Automatic archive and book checking method and system based on artificial intelligence

Abstract

The invention discloses an automatic archive book checking method and system based on artificial intelligence, and relates to the technical field of archive checking, and the method comprises the steps: combining appearance information features with text information features, and obtaining archive book identification information; feature difference comparison is carried out on the file book identification information and historical inventory file books, intelligent duplicate removal and file book missing detection are carried out on suspected duplicate records, and inventory records are formed; inputting the inventory record into an anomaly prediction model, reasoning and analyzing a change trend of the archive book corresponding to the inventory record between a historical state and a current state, and marking an abnormal archive book; performing manual re-checking on the archive book with the abnormal mark, and dynamically adjusting the abnormity prediction model according to a re-checking result; according to the method, the problem of current inventory accuracy is solved by constructing a cross-cycle characteristic difference analysis mechanism and time sequence state evolution analysis based on the LSTM, and intelligent support is provided for long-term archive protection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of archive inventory, in particular to an archive and book automatic inventory method and system based on artificial intelligence. BACKGROUND

[0002] With the development of artificial intelligence and computer vision technology, the inventory of archives and books in management scenarios such as archives and libraries is gradually changing from manual operation to intelligent automatic inventory. The existing method usually uses high-definition camera equipment to collect multi-view images of books or archives in batches, and extracts cover, spine, number and other significant features through convolutional neural network (CNN), optical character recognition (OCR) and other algorithms to realize automatic identification and inventory of archives and books. Under the conditions of standard light, no occlusion and complete appearance, these methods can obtain high recognition accuracy, reduce manual workload and improve inventory efficiency.

[0003] The appearance information, text information and position state of the same archive and book may change slightly in different inventory periods due to long-term storage, transportation or environmental impact, such as slight fading of the cover, wear of the spine, partial loss of text or slight shift of position, etc. The traditional automatic inventory method mainly relies on single-period feature matching or rule judgment, lacks time sequence modeling and prediction ability for the evolution law of the state of archives and books over time, and cannot effectively capture the long-term trend and short-term fluctuations of the state of books; it is difficult to ensure the consistency of cross-period inventory results; especially in the case of degradation of appearance information features, missing of text information or blurred features, it is easy to misjudge as different books, resulting in incomplete inventory records; at the same time, for archives and books that have been damaged, lost or have abnormal positions, the existing method lacks accurate abnormality identification mechanism and cannot timely alarm and mark. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an automatic inventory method for archives and books based on artificial intelligence to solve the problems of repeated records and abnormal omissions caused by lack of dynamic evolution law perception.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides an automatic inventory method for archives and books based on artificial intelligence, which comprises, collecting image data of the archives and books to be inventoried; performing degradation simulation processing on the image data to generate degradation feature enhanced images; extracting information from the degradation feature enhanced images to obtain appearance information features and text information features of the archives and books, respectively; combining the appearance information features and the text information features to obtain archive and book identification information; The archive book identification information is compared with the historical inventory archive books in terms of feature difference, suspected duplicate records are intelligently deduplicated, and archive books missing are detected to form inventory records; The inventory records are input into an abnormality prediction model, the change trend of the archive books corresponding to the inventory records between the historical state and the current state is analyzed by reasoning, and the abnormal archive books are marked out; The archive books marked as abnormal are manually reviewed, and the abnormality prediction model is dynamically adjusted according to the review result.

[0007] As a preferred scheme of the archive book automatic inventory method based on artificial intelligence, wherein: the method for generating the degradation feature enhanced image comprises, Based on the historical data of the age information, the material parameters, the storage environment and the wear mode of the archive books, a set of adjustable degradation parameters is constructed; based on the set of adjustable degradation parameters, degradation simulation is applied to the image data to form the degradation feature enhanced image.

[0008] As a preferred scheme of the archive book automatic inventory method based on artificial intelligence, wherein: the method for respectively acquiring the appearance information features and the text information features of the archive books comprises, The cover, the spine and the numbering area in the degradation feature enhanced image are identified and segmented; the shape contour, the main tone distribution and the detail texture change in the cover, the spine and the numbering area are extracted to obtain the appearance information features; The cover, the spine and the numbering area containing the text and the numbering are acquired to obtain the stroke form, the arrangement order and the layout position of each character to obtain the text information features.

[0009] As a preferred scheme of the archive book automatic inventory method based on artificial intelligence, wherein: the method for combining the appearance information features and the text information features comprises, The appearance information features and the text information features are respectively standardized and dimensionally aligned; The processed appearance information features and the text information features are combined according to a preset fusion strategy to generate the archive book identification information.

[0010] As a preferred scheme of the archive book automatic inventory method based on artificial intelligence, wherein: the method for comparing the archive book identification information with the historical inventory archive books in terms of feature difference comprises, The similarity of the archive book identification information and the archive book identification information in the historical inventory records is calculated, and suspected duplicate records with a similarity reaching a preset candidate threshold are screened out; From the current inventory record and the corresponding archive book in the historical inventory record, the appearance information features and the text information features are extracted respectively, the similarity and the difference are calculated, and the comprehensive similarity and the comprehensive difference across the inventory period are obtained.

