Direct injection English flipping and reading card classification method based on image recognition processing

Abstract

The invention relates to the technical field of image recognition processing, and discloses a direct-injection English read-through card classification method based on image recognition processing, and the method comprises the steps: obtaining a damaged English read-through card image and a historical learning record of a user, recognizing a damaged region, and extracting the features of an individual forgetting mode; calculating a memory attenuation degree and a credibility coefficient based on an improved Ebbinghaus forgetting curve fitting algorithm; utilizing a memory guide content reconstruction algorithm, and taking the memory credibility as a constraint weight to adjust an image restoration parameter; semantic probability distribution is calculated based on the vocabulary features, and restoration candidate schemes are generated in combination with memory constraints; analyzing the matching degree between the candidate scheme and the user memory state by using a time sequence memory verification algorithm; verifying the semantic rationality of the repaired content based on a language model; and performing content identification on the restored image to generate a classification result of comprehensive restoration and time sequence optimization. According to the method, intelligent image restoration of semantic constraint is realized, and integrated processing of restoration and classification is realized.

Description

Technical Field

[0001] This invention relates to the field of image recognition processing technology, and more specifically, to a method for classifying English translation cards based on image recognition processing. Background Technology

[0002] With the rapid development of intelligent education technology, image recognition-based learning aids are widely used in English learning. However, English flashcards, as a traditional and effective vocabulary learning tool, often suffer physical damage during long-term use, including missing, blurred, and torn text, which seriously affects learning outcomes.

[0003] Existing image inpainting techniques are mainly based on pixel-level texture synthesis and structure propagation algorithms. The sample-based image inpainting algorithm proposed by Criminisi et al. fills in damaged areas by sampling similar textures from known regions; this method performs well when processing simple textures in natural images. The partial differential equation-based inpainting method developed by Bertalmio et al. achieves content filling by extending the isoluminance lines of the image. In addition, total variational image inpainting techniques and sparse representation-based inpainting methods are also being applied in related fields.

[0004] However, these traditional methods primarily rely on pixel information surrounding the damaged area for geometric and texture inference, lacking an understanding of the semantics of the image content. When applied to image restoration containing text, they easily produce restoration results that do not conform to the original semantics, especially for content with specific linguistic rules such as English vocabulary, where pure pixel completion often generates meaningless character combinations. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for classifying direct-reading English flashcards based on image recognition processing, comprising the following steps: Acquire images of damaged English flashcards and users' historical learning records, identify damaged areas, and extract individual forgetting pattern features of users; The degree of memory decay and the reliability coefficient of memory are calculated based on the forgetting curve fitting algorithm; The memory-guided content reconstruction algorithm is used to adjust the image restoration parameters by using the memory credibility coefficient as a constraint weight; Based on the semantic probability distribution calculated from lexical features, and combined with memory constraints, candidate repair solutions are generated. The temporal memory verification algorithm is used to analyze the matching degree between candidate solutions and the user's memory state, and the optimal repair content is selected. The repair results are output after verifying the semantic rationality of the repaired content based on a language model. Content recognition is performed on the restored image to generate a classification result that integrates damage restoration and temporal optimization; Among them, the memory-guided content reconstruction algorithm modifies the image repair objective function by establishing a memory constraint matrix, the forgetting curve fitting algorithm adopts an improved Ebbinghaus forgetting curve model combined with individual difference parameters, and the temporal memory verification algorithm calculates the temporal consistency score by constructing a memory state time trajectory diagram.

[0006] Preferably, the extraction of individual user forgetting pattern features includes: extracting learning records from the user's historical learning database and constructing a forgetting pattern feature vector that includes learning time intervals, review frequency, historical accuracy, and vocabulary difficulty level.

[0007] Preferably, the calculation process of the memory reliability coefficient includes: obtaining the memory retention rate value and the context weight adjustment factor value, multiplying the two values ​​to obtain the memory reliability coefficient, which comprehensively reflects the combined impact of time decay effect and individual differences on the accuracy of user memory.

[0008] Preferably, the establishment of the memory constraint matrix includes: multiplying the memory credibility coefficient with the semantic constraint vector constructed based on the user's memory content to form a numerical matrix used to constrain the restoration content to conform to the user's memory during the image restoration process.

