thresholds through incremental learning algorithms, and synchronously feed the optimized strategy back to each operation stage to achieve continuous iterative optimization of the system.

[0106] Specifically, the method in step S5 includes:

[0107] Valid sample selection is achieved using the following formula:

[0108] ;

[0109] in, The overall success rate of single-page assignments is expressed as a percentage (%). This represents the pixel error for distortion correction, expressed in pixels.

[0110] The iCaRL incremental learning algorithm is used to update the lightweight XGBoost sub-model, multi-field coupling control parameters, and OCR semantic threshold. The parameter update formula is as follows:

[0111] ;

[0112] in, The model parameters before and after the update. For learning rate, For loss function, For effective training samples.

[0113] The optimized parameters and strategies are simultaneously distributed to each operational stage to achieve closed-loop optimization of the entire process.

[0114] See Figure 2 This application also provides an integrated module for non-destructive page-turning scanning and restoration of ancient books, including a multi-dimensional state perception module 1, a multi-field collaborative non-destructive page-turning module 2, a high-fidelity scanning and distortion correction module 3, a dual-library driven OCR intelligent restoration module 4, and an edge computing and adaptive optimization module 5.

[0115] The multi-dimensional state perception module 1 includes a binocular vision recognition unit 11, a paper characteristic sensor 12, a temperature and humidity sensor 13, and a MEMS pressure sensor array 14, which are used to collect physical characteristics, environmental parameters, distribution of damaged areas, and real-time stress data of ancient books.

[0116] The multi-field collaborative lossless page turning module 2 includes a soft silicone vacuum suction cup array 21, a silicone spring micro-vibration component 22, a directional airflow nozzle 23, an ion nozzle 24, and a drive control unit 25, which are used to achieve contactless separation of book pages and intelligent avoidance of vulnerable areas;

[0117] The high-fidelity scanning and distortion correction module 3 includes a ≥600dpi linear array scanning sensor 31, a multi-source illumination unit 32, an adaptive pressure elastic leveling roller 33, and an image processing unit 34, which are used for page imaging and distortion correction.

[0118] The dual-library driven OCR intelligent repair module 4 includes a CRNN recognition unit 41, a lightweight semantic coding unit 42, a historical classics corpus 43, an experience knowledge base 44, and a formatting unit 45, which are used for text recognition, incompleteness completion, and semantic calibration.

[0119] The edge computing and adaptive optimization module 5 adopts a multi-core embedded processor 51, and is equipped with a lightweight sub-model module 52 of the lightweight XGBoost sub-model, an incremental learning unit 53, and a real-time closed-loop control unit 54 for parameter matching, algorithm iteration, and module optimization. Each module realizes bidirectional data interaction through an industrial bus, forming a fully closed-loop collaborative system of "perception-decision-execution-feedback-optimization".

[0120] Example 1

[0121] See Figure 3 This application provides an integrated system and method for non-destructive page-turning scanning and restoration of ancient books, as detailed below:

[0122] 1. System initialization and multi-dimensional state awareness

[0123] (1) Hardware platform: This embodiment adopts an integrated ancient book digitization device equipped with an STM32H7 series multi-core embedded processor, and is equipped with a binocular vision recognition module, a paper tensile strength sensor, a temperature and humidity sensor, a 6mm soft silicone suction cup with an integrated 4×4 MEMS pressure sensor array, a multi-field collaborative page turning actuator, a 600dpi linear array scanning sensor, an elastic leveling roller, and a matching algorithm processing and storage unit.

[0124] (2) Ancient book information collection: The ancient book to be processed was identified by visual recognition as a Qing Dynasty thread-bound book. The finished book page size is 260mm×185mm, which conforms to the standard 16-page format commonly used in the domestic thread-bound book industry and is judged to be a large-format book. At the same time, two edge damage areas and three interline annotation areas were identified. The tensile strength of the paper fibers was detected by the paper property sensor. =0.025MPa, and the on-site environmental parameters are temperature 24℃ and relative humidity 42% RH (dry environment).