[0011] As a preferred scheme of the archive book automatic inventory method based on artificial intelligence, the method for forming the inventory record comprises, If the comprehensive similarity exceeds the preset repetition judgment threshold and the comprehensive difference is lower than the preset update judgment threshold, it is determined that the same archive book is repeated and intelligent deduplication is performed; if only one of the conditions is met and the other is not met, it is determined that the archive books are different, that is, no deduplication operation is performed, and it is considered as a new inventory record; Match all current archive book identification information with archive books in the historical inventory, and any book that exists in the historical inventory but is not matched successfully in the current inventory is marked as missing; Integrate the archive book state after intelligent deduplication and missing detection processing to form the inventory record.

[0012] As a preferred scheme of the archive book automatic inventory method based on artificial intelligence, the method for forming the inventory record comprises, Construct a book state evolution sequence according to the time sequence of the current inventory record and the archive book identification information in the historical inventory record; Based on the long short-term memory network, the time sequence feature of the book state evolution sequence is extracted by using the forget gate, the input gate and the output gate, and the long-term evolution trend and the short-term state fluctuation reflecting the appearance information features, the text information features and the position state features changing with time are obtained respectively; Combine the long-term evolution trend and the short-term state fluctuation features with the normal sample label and the abnormal sample label in the historical inventory record, supervise the training of the long short-term memory network, and obtain an abnormal prediction model; Based on the abnormal prediction model, the expected evolution law of each archive book is established, including the change trend of the appearance information features, the text information features and the position state features with time; Compare the archive book state obtained by the current inventory with the expected evolution law; when the archive book exceeds the preset allowed range in any of the appearance information features, the text information features and the position state features, the archive book is determined to be an abnormal archive book, and the abnormal feature category is labeled.

[0013] As a preferred embodiment of the artificial intelligence-based automatic inventory method for archives and books described in this invention, the method for manually reviewing archives and books marked with anomalies includes: Extract the abnormal files and books in the order of the inventory record; provide them to the manual reviewers for status verification; Manual reviewers compare the actual appearance, text, and location characteristics of the archives and books with the identification information in the inventory records. If any characteristic does not match the inventory records, the record is marked as correct. If all three characteristics match the inventory records, the record is marked as incorrect, and the review result is obtained.

[0014] As a preferred embodiment of the AI-based automatic inventory method for archives and books described in this invention, the method for dynamically adjusting the anomaly prediction model based on the review results includes: For archives and books that are correctly marked as anomalies, the anomaly weight values ​​of the corresponding appearance information features, text information features, and location status features are adjusted upward by a first preset ratio, and the features of the archives and books are added as positive samples to the anomaly prediction model training set. For archives and books marked as erroneous, the anomaly weight values ​​of the corresponding appearance information features, text information features, and location status features are adjusted downward according to the second preset ratio, and the features of the archives and books are added as negative samples to the anomaly prediction model training set. Incremental training is performed on the anomaly prediction model based on the updated training set.

[0015] Secondly, the present invention provides an automatic inventory system for archives and books based on artificial intelligence, including a degradation enhancement module for collecting image data of archives and books to be inventoried; performing degradation simulation processing on the image data to generate a degradation feature enhanced image; The feature extraction module is used to extract information from the degraded feature-enhanced image, and to obtain the appearance information features and text information features of the archives and books respectively; the appearance information features and text information features are combined to obtain the archives and books identification information; The inventory module is used to compare the archival and book identification information with the feature differences of archival and book inventories across cycles, intelligently deduplicate suspected duplicate records and detect missing archival and book records to form inventory records. The anomaly marking module is used to input inventory records into the anomaly prediction model, infer and analyze the changing trend of the archives and books corresponding to the inventory records between their historical and current states, and mark the archives and books as abnormal. The review and optimization module is used to manually review archives and books marked with anomalies, and dynamically adjust the anomaly prediction model based on the review results.

[0016] The beneficial effects of this invention are as follows: By constructing a cross-cycle feature difference analysis mechanism, screening candidate objects based on similarity thresholds, and combining the dual verification of appearance and text feature differences, this invention achieves dynamic and accurate identification of the identity of archive books, effectively distinguishing feature changes caused by natural aging from genuinely newly added books, solving the problem of "different states of the same book" that traditional single-cycle matching cannot handle, avoiding misjudgment of duplicate records and missed detection of newly added books, and significantly improving the accuracy of inventory results; Furthermore, by using LSTM-based time-series state evolution analysis, the evolution trajectory of state features is constructed according to the time series, extracting long-term evolution patterns and short-term fluctuation characteristics. This enables intelligent monitoring and prediction of the health status of archival books, breaking through the limitation of traditional inventory systems that can only identify the current state. It can provide early warning of non-natural degradation phenomena, providing a scientific basis for book preservation decisions and achieving a technological leap from passive identification to proactive prediction. The two core technologies work together to form a complete closed loop of "data acquisition - feature analysis - state prediction - feedback optimization," which not only solves the current problem of inventory accuracy but also provides intelligent support for long-term archival preservation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-based automated inventory method for archives and books.

[0019] Figure 2 This is a schematic diagram of an AI-based automated inventory system for archives and books.

[0020] Figure 3 A flowchart for obtaining archival book identification information.