[0009] Preferably, the calculation of the semantic probability distribution includes: using an improved TF-IDF algorithm to calculate the semantic relevance of candidate words in the current context. This algorithm introduces a semantic similarity factor and a user learning level matching factor to ensure that the recommended candidate words match the user's current learning level.

[0010] Preferably, the calculation of the temporal consistency score includes: constructing a continuous memory state curve using cubic spline interpolation, calculating the time weight using an exponential decay function, and obtaining the matching degree by comparing the similarity between the candidate scheme and the historical memory state.

[0011] Preferably, the language model verification uses a pre-trained BERT model, and the semantic rationality is evaluated by calculating the perplexity of the repaired content in the language model. When the perplexity exceeds a preset threshold, the candidate solution is returned to be regenerated.

[0012] Preferably, the generation of the classification result includes: using optical character recognition technology to identify the text content in the restored image, and based on the identified complete content and the user's memory reinforcement timing prediction results, generating a personalized English flip card classification scheme that includes difficulty level, review priority and recommended study time.

[0013] Preferably, the forgetting curve fitting algorithm uses the least squares method to fit the user's individualized forgetting parameters and determines the optimal value of the individualized parameters through nonlinear regression analysis. The individualized parameters reflect the user's basic forgetting rate and the shape characteristics of the forgetting curve.

[0014] A direct-reading English flashcard classification system based on image recognition processing includes: The image acquisition module is used to acquire digital images of damaged English translation cards; The database management module is used to store and manage user learning records, vocabulary knowledge base, image sample library, and model parameter library; The image processing module is used to identify the boundaries of damaged areas and extract forgetting pattern features; The memory modeling module is used to calculate the degree of memory decay and the memory reliability coefficient; The content reconstruction module is used to execute the memory-guided content reconstruction algorithm, which transforms memory information into image restoration constraints. The semantic analysis module is used to calculate the semantic probability distribution and generate candidate repair solutions; The timing verification module is used to execute the timing memory verification algorithm to select the optimal repair content; The results output module is used to generate the final repair results and classification scheme; The modules interact with each other through data interfaces to collaboratively complete the intelligent repair and classification tasks of damaged English translation cards.

[0015] The beneficial effects of this invention are as follows: First, it achieves intelligent image restoration with semantic constraints: through a memory-guided content reconstruction algorithm, it transforms the user's memory information into restoration constraints, thereby overcoming the key bottleneck of traditional image restoration methods lacking semantic understanding. This algorithmic breakthrough ensures that the image restoration result is no longer random pixel filling, but content reconstruction based on semantic logic, thus significantly improving the accuracy and reliability of the restored content.

[0016] Secondly, a personalized memory decay prediction model was established: by using an improved Ebbinghaus forgetting curve model combined with individual user learning characteristics, the precise quantification and prediction of memory states were achieved. This innovative algorithm can scientifically assess the reliability of user memory information, providing an objective constraint for image restoration and thus avoiding the blind use of subjective memory information.

[0017] Third, a temporal memory verification algorithm was constructed: by analyzing the temporal evolution trajectory of user memory, the algorithm verifies the temporal consistency between the candidate repair solutions and the memory state, ensuring that the repaired content conforms to the user's cognitive patterns. This verification algorithm effectively solves the problem of discrepancies between the repaired content and the user's actual memory state, thus improving the acceptability of the repair results.

[0018] Fourth, it achieves integrated processing of restoration and classification: the image restoration process is optimized in a unified manner with the subsequent content classification task. The restoration stage considers classification requirements, and the classification stage provides feedback on restoration quality, thus forming a closed-loop optimization system. This integrated processing method avoids the information loss caused by traditional separate processing, thereby improving the overall system performance and efficiency.

[0019] Fifth, it provides technological innovation based on cognitive science: for the first time, it deeply integrates the memory theory of cognitive science with computer vision technology, pioneering a new technological path for image processing based on the user's cognitive state. This innovation not only solves the specific problem of repairing damaged English flashcards, but also provides a theoretical foundation and technological paradigm for other image processing tasks that require combining user cognitive information. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 It is a line graph validating the effectiveness of the personalized forgetting curve model; Figure 3 This is a grouped bar chart comparing the performance of image restoration methods; Figure 4 It is a heatmap of the vocabulary difficulty-user level repair success rate matrix; Figure 5 It is a radar chart for a multi-dimensional performance radar comparative analysis of the system; Figure 6 It is a stacked area diagram showing the evolution trend of time series consistency scores. Detailed Implementation

[0021] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0022] Before providing a detailed description of this application, the terms used in this application will be explained below to aid in understanding the technical solutions: English flashcards: These are learning cards that contain English words, phrases, and their Chinese translations, and are typically used for English vocabulary learning and memorization training.