[0125] (3) Benchmark parameter matching: Call the pre-trained lightweight XGBoost sub-model corresponding to the large format of Ming and Qing period, input the five core features of format type, fiber tensile strength, aging degree, temperature and humidity, complete a single inference, the inference time is 78ms, and output the full process benchmark operation parameters: number of adsorption points 4, benchmark adsorption negative pressure 0.018MPa, micro-vibration initial frequency 75Hz, airflow initial velocity 0.38m / s, scanning resolution 600dpi, leveling roller benchmark pressure 0.04N.

[0126] 2. Lossless page turning with real-time coupling of multiple fields

[0127] (1) Adsorption point planning and vulnerable area avoidance: Four adsorption points are set according to the parameters of large format. The damaged area is initially avoided by visual recognition. The planned points are distributed in the non-text area of ​​the page, 2.5cm away from the edge of the page and 3cm away from the spine. When the suction cup comes into contact with the paper, the force distribution is detected in real time by the MEMS pressure sensor array. The force of one edge sensing unit of the suction cup exceeds the safety threshold of 0.02N for brittle ancient books. The system automatically adjusts the position of the suction cup by 2mm and then detects the force distribution again to be evenly distributed in the range of 0.01-0.015N, confirming that the vulnerable area is completely avoided.

[0128] (2) Dynamic negative pressure adjustment: Substituting the humidity-fiber strength dual-factor dynamic negative pressure formula and combining the on-site drying environment parameters, the real-time adsorption negative pressure is calculated as follows:

[0129] The adsorption negative pressure is finely adjusted based on the calculation results to adapt to the paper embrittlement characteristics under dry conditions and avoid fiber stretching damage.

[0130] (3) Multi-field collaborative separation: Simultaneously start micro-vibration, ion wind neutralization, and airflow-assisted multi-unit collaborative operation. The micro-vibration frequency is dynamically adjusted to 70Hz according to the negative pressure parameters and paper characteristics. The airflow speed is matched with the negative pressure adjustment to 0.36m / s. The ion wind intensity is appropriately increased according to the drying environment. The multi-parameter collaboration forms a "suction from above and support from below" separation force to complete the non-contact and gentle separation of the book pages.

[0131] (4) Penetration detection and leveling: The motion vectors of the target page and the lower page are detected by Farneback optical flow method. If there is no penetration problem, the page turning action is completed. The elastic leveling roller adjusts the pressure to 0.038N based on the data of the damaged area, gently smoothing the page to prepare for scanning.

[0132] 3. High-fidelity scanning and graded distortion correction

[0133] (1) Adaptive scanning imaging: The data of the damaged area during the page turning process is synchronized to the scanning module. The system automatically reduces the exposure of the identified damaged area by 20% to avoid damage to the fragile paper by strong light. After the page is left to stand still for 1.5 seconds, a high-resolution 600dpi scan is started. The multi-light source system of cold light source and ring light eliminates the dark corner of the spine and the reflection of the page, and obtains a high-definition original image of the page. The scanning parameters fully comply with the requirements of the "Technical Specification for Digitization of Ancient Books" WH / T 78-2016.

[0134] (2) Distortion region segmentation and hierarchical correction: The semantic features of the spine are extracted by MobileNetV2 and combined with the dynamic spectrum transformation to enhance the features, so as to accurately segment the distortion region of the spine and calculate the distortion value. =0.32, which is considered mild to moderate distortion; correction was performed using an affine transformation matrix, and the pixel-level error was calculated after correction. =0.7 pixels, which satisfies The accuracy requirement is ≤1, which is imperceptible to the human eye and does not affect subsequent OCR recognition.

[0135] (3) Image optimization and data feedback: The residual dark corners of the spine are eliminated by segmented histogram correction, and the image is optimized by deblurring algorithm to output a high-definition distortion-free image; at the same time, the distortion and page flatness data of this correction are fed back to the page turning module to optimize the flattening parameters and page turning strategy of subsequent pages, forming a collaborative closed loop.

[0136] 4. Dual-library driven OCR intelligent repair and semantic calibration

[0137] (1) Text recognition and feature encoding: The text content and position coordinates of the vertical text and inline annotations on the page were extracted by CRNN algorithm. Two missing text blocks were identified, one in the text area and one in the inline annotation area. ViT feature encoding was performed on the two missing text blocks to obtain the corresponding feature vectors.