[0021] Figure 4 This is a flowchart for dynamically adjusting the anomaly prediction model. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many 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 different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figure 1 , Figure 2 , Figure 3 and Figure 4 As one embodiment of the present invention, this embodiment provides an automatic inventory method for archives and books based on artificial intelligence, including the following steps: Methods for generating degraded feature-enhanced images include, Existing technologies typically only acquire clear images from a standard perspective when collecting archival book image data, lacking simulation processing of the degradation characteristics of books caused by long-term storage, wear and tear, or environmental impact. Therefore, when images have degradation conditions such as fading, stains, spine wear, or missing text, existing methods struggle to maintain the accuracy and stability of identification, and cannot effectively support intelligent judgment of cross-period inventory and abnormal books.

[0026] Based on historical data such as the age information, material parameters, storage environment and wear patterns of archival books, an adjustable degradation parameter set is constructed; based on the adjustable degradation parameter set, degradation simulation is applied to the image data to form a degradation feature enhanced image.

[0027] It should be noted that when constructing the adjustable degradation parameter set, the possible aging characteristics of the archival books are determined based on their age information, such as yellowing of paper and fading of covers. Material parameters, such as paper type and cover material, are used to determine the sensitivity of different materials to wear, fading, and stains. The potential impact of the storage environment, such as light intensity, temperature, humidity, and air quality, on the appearance and text of the archival books is analyzed, such as high light intensity areas are prone to cover fading, and excessive humidity is prone to mold or paper warping. Based on historical wear pattern data, common types and locations of physical damage are analyzed, such as wear on the spine corners, scratches on the cover, stains in the numbered area, or text loss. The adjustable degradation parameter set includes light attenuation parameters, color fading parameters, physical wear parameters, stains, occlusion parameters, and text loss parameters; Degradation simulation is applied to image data, including image processing algorithms (such as convolutional filtering, color transformation, and texture synthesis) or depth generation models. Based on the light attenuation parameters and color fading parameters in the adjustable degradation parameter set, the brightness of the whole or local areas of the image is reduced, the contrast is adjusted, and the saturation is attenuated. By simulating the cover fading and paper yellowing effects caused by light exposure, the light aging characteristics of archival books from different eras during long-term storage are reflected. Based on physical wear parameters, scratches, worn edges, bends, or creases are generated for key areas in the image (such as the spine, cover corners, and numbering areas). The physical wear parameters are adjusted according to the sensitivity of the archival book material and historical wear patterns. For example, scratches may appear on hard covers, while paper covers are prone to bending and damage. By utilizing stain and occlusion parameters, localized occlusion areas such as dust, water stains, and oil stains are randomly generated in the image, with varying transparency and shape. The stain distribution density and coverage area are adjusted based on storage environment parameters (such as humidity and air quality) to make the image more closely resemble the potential contamination conditions in actual storage environments. Based on text loss parameters, text and numbering in images are partially missing, blurred, have broken strokes, or are slightly deformed. The degradation intensity is adjusted according to the age, material, and historical wear patterns of the archives and books; for example, old paper or frequently used numbering areas are more prone to text wear or detachment. The degradation features enhancement image is generated by combining the effects of light attenuation, color fading, physical wear, stains and occlusion, and text degradation according to the weights of an adjustable degradation parameter set. The weights of the various types of the adjustable degradation parameter set can be exemplified as follows: light attenuation accounts for 20%~30%, color fading accounts for 15%~25%, physical wear accounts for 20%~30%, stains and occlusion account for 10%~20%, and text degradation accounts for 10%~20%. The weights can be dynamically adjusted according to the age, material, storage environment, and historical wear patterns of the archives and books to make the degradation feature enhancement image more consistent with the actual aging and damage.

[0028] The degradation simulation results in enhanced images that can fully simulate the aging, wear, stains, and missing text that may occur in actual archival books, providing a reliable data foundation for subsequent extraction of appearance and text information features, cross-period inventory, and anomaly detection.

[0029] Methods for separately acquiring the appearance and textual information features of archival books include, The cover, spine, and numbering regions in the degraded feature-enhanced image are identified and segmented; shape contours, main color distribution, and detail texture changes are extracted from the cover, spine, and numbering regions to obtain appearance information features.

[0030] Specifically, the structural features of the archive books in the degraded feature enhancement image are analyzed. First, the location of the cover, spine and number is determined. Then, the boundary range is gradually expanded by using the differences in brightness, color and texture continuity of these regions until the boundary is highly consistent with the real physical contour, so as to obtain accurate segmentation results. When acquiring shape contour features, the edge pixels of the segmented region are scanned sequentially, the directional change amplitude and curvature change rate of the edge pixels at adjacent positions are analyzed, and the extension length of the edge in the horizontal and vertical directions is recorded. Combined with the aspect ratio, symmetry and corner distribution of the overall region, the geometric contour characteristics of the region are described. When acquiring overall color information, all pixels within the segmented area are converted to a unified color representation space. The frequency of color components of each pixel is counted. Based on the frequency and distribution range of color occurrence, similar colors are aggregated to form a set of main colors, and the proportion of each main color in the area is marked. When extracting detailed texture features, the segmented region is divided into several directional windows (such as horizontal, vertical, and diagonal directions) and multi-scale windows (such as different pixel spacing or region size). The variation amplitude and frequency distribution of pixel grayscale or color in the corresponding direction within each window are calculated respectively. Texture repeatability, roughness, and detail variation features under different directions and scales are extracted. The above geometric contour characteristics, color information, and detailed texture features are combined to form a unified appearance information feature vector.