[0023] Damaged area: refers to the area where the image content is missing, blurred, or incomplete due to physical wear, tearing, fading, or other reasons.

[0024] Memory-guided content reconstruction algorithm: The innovative algorithm proposed in this application analyzes the user's memory state and forgetting patterns, and uses memory information as a constraint to guide the intelligent restoration process of image content.

[0025] Forgetting Curve Fitting Algorithm: A mathematical model based on the Ebbinghaus forgetting curve theory and individual learning data to predict the degree of memory decay in users.

[0026] Memory reliability coefficient: A numerical parameter that quantifies the accuracy of a user's memory of specific content. It ranges from 0 to 1 and is used to constrain content generation during the image restoration process.

[0027] Temporal memory verification algorithm: This algorithm verifies the consistency between the repaired content and the user's memory state by analyzing the relationship between the time node where the user's memory was formed and the current time.

[0028] Semantic probability distribution: A probability distribution matrix of the contents that the damaged area may contain, calculated based on lexical features and language models.

[0029] Application scenario description The technical solution of this application is mainly applied to the fields of intelligent education and language learning, with specific application scenarios as follows: Implementation Environment: This application can be implemented in a computer system equipped with image acquisition devices (such as cameras and scanners), including personal computers, mobile devices, and tablets. The system needs to have image processing capabilities, data storage functions, and a user interface, capable of acquiring and processing the user's historical learning records.

[0030] Main application scenarios: Users frequently use English flashcards for vocabulary memorization and review during their long-term English learning process. Due to frequent use and time factors, these English flashcards may suffer the following damage: folded corners causing missing text, faded ink making the text illegible, torn paper causing incomplete content areas, and liquid contamination making the text unreadable.

[0031] Specific usage process: Users acquire digital images of damaged English flip cards using an image acquisition device. The system automatically identifies the location and type of damage. Simultaneously, the system accesses the user's historical learning data, including the frequency of use, study time, and review records of the English flip card. Based on this data, the system analyzes the user's memory status of the English flip card content and performs intelligent repair based on the image information. After repair, the system identifies and categorizes the repaired English flip card, providing users with personalized learning suggestions and review schedules.

[0032] System environment requirements: The system needs to integrate an image processing module, a machine learning module, a database management module, and a user interface module.

[0033] The image processing module consists of an image acquisition unit, a preprocessing unit, an edge detection unit, and an image inpainting unit. The image acquisition unit acquires digital images from a camera or scanner; the preprocessing unit performs image denoising, size normalization, and color correction; the edge detection unit uses the Canny algorithm to identify the boundaries of damaged areas; and the image inpainting unit executes a memory-guided content reconstruction algorithm. These units are connected via an image data pipeline to ensure sequential data processing and consistent formatting.

[0034] Furthermore, the specific processing method of the image preprocessing unit is as follows: First, image denoising employs a bilateral filtering algorithm, with the filtering function being... ,in Weighting function based on spatial distance , Weighting function based on pixel value differences Spatial standard deviation grayscale standard deviation , The first step is to use a normalization constant; the second step is to standardize the size by adjusting the input image to a fixed resolution. For each pixel, a bilinear interpolation algorithm is used, and the interpolation formula is: ,in The interpolation point coordinates are used; the third step, color correction, employs white balance adjustment and gamma correction. White balance is adjusted using the grayscale world assumption to adjust the RGB channels, and the adjustment formula is... The gamma correction formula is ,in Using standard display gamma values, this preprocessing process ensures that the input images have uniform quality standards and format specifications, providing an optimal data foundation for subsequent edge detection and repair algorithms.

[0035] The machine learning module consists of a personalized memory decay prediction model, a user memory time series model, a semantic probability calculation unit, and a time series memory verification unit. Data interaction between these units is achieved through feature vector transmission.