[0138] (2) Intelligent completion of missing text: For missing text in the main text area, the context semantic sequence is extracted, and the corresponding literature content is matched in the authoritative corpus of historical classics through the RAG search engine. The completion probability is calculated by combining semantic encoding to determine the optimal completion text; For missing text in the annotation area, the retrieval and completion are completed by matching the exclusive corpus of ancient book annotations.

[0139] (3) Layout Adaptive Threshold Semantic Calibration: For the completion results of the main text area, a semantic similarity threshold of ≥0.8 was adopted, and the semantic similarity was verified to be 0.91, which meets the requirements and is adopted; For the completion results of the annotation area, the system automatically recognizes it as the annotation layout, and an adaptive threshold of ≥0.7 is adopted, and the semantic similarity was verified to be 0.76, which meets the requirements and is adopted.

[0140] (4) Knowledge accumulation and format output: Two rare variant characters were identified in this assignment. After the correspondence between the characters was confirmed by matching with the historical classic corpus, the relevant information was stored in the experience knowledge base for rapid identification in similar scenarios in the future. Finally, according to the original vertical double-column layout of the ancient book, a PDF file with text and image comparison and an editable text file were generated to complete the standardized output format.

[0141] 5. The end-to-end collaborative adaptive optimization system automatically collects all data from the entire process of this task, including a 100% page-turning success rate, a pixel-level error correction of 0.7 pixels, and an OCR recognition accuracy of 96.8%, which meets the valid sample selection criteria and is marked as a valid training sample. Through the iCaRL incremental learning algorithm, the corresponding sub-model and multi-field coupling control parameters of the Ming and Qing large-format books are fine-tuned and optimized. The updated parameters are embedded into the system control module for subsequent optimization of similar ancient books, realizing the system's autonomous evolution.

[0142] The beneficial effects of this application include: Significantly improved non-destructive protection: 100% success rate for turning pages of fragile ancient books, with a fiber damage rate of <0.5%. Significantly improved overall efficiency: Processing time per book reduced by over 60%, with a manual intervention rate of <2%. Perfect adaptation to edge devices: Lightweight model, fast inference, low computational power, and can be embedded. More accurate and authoritative text restoration: Incomplete text completion accuracy reaches 94%, and overall recognition rate ≥98%. System self-evolution: Iterative optimization with each task, long-term accuracy improvement of 5%–8%, and significantly reduced maintenance costs. Compliant and directly archiveable: Scanning resolution, book format, and layout output conform to industry standards.

[0143] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for integrated page-turning scanning and restoration of ancient books without damage, characterized in that, include: S1. Collect the physical characteristics and environmental parameters of ancient books, and output the baseline operation parameters for page turning, scanning and correction through a lightweight XGBoost model; S2. Based on the aforementioned benchmark operating parameters, the suction cup-integrated MEMS pressure sensor array is used to avoid vulnerable areas, dynamically adjust the adsorption negative pressure, and complete the non-destructive separation of pages using a negative pressure-vibration-airflow-electrostatic neutralization coupling mechanism, and perform page-turning penetration detection. S3. Synchronize vulnerable area data to adaptively complete high-fidelity scanning, correct spine distortion according to the distortion value, and feed back the correction error to optimize page turning parameters; S4. The CRNN model is used to recognize text, and the missing text is completed by combining an authoritative ancient book corpus and contextual semantics. The recognition results are adaptively calibrated according to the format. S5. Collect full-process operation data, and update the CRNN model and the page turning parameters through incremental learning to achieve closed-loop optimization of the CRNN model.

2. The method according to claim 1, characterized in that, In step S1, The lightweight XGBoost model takes the ancient book format, fiber strength, aging degree, temperature and humidity as inputs, and trains 6 lightweight XGBoost sub-models according to the age and the ancient book format. The baseline parameters are output only once during material loading.