[0031] For the cover, spine, and numbered area containing text and numbers, obtain the stroke shape, arrangement order, and page position of each character to obtain text information features.

[0032] Specifically, the process involves identifying the overall style and subtle differences in the strokes of characters; recording the arrangement order of characters according to their natural reading direction within the segmented area to ensure that the logical structure of the text or number is reflected; marking the specific position of each character in the entire area of ​​the page, including character spacing, line spacing, offset relative to the edge, and alignment; and finally combining multi-dimensional information such as stroke shape, arrangement order, and page position to form complete text information features.

[0033] Methods for combining appearance information features with textual information features include,

[0034] The appearance information features and text information features are standardized and dimensionally aligned respectively.

[0035] Specifically, the values ​​and distribution characteristics of appearance information features and text information features in historical inventory records are statistically analyzed across each dimension. The values ​​of each dimension are normalized according to the mean and standard deviation, so that the mean is 0 and the variance is 1. Outliers exceeding, for example, an absolute z-score of 3 are truncated or replaced. A unified target dimension is set based on the cumulative dimensions of appearance information features and text information features in historical inventory records, such as 256 dimensions. When the feature dimension is greater than the target dimension, the main change information components are retained and redundant components are discarded. When the feature dimension is less than the target dimension, the statistical summary is added in sequence or zero values ​​are filled until the target dimension is reached.

[0036] The processed appearance information features and text information features are combined according to a preset fusion strategy to generate archive and book identification information.

[0037] It should be noted that the preset fusion strategy is as follows: First, fusion weights are assigned based on the stability and discriminative power of appearance information features and text information features in historical recognition tasks, for example, appearance information features are weighted at 0.6 and text information features at 0.4; then, the feature values ​​of the corresponding dimensions are summed according to their weights to obtain the preliminary fused vector; for dimensions with large fluctuations in the fused vector, weight smoothing is added to reduce the impact of abnormal fluctuations of a single feature on the overall recognition result; finally, the fusion result is normalized once to keep its range within a set range, such as 0~1, to ensure that the identification information of different archives and books is comparable within the same feature space.

[0038] For example, when the appearance information features are 0.2, 0.5, and 0.8, and the text information features are 0.6, 0.4, and 0.7, the fusion vectors are obtained by weighting them with weights of 0.6 and 0.4, resulting in fusion vectors of 0.36, 0.46, and 0.78. After smoothing and normalization, the archive and book identification information is generated.

[0039] Methods for comparing the characteristic differences between archival and book identification information and historically inventoried archival and book records include: The similarity between the archival book identification information and the archival book identification information in the historical inventory records is calculated, and suspected duplicate records that reach the preset candidate threshold are screened out.

[0040] It should be noted that the expression for similarity calculation is: ; in, This represents the similarity between archival book identification information A and archival book identification information B, with a value ranging from 0 to 1. The closer the value is to 1, the more similar the two are. This is the first piece of archival book identification information in [the following text is missing]. Feature values ​​in each dimension The second piece of archival book identification information is in the [number]th [section]. Feature values ​​in each dimension This represents the total number of dimensions of the feature vectors. For example, 256 means that the feature vectors have 256 dimensions. It is an index variable for the feature dimension; Furthermore, randomly select, for example, 500 pairs of archival book pairs (where, for example, 250 pairs are multiple versions of the same book, and for example, 250 pairs are different book records), and calculate their similarity distribution; set the value corresponding to the 5th percentile in the similarity distribution of multiple versions of the same book as the lower limit of recognition, and set the value corresponding to the 95th percentile in the similarity distribution of different book records as the upper limit of distinction; select a value between the two that optimizes the false positive rate and the false negative rate as the candidate threshold, for example, 0.85, and dynamically fine-tune this value based on the results of manual review in subsequent runs.

[0041] For each suspected duplicate record, appearance features and text features are extracted from the current inventory of archives and books and the corresponding archives and books in the historical inventory records, respectively. The similarity and difference are calculated to obtain the comprehensive similarity and comprehensive difference across the inventory period.

[0042] It should be noted that the expressions for extracting appearance information features and text information features separately, and calculating similarity and difference, are as follows: Similarity of appearance information features: ; in, The current inventory of archival books features an appearance characteristic. Appearance characteristics compared with historical inventory records similarity, This is the current inventory of appearance information characteristics. In the Feature values ​​in each dimension It is the appearance information characteristics in the historical inventory records. In the Feature values ​​in each dimension; Text information feature similarity: ; in, The characteristics of the current inventory text information of archives and books Features of textual information in historical inventory records similarity, The characteristics of the current inventory text information In the Feature values ​​in each dimension It is a feature of textual information in historical inventory records. In the Feature values ​​in each dimension; Difference in appearance information features: ; in, The current inventory of archival books features an appearance characteristic. Appearance characteristics compared with historical inventory records Difference Textual information feature difference: ; in, The characteristics of the current inventory text information of archives and books Features of textual information in historical inventory records Difference Furthermore: the expressions for obtaining the comprehensive similarity and comprehensive difference across inventory periods are as follows: Overall similarity across inventory periods: ; in, It is the comprehensive similarity between the current inventory of archives and books and the corresponding archives and books in historical inventory records across inventory periods. This refers to the weight of appearance information features in the overall similarity calculation, for example, 0.6. This represents the weight of textual information features in the overall similarity calculation, for example, 0.4; Overall Difference Across Inventory Periods: ; in, It represents the overall difference between the current inventory of archives and books and the corresponding archives and books in historical inventory records across different inventory periods.