[0036] Detailed Description of the Personalized Memory Decline Prediction Model: This is the core innovative model of this application. Based on the improved Ebbinghaus forgetting curve theory, it incorporates quantitative analysis of individual differences and the review effect. The model architecture includes a parameter learning layer, a memory state prediction layer, and a reliability assessment layer. The parameter learning layer automatically learns individualized parameters by analyzing the user's historical learning data using the least squares method and gradient descent algorithm. , and review effect coefficient The memory state prediction layer calculates the memory retention rate at any given time point based on the learned parameters; the credibility assessment layer combines contextual information and time decay characteristics to output the final memory credibility coefficient. The model's training process employs incremental learning, continuously optimizing prediction accuracy as the user progresses. In practical applications, this model can accurately predict a user's memory state of specific words at the current moment based on their past learning performance, providing a scientific constraint for image restoration.

[0037] Detailed Description of the User Memory Temporal Model: This application presents a unique temporal modeling technique to describe the dynamic changes in a user's memory state over time. The model employs a combination of piecewise linear interpolation and cubic spline interpolation to construct a continuous memory trajectory curve. The model structure includes a data preprocessing layer, an interpolation calculation layer, and a smoothing optimization layer. The data preprocessing layer cleans and standardizes the user's historical memory test data; the interpolation calculation layer constructs a continuous function between known discrete time points to ensure the smoothness of the memory trajectory; the smoothing optimization layer reduces curve oscillations by minimizing the integral of the second derivative. The innovation of this model lies in the introduction of a memory reinforcement point identification algorithm, which can automatically detect time points where the user's memory is significantly enhanced (such as review times) and adjust the curve shape at these points. During temporal memory verification, this model provides an estimate of the user's memory state at any given time point, making it possible to evaluate the temporal consistency of repair candidate solutions.

[0038] The database management module consists of the following components: a user learning record database, a vocabulary knowledge base, an image sample library, and a model parameter library. The user learning record database stores an individual's learning history, test results, and review records; the vocabulary knowledge base contains information on the difficulty level, semantic relationships, and usage frequency of English vocabulary words; the image sample library stores standard image templates and font styles for English flashcards; and the model parameter library stores each user's personalized model parameters and configuration information. Relational connections between the databases enable data association queries and maintain consistency.

[0039] The user interface module consists of an image upload unit, a results display unit, a user feedback unit, and a parameter adjustment unit. The image upload unit provides drag-and-drop or file selection functionality; the results display unit presents the restored image and classification results in a visual manner; the user feedback unit collects user evaluations of the restoration quality; and the parameter adjustment unit allows users to customize key parameters such as memory reliability. Each unit responds to user actions and synchronizes its status through a message passing mechanism.

[0040] The method of this embodiment includes the following steps. This embodiment proposes a direct-reading English flashcard classification method based on image recognition processing, which includes the following steps: Step 1: Image Acquisition and Damaged Area Identification According to embodiments of this application, an image of a damaged English flashcard and the user's historical learning records are obtained, an edge detection algorithm is used to identify the boundary of the damaged area, and the user's individual forgetting pattern features are extracted.

[0041] Specifically, the system acquires a digital image of the damaged flip card using an image acquisition device. In some implementations, the Canny edge detection algorithm may be optionally used to detect discontinuous boundaries in the image and identify the contour of the damaged area. Furthermore, the damaged area identification may optionally employ a gradient analysis-based method, locating the damaged boundary by calculating abrupt changes in pixel gradients.

[0042] It should be understood that the extraction of individual user forgetting pattern features involves the system extracting the user's learning records from the user's historical learning database, such as vocabulary learning time, review frequency, and answer accuracy. Based on this data, a forgetting pattern feature vector for each individual user is constructed. ,in Indicates the learning time interval. Indicates the frequency of review. Indicates historical accuracy. Indicates the level of vocabulary difficulty.

[0043] Data Preprocessing Explanation: Since the features in the forgetting pattern feature vector have different dimensions and numerical ranges, the system uses the Z-score standardization method to preprocess the feature vector. Specifically, for each feature dimension, the mean and standard deviation of that feature in the user's historical data are calculated. Then, the current feature value is subtracted from the mean and divided by the standard deviation to obtain the standardized dimensionless feature value. This standardization process ensures that features with different dimensions have the same importance weight in subsequent calculations, avoiding biases caused by differences in numerical ranges such as time intervals (in days) and review frequencies (times / day).

[0044] Step 2: Forgetting curve fitting and memory reliability calculation According to an embodiment of this application, based on a forgetting curve fitting algorithm, the degree of memory decay and the memory reliability coefficient of the user's current damaged English flipbook content are calculated.