3. The method according to claim 2, characterized in that, The ancient book's format is defined as 260mm × 185mm as the first threshold; if the size of the ancient book is smaller than the first threshold, it is considered a small format; if the size of the ancient book is greater than or equal to the first threshold, it is considered a large format.

4. The method according to claim 1, characterized in that, The suction cup integrates a 4×4 MEMS pressure sensor array, which automatically adjusts by 1-3mm when the force exceeds the safety threshold to avoid vulnerable areas.

5. The method according to claim 1, characterized in that, In step S2, Based on formula (1), dynamic negative pressure adjustment is performed; wherein, formula (1) is: （1）； in, To absorb negative pressure in real time; For safety factor; The tensile strength of the paper fibers being tested; This refers to the humidity adjustment coefficient. Real-time ambient relative humidity, The standard reference humidity is 50%RH.

6. The method according to claim 1, characterized in that, In step S2, the method for implementing the negative pressure-vibration-airflow-electrostatic neutralization coupling mechanism includes: Based on real-time force feedback, paper separation state, and ambient humidity as input variables, a multi-parameter coupled linkage model is established by real-time closed-loop adjustment of coupled control parameters including negative pressure, vibration, airflow, and electrostatic neutralization parameters. Based on formulas (2)-(3), the micro-vibration parameters and airflow parameters are obtained; wherein, formulas (2)-(3) are: （2）； （3）； in, For micro-vibration displacement function, For vibration amplitude and vibration frequency The frequency is dynamically adjusted within the 50-100Hz range based on the paper fiber strength and adsorption negative pressure. Time, in seconds; air density; This is the air pressure difference; This represents the airflow velocity.

7. The method according to claim 1, characterized in that, The method for implementing graded distortion collaborative correction in step S3 is as follows: S31. Extract semantic features of the spine based on the MobileNetV2 network, segment the spine distortion region and calculate the distortion variation value; S32. Based on the distorted value Implement graded correction, the formula is: （4）； in, Let be the affine transformation matrix. It is a 3D surface reconstruction algorithm, corresponding to severe distortion correction; the pixel error after correction is controlled within 1 pixel.

8. The method according to claim 1, characterized in that, The method of step S4 includes: S41. It adopts a lightweight architecture that combines CRNN model text recognition, ViT feature encoding, lightweight BiLSTM semantic encoding and RAG retrieval, and enhancement engine to replace the large parameter generative language model, thereby reducing computing power requirements and inference latency. S42. A pre-built authoritative corpus of historical classics is constructed, including a full-text database of ancient books, a high-frequency word database, a variant character-orthodox character mapping database, a taboo character database, and a dedicated corpus of annotations for ancient books; S43. Based on formulas (5)-(6), complete the missing text; wherein, formulas (5)-(6) are: （5）； （6）； in, For contextual semantic sequence, Here, K represents the feature vector of the incomplete text, and K represents the matching result retrieved from the classical text corpus. It is a lightweight semantic coding model. To provide the optimal text completion; S44. Based on the completed incomplete text, set differentiated semantic thresholds according to the main text format, annotation format, and chart / figure annotation format, and obtain adaptive calibration recognition results.

9. The method according to claim 1, characterized in that, In step S5, valid samples are defined as those with a job success rate ≥ 95% and a correction error ≤ 1 pixel. The iCaRL incremental learning algorithm is used to update the lightweight XGBoost sub-model, the coupling control parameters, and the semantic threshold.

10. An integrated system for non-destructive page-turning scanning and restoration of ancient books for implementing the method of any one of claims 1-9, characterized in that, include: The multi-dimensional state perception module is used to collect physical properties, environmental parameters, distribution of damaged areas, and real-time stress data of ancient books; A multi-field collaborative lossless page-turning module is used to achieve contactless separation of pages and intelligent avoidance of vulnerable areas; High-fidelity scanning and distortion correction module for page imaging and distortion correction; A dual-library driven OCR intelligent repair module is used for text recognition, incomplete character recognition, and semantic calibration. The edge computing and adaptive optimization module is used for parameter matching, algorithm iteration and module optimization; each module realizes bidirectional data interaction through the industrial bus, forming a closed-loop collaborative system of "perception-decision-execution-feedback-optimization".

Images