[0043] Methods for creating inventory records include, If the overall similarity exceeds the preset duplicate judgment threshold and the overall difference is lower than the preset update judgment threshold, it is judged as a duplicate record of the same archive book and intelligent deduplication is performed; if only one condition is met and the other is not met, it is judged as a different archive book, that is, no deduplication operation is performed and it is regarded as a new inventory record.

[0044] It should be noted that the source of the duplicate judgment threshold is based on the similarity distribution statistics of historical inventory records, and the 85th percentile of the comprehensive similarity is selected as the threshold. The source of the updated judgment threshold is based on the statistical analysis of the difference distribution, and the 15th percentile of the comprehensive difference is selected as the threshold. Furthermore, intelligent deduplication refers to merging historical records, retaining the latest inventory information, or updating identification information when the overall similarity exceeds the preset duplication threshold and the overall difference is lower than the preset update threshold.

[0045] Match all current archival book identification information with the archival books in the historical inventory. Any book that exists in the historical inventory but cannot be matched in the current inventory is marked as missing.

[0046] It should be noted that when a current inventory of archival book identification information has matching candidates with multiple historical inventory of archival books, the archival book in the historical inventory with the highest comprehensive similarity and the lowest comprehensive difference is selected as the matching object; the matching order can be carried out in the order of the current inventory of archival books. It should be noted that the criteria for determining a matching failure are as follows: when a certain archival book in the historical inventory is compared with the identification information of all archival books in the current inventory, there is no record that simultaneously satisfies the condition that the overall similarity is greater than or equal to the duplication threshold and the overall difference is less than or equal to the update threshold. For example, if the duplication threshold is set to 0.85 and the update threshold is set to 0.15, then when a certain archival book in the historical inventory is compared with all current inventory records, and there is no inventory record that simultaneously satisfies the condition that the overall similarity is greater than or equal to 0.85 and the overall difference is less than or equal to 0.15, the archival book is determined to be missing and is marked in the missing list.

[0047] The status of archives and books after intelligent deduplication and missing item detection is integrated to form an inventory record.

[0048] Specifically, for archive and book records that have undergone intelligent deduplication, the identification information, appearance information features, and text information features of the latest inventory cycle are retained and marked as "retained" in the status field; for archives and books that are determined to be missing in the missing detection, they are marked as "missing" in the status field, and the reason for the missing and the timestamp of the missing detection are attached; for archives and books that appear in the current inventory and are not matched in the historical records, they are marked as "new" in the status field.

[0049] Methods for reasoning and analyzing the changing trends of archives and books corresponding to inventory records between their historical and current states include... Existing methods for inventorying archives and books typically rely solely on static information such as appearance or location collected during a single inventory check, lacking modeling of cross-period state evolution and failing to effectively identify potential anomalies caused by long-term changes or short-term fluctuations. Furthermore, traditional methods often rely on human experience to set judgment rules, failing to fully utilize supervised learning from normal and abnormal samples in historical inventory data. This results in vague judgment criteria that are prone to misjudgment or omission. Simultaneously, the lack of differentiated allowable ranges for different feature categories such as appearance, text, and location makes it difficult to accurately pinpoint the type and severity of problems in anomaly detection results, impacting the automation and intelligence level of inventory checks.

[0050] Construct a sequence of book status evolution by combining the current inventory record with the archival book identification information in the historical inventory record in chronological order.

[0051] It should be noted that when constructing the book status evolution sequence, the identification information of the same archival books in the current inventory record and the historical inventory record is sorted by timestamp to form an identification information sequence arranged in ascending order of inventory time; based on the feature vector of the archival books at each time point, the feature difference degree between adjacent time nodes is calculated. The calculation of the difference degree can combine the difference in appearance features, the difference in text features, and the change in location status; for nodes whose difference degree exceeds the preset change threshold, they are marked as status change nodes, and their change type is recorded, such as the addition of appearance damage, the loss of number information, and the change in location, etc. Furthermore, a sliding window can be used to smooth the identification information sequence, eliminating noise caused by a single anomaly identification, thereby obtaining the book state evolution sequence.

[0052] Based on the Long Short-Term Memory Network, the forget gate, input gate and output gate are used to extract temporal features from the book state evolution sequence, and the long-term evolution trend and short-term state fluctuations reflecting the changes of appearance information features, text information features and positional state features over time are obtained respectively.

[0053] It should be noted that the forgetting gate compares the current state of a book with its historical state to determine which historical information is still valuable. For example, if the book number has not changed and the degree of wear and tear on the exterior remains stable, the historical information is retained; if new stains appear on the cover or the number is missing, the weight of the past related state is reduced. The forgetting gate reduces the importance of unimportant information by setting weight ratios for each type of feature, for example: 70% of historical appearance information is retained, 90% of historical text information is retained, and 80% of the location status is retained. The input gate evaluates the book status data recorded in the current inventory, introduces new information, and determines its impact on the overall status update. For example, if a newly scanned cover image shows fading or stains, the input gate will increase the weight of the appearance information; if the text number is missing or updated, the weight of the text information will be increased. The input gate controls the contribution of new information to the overall status by allocating weights, for example: new appearance information weight 30%~50%, text information weight 50%~70%, and location status weight 20%~40%. The output gate integrates the historical information filtered by the forget gate with the new information introduced by the input gate to generate a comprehensive state feature vector at the current time point; the output gate highlights long-term trends (such as the appearance gradually fading) and short-term fluctuations (such as the temporary movement of the book or the error in the text number) by combining the current state weights. Through the coordinated processing of the forget gate, input gate, and output gate, the short-term memory network can dynamically adjust the weights of historical information and new data in each inventory, comprehensively reflecting the long-term trend and short-term fluctuations of the book's status evolution, and providing reliable temporal feature support for automatic inventory.