[0045] It should be noted that a personalized memory decay prediction model is established by using an improved Ebbinghaus forgetting curve model and incorporating individual difference parameters. in, Indicates time Memory retention rate afterward and For individualized parameters, For the first The effect coefficient of the second review. Indicates the first The reinforcement effect of the second review This refers to the number of times you need to review.

[0046] Furthermore, to ensure the mathematical validity of the memory retention rate, each parameter in the model has explicit constraints: memory retention rate The value of is strictly limited to the interval [0,1], and when the calculated result exceeds 1, it is automatically truncated to 1; individualized parameter The range of values ​​for the parameter is constrained to (0,2], and the parameter... The range of values ​​for is constrained to (0,3], which ensures that the forgetting curve has reasonable decay characteristics; the review effect coefficient The value range is constrained to [0,1], enhancing the effect. The value range is constrained to [0,1], when the number of reviews When the number of reviews exceeds 20, the system only considers the effect of the most recent 20 reviews to avoid computational overflow.

[0047] Furthermore, time variables It has clear constraints and physical meaning: time variable The range of values ​​for is strictly limited to . (non-negative real numbers), where Indicates the initial learning moment. The unit is day, with a minimum time precision of 0.1 day (i.e., 2.4 hours); to ensure the stability of numerical calculations, when When the time exceeds 365 days, the system automatically invokes the long-term memory model for processing; the time variable is reflected in the product term of the review effect as the review time series. ,in Indicates the first The timing of this review must meet the following requirements. The strict increasing relationship ensures the logical consistency of review time and the temporal continuity of memory decay.

[0048] The memory reliability coefficient is calculated based on this model. The calculation process is as follows: obtain the memory retention rate (representing the strength of a user's memory of specific content at the current time point, a dimensionless ratio between 0 and 1) and the context weight adjustment factor (reflecting the influence of content features and user features on memory assessment, usually between 0.5 and 1.5). Multiply these two values, and the result is the memory reliability coefficient. This coefficient comprehensively reflects the combined influence of time decay effect and individual differences on the accuracy of user memory, and is used to determine the constraint weights in the subsequent image restoration process.

[0049] It should be noted that the specific calculation method of the context weight adjustment factor is as follows: First, calculate the base weight based on the content characteristics of the current English flashcard (such as vocabulary type, grammatical complexity, semantic abstraction level, etc.), and then adjust it based on the user's mastery of this type of content in their learning history. The specific calculation process is as follows: Obtain the base weight value (usually set as a floating-point value between 0.8 and 1.2), and simultaneously obtain the user's mastery level value of the same type of content (a dimensionless ratio between 0 and 1) and the normalized value of the usage frequency of this type of content (a dimensionless frequency index). Then, multiply the mastery level value with its corresponding adjustment coefficient of 0.3, and multiply the usage frequency value with its corresponding adjustment coefficient of 0.2. Add the two multiplication results and add 1. This result reflects the adjustment factor of the user's characteristics on the base weight. Finally, multiply the base weight with the adjustment factor to obtain the final context weight adjustment factor, which quantifies the degree of influence of the user's individual characteristics on the reliability of memory.

[0050] Furthermore, the specific calculation method for the base weight values ​​is as follows: A multi-factor scoring model is adopted, with the following steps: First, the content features of the English translation cards are quantitatively evaluated, including vocabulary type weights (nouns 0.9, verbs 1.0, adjectives 0.8, adverbs 0.7), grammatical complexity scores (calculated based on sentence length and clause level, ranging from 0.5 to 1.5), and semantic abstraction scores (concrete nouns 0.7, abstract concepts 1.3, action verbs 1.0); Second, the scores of each feature are weighted and summed, with the weights allocated as follows: vocabulary type 40%, grammatical complexity 35%, and semantic abstraction 25%; Third, the weighted sum is mapped to the 0.8-1.2 range using the Sigmoid function, with the mapping formula being BaseWeight = 0.8 + 0.4 × (1 / (1 + ... e^(-2×(score-1)))), where score is the weighted sum of feature scores; the fourth step is to fine-tune according to the subject area to which the content belongs, keeping the original value of general vocabulary, multiplying professional terms by 1.1 times the adjustment coefficient, and multiplying colloquial expressions by 0.9 times the adjustment coefficient; the fifth step is to output the final basic weight value, which objectively reflects the learning difficulty and memory complexity of the English flip card content.

[0051] The specific implementation of the forgetting curve fitting algorithm: According to an embodiment of this application, the algorithm optionally employs the least squares method to fit the user's individualized forgetting parameters. First, the system collects the user's historical learning and testing data, such as the time points of each learning session, review time points, and corresponding memory test accuracy rates. Then, through nonlinear regression analysis, the individualized parameters are determined. and The optimal value. It should be noted that the parameter... Parameters that reflect the user's baseline forgetting rate Controlling the shape of the forgetting curve. Additionally, the review effect coefficient. This was determined by analyzing the relationship between review frequency and memory retention rate, showing that frequently reviewed content has a higher retention rate. value.

[0052] Furthermore, individualized parameters and The specific calculation method is as follows: The Levenberg-Marquardt optimization algorithm is used to solve the nonlinear least squares problem. The specific implementation steps are as follows: First, the user's historical learning data is constructed into a time series. ,in Indicates the first time after learning At a certain point in time, The first step is to represent the memory retention rate measured at that time point; the second step is to construct the objective function. This function measures the fitting error between the theoretical forgetting curve and the actual measurement data; the third step is to calculate the objective function's effect on... and The partial derivatives are used to obtain the Jacobian matrix; the fourth step is to update the parameter values ​​using the damped least squares method, with the update formula being... ,in The damping coefficient is... For Jacobian matrices, The fifth step is to repeat the iteration until the objective function value converges or the maximum number of iterations (100) is reached, and finally the optimized value is output. and Parameter value.

[0053] Furthermore, the review effect coefficient The specific calculation method is as follows: A combination of correlation analysis and regression fitting is used, with the following steps: First, extract review-related information from the user's historical learning data, including review intervals, review duration, review methods (active recall vs. passive reading), and post-review test performance; Second, calculate the Pearson correlation coefficient between each review session and the degree of improvement in subsequent memory retention. The correlation coefficient calculation formula is... ,in As a review intensity indicator, The third step involves determining the basic weight of the review effect based on the magnitude and significance of the correlation coefficient; the higher the correlation coefficient, the greater the weight. The fourth step considers the timing of the review and adjusts the weights using the forgetting curve theory; the closer the review time is to the optimal forgetting point, the higher the effect coefficient. The fifth step uses exponential smoothing to calculate a weighted average of the historical review effects, using the following formula: ,in The system scores the immediate effect of the current review session and ultimately outputs a personalized sequence of review effect coefficients.

[0054] Application example in a specific scenario: Suppose user Zhang learned the word "sophisticated" on January 1, 2024, and reviewed it three times on January 3, January 7, and January 15, with accuracy rates of 90%, 85%, and 80% respectively. The current time is January 25, and the user uploaded a damaged English flashcard for the word. The system first calculates the time variable. Days, then based on the user's historical data, fit the results. , The review effect coefficients are respectively , , The current memory retention rate was finally calculated. Combined with context weights The memory reliability coefficient is obtained. .

[0055] Furthermore, to ensure consistency in variable naming, all time-related variables adopt the same naming convention throughout the entire technical solution. To represent time variables, avoid using different variable names such as t_current and time to prevent confusion; all weight coefficients should be named with the prefix W_, such as W_context for context weight; all coefficient parameters should be named with the prefix C_, such as C_memory for memory reliability coefficient, to ensure the standardization and consistency of variable naming in technical documents.

[0056] Step 3: Memory-guided content reconstruction According to an embodiment of this application, a memory-guided content reconstruction algorithm is used, with the memory credibility coefficient as a constraint weight, to adjust the content generation parameters of the image restoration algorithm.

[0057] It should be noted that the algorithm provided in this application is as follows: Traditional image restoration algorithms rely solely on surrounding pixel information, while this application innovatively transforms user-remembered information into restoration constraints. The specific implementation process is as follows: Step 3-1: Establish the memory constraint matrix ,in It is a semantic constraint vector constructed based on user-remembered content.

[0058] Step 3-2: Modify the objective function of traditional image inpainting by adding a memory constraint term: in, For the repaired image, The original damaged image, For time-related memory constraint weighting coefficients, This is a time-dependent memory constraint matrix.