[0054] By combining long-term evolution trends, short-term state fluctuation characteristics, and normal and abnormal sample labels from historical inventory records, a supervised training of the Long Short-Term Memory Network is conducted to obtain an anomaly prediction model.

[0055] Specifically, the long-term evolution trend features and short-term state fluctuation features extracted from each archival book are paired with their corresponding normal or abnormal sample labels in historical inventory records to form a supervised training dataset. During training, the long-term evolution trend features and short-term state fluctuation features are input into a Long Short-Term Memory (LSTM) network. The LTM network learns the mapping relationship between state evolution and the occurrence of anomalies based on historical labels. By iteratively optimizing the LTM network parameters, it is enabled to accurately capture the correlation between long-term trends and short-term fluctuations and anomaly labels, thereby generating a model capable of predicting abnormal book states.

[0056] Based on the anomaly prediction model, the expected evolution pattern of each archival book is established, including the changing trends of appearance information features, text information features, and location status features over time.

[0057] Specifically, the sequence of book status evolution is input into a trained anomaly prediction model to extract long-term evolution trends and short-term status fluctuation features. The extracted long-term trends and short-term fluctuation features are then comprehensively analyzed: the long-term trends describe the overall change trend of each feature over time, such as the cumulative changes in color fading and spine wear; the short-term fluctuations reflect the minor fluctuations of each feature between adjacent inventory cycles, such as slight positional shifts or changes in text markings; the long-term trends and short-term fluctuations are integrated to generate the expected evolution pattern of each archival book in future inventory cycles.

[0058] The current status of the archives and books obtained from the inventory is compared with the expected evolution pattern. When any feature of the archives or books in terms of appearance, text, or location exceeds the preset allowable range, the archives or books are determined to be abnormal archives or books, and the abnormal feature category is marked.

[0059] Specifically, for each archival book, the appearance, textual, and location characteristics obtained from the current inventory are compared item by item with the expected evolution pattern of the book. Appearance characteristics include indicators such as color, cover integrity, and degree of wear; the allowable deviation range can be exemplarily set as color deviation ±5%, and minor cover wear not exceeding two locations. Textual characteristics include numbering, text integrity, and readability; the allowable deviation range can be exemplarily set as numbering error within one character, localized text wear not affecting overall readability, and a decrease in readability score not exceeding 5%. Location characteristics include actual storage location; the allowable deviation range can be exemplarily set as a horizontal / vertical offset of one shelf unit or ±10 cm from the expected location. Subsequently, the actual values ​​of each characteristic are compared with their corresponding allowable ranges. If any characteristic exceeds the allowable range, the archival book is considered to be in an abnormal state, and the abnormal characteristic category is marked.

[0060] By constructing a sequence of book status evolution and utilizing the forgetting, input, and output gates of a Long Short-Term Memory (LSTM) network to extract long-term trends and short-term fluctuations, dynamic modeling of cross-cycle changes in archival books was achieved. Supervised training, incorporating historical normal and abnormal sample labels, yielded a model capable of accurately predicting anomalies. Furthermore, quantified allowable deviation ranges were set for three types of features: appearance, text, and location, enabling differentiated anomaly identification. This approach not only identifies significant anomalies but also detects subtle changes such as minute shifts or slight wear, automatically labeling anomaly categories. This improves the accuracy, stability, and intelligence of inventory checks, while reducing manual intervention and misjudgment rates.

[0061] Methods for manually reviewing archives and books marked with anomalies include: Extract the abnormal files and books in the order of the inventory record; provide them to the manual reviewers for status verification.

[0062] Manual reviewers compare the actual appearance, text, and location characteristics of the archives and books with the identification information in the inventory records. If any characteristic does not match the inventory records, the record is marked as correct. If all three characteristics match the inventory records, the record is marked as incorrect, and the review result is obtained.

[0063] Methods for dynamically adjusting anomaly prediction models based on review results include: For archives and books whose anomalies are correctly identified, the anomaly weight values ​​of the corresponding appearance information features, text information features, and location status features are adjusted upwards according to a first preset ratio, and the features of the archives and books are added as positive samples to the anomaly prediction model training set.

[0064] It should be noted that manual verification and comparison confirmed that the archives and books marked as abnormal during the current inventory did indeed have anomalies such as physical damage, missing text, or misaligned positions. For these books, the anomaly weights of the corresponding physical appearance features, textual information features, and positional status features were adjusted upwards by a first preset ratio, for example, increasing the original weights by 10% to 20%, to enhance the contribution of these features in the anomaly prediction model's anomaly detection. Subsequently, the physical appearance features, textual information features, and positional status features of the archive / book were combined to form a vector, which was then added as a positive sample to the anomaly prediction model's training set.