[0059] Furthermore, the time-related constraint weighting coefficient The specific calculation method is as follows: ,in The basic constraint weights (value range [0.1, 2.0]) For time The memory retention rate (calculated by the forgetting curve model). The time confidence adjustment factor (calculated using the following formula) ),in The optimal time for repair is usually within 1-3 days after learning. Using a time-sensitive parameter (typically set to 2.0 days), this design ensures that the constraint weights decay over time in accordance with the forgetting curve, with the strongest constraint effect near the optimal repair time; Time-dependent memory constraint matrix. The calculation method is as follows: ,in The reliability coefficient of time-related memory. The semantic constraint vector takes into account time decay. The element weights of this vector are dynamically adjusted according to the time distance to ensure that the repair constraints are consistent with the time of the user's current memory state.

[0060] Furthermore, memory constraint weight coefficients It has strict value constraints: The value range is limited to the interval (0,2]. When the memory reliability coefficient is high (>0.8), The value range is [0.6, 1.2] to ensure the dominant role of memory constraints; when the memory reliability coefficient is moderate (0.4-0.8), The value range is [0.4, 0.8] to balance visual consistency and memory constraints; when the memory reliability coefficient is low (<0.4), The value range is [0.1, 0.4] to avoid the negative impact of unreliable memory information on the repair results.

[0061] Step 3-3: Solve the above objective function using an iterative optimization algorithm to generate the image restoration result of memory-guided content reconstruction.

[0062] It should be noted that the specific implementation of the iterative optimization algorithm is as follows: the objective function is minimized using gradient descent. The specific iterative process is as follows: First, the image data (a numerical array stored in the form of a pixel matrix) and the learning rate parameter (usually set to a positive value between 0.01 and 0.1 to control the parameter update step size) of the current iteration are obtained. At the same time, the gradient of the objective function with respect to the image is calculated. The gradient calculation process is as follows: the current image is subtracted from the original damaged image to obtain the pixel difference matrix, which is multiplied by a constant 2 to obtain the first gradient component. The current image is subtracted from the memory constraint matrix to obtain the constraint difference matrix, which is multiplied by a constant 2 and the memory constraint weight coefficient to obtain the second gradient component. The two gradient components are added to obtain the total gradient vector. Then, the learning rate and the total gradient vector are multiplied, and the product is subtracted from the current image to obtain the image data for the next iteration. This update process adjusts the image in the direction of decreasing the objective function value. The iteration termination condition is that the change in the objective function value is less than a preset threshold of 0.001 for three consecutive iterations or the maximum number of iterations of 200 is reached. The image at the termination is the optimal repair result.

[0063] Furthermore, the threshold setting in the iteration termination condition has a clear theoretical basis: the convergence threshold of 0.001 is chosen based on the image restoration accuracy requirements, and this value corresponds to a single pixel grayscale value change of less than 0.255 (i.e., one-thousandth of an 8-bit image), ensuring the visual stability of the restoration result; the maximum number of iterations of 200 is set based on the balance of computational complexity. According to empirical statistics, more than 95% of restoration tasks converge within 200 iterations. This limitation avoids the algorithm from infinitely looping in ill-conditioned cases; convergence is determined when the objective function value changes less than the threshold for three consecutive times. This condition ensures that the algorithm has reached a local optimum rather than a temporary numerical fluctuation.

[0064] The specific implementation of the memory-guided content reconstruction algorithm is as follows: According to an embodiment of this application, the algorithm realizes the numerical representation of memory information by constructing a semantic constraint vector. Specifically, the system first converts the user's fuzzy memory content into a set of candidate characters, and assigns a weight value based on memory credibility to each candidate character.

[0065] Furthermore, the specific method for converting user's fuzzy memory content into a candidate character set is as follows: a matching algorithm based on edit distance and speech similarity is adopted, and the specific implementation steps are as follows: The first step is to obtain vague memory descriptions input by users through the user interface, such as "long words that start with soph" or "words that contain the ending icated". The second step is to parse the fuzzy description into structured query conditions and extract information such as key character fragments, position constraints, and length ranges. The third step is to perform a fuzzy matching query in the vocabulary database and use the Levenshtein edit distance algorithm to calculate the similarity between candidate words and memory fragments. Words with an edit distance of less than or equal to 2 are included in the candidate set. The fourth step is to use the Soundex speech coding algorithm to calculate speech similarity and include words with similar speech into the candidate set. The fifth step is to assign weight values ​​to candidate characters based on the memory credibility coefficient. The weight calculation formula is as follows: ,in For the first Similarity score of candidate characters This represents the frequency of the character in the user's learning record.