[0065] For archives and books marked as erroneous, the anomaly weight values ​​of the corresponding appearance information features, text information features, and location status features are adjusted downward according to a second preset ratio, and the features of the archives and books are added as negative samples to the anomaly prediction model training set.

[0066] It should be noted that manual review and comparison revealed that books marked as abnormal were actually not abnormal; for example, their appearance was intact, the text was complete, and their storage location was correct. For these books, the anomaly weights of their corresponding appearance, text, and location features were adjusted downwards by a second preset ratio, for example, reducing the original weights by 10% to 20%, to reduce the impact of these features on the anomaly prediction model's identification of anomalies. Simultaneously, the appearance, text, and location features of these books were combined into a vector and added as a negative sample to the anomaly prediction model's training set to optimize the model's recognition ability and reduce false positives. Incremental training is performed on the anomaly prediction model based on the updated training set.

[0067] It should be noted that when the error rate of a single batch review is greater than 10%, incremental training is started immediately; the positive and negative samples obtained after review and annotation are merged to form an updated training set, in which each sample contains a multi-dimensional feature vector containing appearance information features, text information features, and position and state features; the feature vectors of all samples are normalized in numerical range using a unified standardization method (such as Z-score standardization or Min-Max normalization) to ensure comparability between different feature dimensions; When performing incremental training, all parameters of the existing anomaly prediction model are loaded as initial weights, and only the newly added positive and negative samples are used in the training iteration. During the training process, a mini-batch gradient descent method is adopted, with a fixed number of samples in each batch (e.g., 32 or 64 samples) and a small learning rate (e.g., 0.0001 to 0.001) is set to control the magnitude of the anomaly prediction model parameter update. After each training round, the accuracy, recall, and F1 score of the current anomaly prediction model are evaluated using a validation set. Training is terminated early when the performance metrics meet the following conditions: accuracy ≥ 95%, recall ≥ 90%, or the improvement in metrics for multiple consecutive rounds is less than a set value (e.g., 0.1%). Finally, the parameters of the anomaly prediction model that has completed incremental training are saved and used to replace the original anomaly prediction model for subsequent anomaly prediction tasks.

[0068] This embodiment also provides an artificial intelligence-based automatic inventory system for archives and books, including: The degradation enhancement module is used to collect image data of archives and books to be inventoried; it performs degradation simulation processing on the image data to generate degradation feature enhanced images; The feature extraction module is used to extract information from the degraded feature-enhanced image, and to obtain the appearance information features and text information features of the archives and books respectively; the appearance information features and text information features are combined to obtain the archives and books identification information; The inventory module is used to compare the archival and book identification information with the feature differences of archival and book inventories across cycles, intelligently deduplicate suspected duplicate records and detect missing archival and book records to form inventory records. The anomaly marking module is used to input inventory records into the anomaly prediction model, infer and analyze the changing trend of the archives and books corresponding to the inventory records between their historical and current states, and mark the archives and books as abnormal. The review and optimization module is used to manually review archives and books marked with anomalies, and dynamically adjust the anomaly prediction model based on the review results.

[0069] This embodiment also provides a computer device applicable to the case of an automatic inventory method for archives and books based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the automatic inventory method for archives and books based on artificial intelligence as proposed in the above embodiment.

[0070] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0071] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based automatic inventory method for archives and books as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0072] In summary, this invention achieves dynamic and accurate identification of the identity of archive books by constructing a cross-period feature difference analysis mechanism, screening candidate objects based on similarity thresholds, and combining dual verification of appearance and text feature differences. It effectively distinguishes feature changes caused by natural aging from genuinely newly added books, solves the problem of "different states of the same book" that traditional single-period matching cannot handle, avoids misjudgment of duplicate records and missed detection of newly added books, and significantly improves the accuracy of inventory results.

[0073] Furthermore, by using LSTM-based time-series state evolution analysis, the evolution trajectory of state features is constructed according to the time series, and long-term evolution patterns and short-term fluctuation characteristics are extracted. This enables intelligent monitoring and prediction of the health status of archival books, breaking through the limitation of traditional inventory systems that can only identify the current state. It can provide early warning of non-natural degradation phenomena, providing a scientific basis for book preservation decisions and achieving a technological leap from passive identification to proactive prediction. The two core technologies work together to form a complete closed loop of "data acquisition - feature analysis - state prediction - feedback optimization", which not only solves the current problem of inventory accuracy, but also provides intelligent support for long-term archival preservation.

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic inventory method for archives and books based on artificial intelligence, characterized in that: include, Collect image data of the archives and books to be inventoried; perform degradation simulation processing on the image data to generate images with enhanced degradation features; Information is extracted from the degraded feature-enhanced image to obtain the appearance information features and text information features of the archives and books respectively; the appearance information features and text information features are combined to obtain the archives and books identification information; The archival and book identification information is compared with the characteristic differences of historically inventoried archival and book records. Intelligent deduplication and missing archival and book detection are performed on suspected duplicate records to form an inventory record. The inventory records are input into the anomaly prediction model, and the changing trend of the archives and books corresponding to the inventory records between their historical and current states is analyzed to mark the abnormal archives and books. The archives and books marked with anomalies are manually reviewed, and the anomaly prediction model is dynamically adjusted based on the review results.

2. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 1, characterized in that: The method for generating degraded feature-enhanced images includes, Based on historical data on the age, material parameters, storage environment, and wear patterns of archival books, an adjustable degradation parameter set is constructed. Based on an adjustable set of degradation parameters, degradation simulation is applied to image data to form an image with enhanced degradation features.

3. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 1, characterized in that: The methods for obtaining the appearance information features and textual information features of archival books respectively include: The cover, spine, and serial number regions in the degraded feature-enhanced image are identified and segmented; shape contours, main color distribution, and detail texture changes are extracted from the cover, spine, and serial number regions to obtain appearance information features; For the cover, spine, and numbered area containing text and numbers, obtain the stroke shape, arrangement order, and page position of each character to obtain text information features.

4. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 3, characterized in that: The method for combining appearance information features and text information features includes, The appearance information features and text information features are standardized and dimensionally aligned respectively; The processed appearance information features and text information features are combined according to a preset fusion strategy to generate archive and book identification information.

5. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 4, characterized in that: The method for comparing the feature differences between archival book identification information and historically inventoried archival books includes: The similarity between the archival and book identification information and the archival and book identification information in historical inventory records is calculated, and suspected duplicate records whose similarity reaches the preset candidate threshold are screened out. For each suspected duplicate record, appearance features and text features are extracted from the current inventory of archives and books and the corresponding archives and books in the historical inventory records, respectively. The similarity and difference are calculated to obtain the comprehensive similarity and comprehensive difference across the inventory period.

6. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 5, characterized in that: The method for generating inventory records includes: If the overall similarity exceeds the preset duplicate judgment threshold and the overall difference is lower than the preset update judgment threshold, it is judged as a duplicate record of the same archive book and intelligent deduplication is performed; if only one condition is met and the other is not met, it is judged as a different archive book, that is, no deduplication operation is performed and it is regarded as a new inventory record. Match all current archival book identification information with the archival books in the historical inventory. Any book that exists in the historical inventory but cannot be matched in the current inventory is marked as missing. The status of archives and books after intelligent deduplication and missing item detection is integrated to form an inventory record.

7. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 6, characterized in that: The methods for reasoning and analyzing the changing trends of the archives and books corresponding to the inventory records between their historical and current states include... Construct a sequence of book status evolution by combining the current inventory record with the archival book identification information in the historical inventory record in chronological order; Based on the Long Short-Term Memory Network, the forget gate, input gate and output gate are used to extract the temporal features of the book state evolution sequence, and the long-term evolution trend and short-term state fluctuations reflecting the changes of appearance information features, text information features and positional state features over time are obtained respectively. By combining long-term evolution trends, short-term state fluctuation characteristics with normal and abnormal sample labels in historical inventory records, the long short-term memory network is trained under supervision to obtain an anomaly prediction model. Based on the anomaly prediction model, the expected evolution pattern of each archival book is established, including the changing trends of appearance information features, text information features, and location status features over time. The current status of the archives and books obtained from the inventory is compared with the expected evolution pattern. When any feature of the archives or books in terms of appearance, text, or location exceeds the preset allowable range, the archives or books are determined to be abnormal archives or books, and the abnormal feature category is marked.

8. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 1, characterized in that: The method for manually reviewing archives and books marked with anomalies includes: Extract the books with abnormal records in the order of the inventory records; Provided to manual reviewers for status verification; Manual reviewers compare the actual appearance, text, and location characteristics of the archives and books with the identification information in the inventory records. If any characteristic does not match the inventory records, the record is marked as correct. If all three characteristics match the inventory records, the record is marked as incorrect, and the review result is obtained.

9. The method for automatic inventory of archives and books based on artificial intelligence as described in claim 8, characterized in that: The method for dynamically adjusting the anomaly prediction model based on the review results includes: For archives and books that are correctly marked as anomalies, the anomaly weight values ​​of the corresponding appearance information features, text information features, and location status features are adjusted upward by a first preset ratio, and the features of the archives and books are added as positive samples to the anomaly prediction model training set. For archives and books marked as erroneous, the anomaly weight values ​​of the corresponding appearance information features, text information features, and location status features are adjusted downward according to the second preset ratio, and the features of the archives and books are added as negative samples to the anomaly prediction model training set. Incremental training is performed on the anomaly prediction model based on the updated training set.

10. An AI-based automatic inventory system for archives and books, based on the AI-based automatic inventory method for archives and books as described in any one of claims 1 to 9, characterized in that: include, The degradation enhancement module is used to collect image data of archives and books to be inventoried; it performs degradation simulation processing on the image data to generate degradation feature enhanced images; The feature extraction module is used to extract information from the degraded feature enhancement image, and to obtain the appearance information features and text information features of the archives and books respectively; By combining appearance information features with text information features, the identification information of archives and books is obtained; The inventory module is used to compare the archival and book identification information with the feature differences of archival and book inventories across cycles, intelligently deduplicate suspected duplicate records and detect missing archival and book records to form inventory records. The anomaly marking module is used to input inventory records into the anomaly prediction model, infer and analyze the changing trend of the archives and books corresponding to the inventory records between their historical and current states, and mark the archives and books as abnormal. The review and optimization module is used to manually review archives and books marked with anomalies, and dynamically adjust the anomaly prediction model based on the review results.

Images