[0066] Decoding and conversion of semantic constraint vectors to pixel constraints: The system uses character template matching and pixel mapping algorithms to convert semantic constraint vectors into specific pixel-level constraint matrices. The specific process is as follows: First, standard character templates are extracted from the font library, and the glyph information of each candidate character is converted into a binary pixel matrix; then, based on the character probability weights in the semantic constraint vector, the corresponding character template pixel matrices are weighted and fused; finally, the fused pixel matrix is ​​normalized to a grayscale value range of 0-255 to form the final memory constraint matrix, which is directly used to calculate the constraint terms of the image inpainting objective function.

[0067] In some implementations, the iterative optimization process employs gradient descent, considering both visual consistency and semantic constraints of the image in each iteration, by adjusting the weight coefficients. To balance the importance of both. It should be noted that when the change in the objective function value is less than the preset threshold of 0.001 for three consecutive iterations, the algorithm converges and outputs the final repair result.

[0068] Application example in a specific scenario: Continuing from the previous example, in user Zhang's "sophisticated" card, the "soph" part is clearly visible, but the "isticated" part is severely damaged. Based on the memory reliability coefficient... The system determines that the user has a moderate level of memory for the word. During the semantic constraint vector construction process, the system assigns probability weights to the positions of each letter in "isticated": 'i' (0.85), 's' (0.80), 't' (0.75), 'i' (0.78), 'c' (0.82), 'a' (0.70), 't' (0.73), 'e' (0.68), 'd' (0.65). Memory constraint weight coefficients. The value was set to 0.65 to match the memory reliability coefficient. After 15 iterations of optimization, the algorithm successfully reconstructed the complete image of the word "sophisticated," with the reconstructed "isticated" portion visually consistent with the original font style.

[0069] Step 4: Semantic probability distribution calculation and repair candidate generation According to embodiments of this application, based on vocabulary features and learning difficulty data, the semantic probability distribution of possible content in the damaged area is calculated, and a repair candidate scheme is generated by combining memory constraints.

[0070] It should be noted that, unlike traditional content filling methods based on pixel similarity, this application predicts damaged content based on linguistic features and learning data.

[0071] Step 4-1: Retrieve a set of candidate words related to the user's learning history from the vocabulary database, and establish a semantic association graph of the candidate words based on the context and word co-occurrence frequency.

[0072] It should be understood that the semantic association graph is constructed as follows: First, based on the visible character fragments (such as "soph") in the damaged area, all candidate words containing the fragments are retrieved from the vocabulary database; then, the co-occurrence probability of each candidate word with the context is calculated. The co-occurrence probability calculation process is as follows: count the number of times the candidate word and the specific context appear simultaneously (integer frequency value), and simultaneously count the total number of times the context appears alone (integer frequency value). Divide the number of simultaneous occurrences by the total number of occurrences in the context. The result of the division operation is the co-occurrence probability value (a dimensionless ratio between 0 and 1). This probability reflects the association strength between the word and the context; finally, a weighted undirected graph is constructed based on the co-occurrence probability and semantic similarity. The nodes in the graph are candidate words. The edge weight calculation process is as follows: obtain the co-occurrence probability value of the candidate word and the semantic similarity value between words (both are dimensionless values ​​between 0 and 1). Multiply the co-occurrence probability by a weight coefficient of 0.6 and the semantic similarity by a weight coefficient of 0.4. Then, add the two products. The result is used as the weight value of the corresponding edge in the graph. This weight comprehensively reflects the statistical association and semantic association strength between words.

[0073] Furthermore, the specific definition and acquisition method of the context environment are as follows: the sliding window technique is used to extract the context information of the English flashcards, and the specific implementation steps are as follows: The first step is to identify the coordinates of the damaged area in the English translation card and determine the boundaries of the text before and after the damaged text. The second step is to set a sliding window range of 5 words before and after the damaged area as the center, and extract all visible text content within the window as the context. The third step is to perform part-of-speech tagging and syntactic analysis on the extracted context text to identify part-of-speech information such as nouns, adjectives, and verbs, as well as syntactic structures such as subject, verb, and object. The fourth step is to construct a contextual feature vector, including features such as word frequency statistics, N-gram patterns, and syntactic dependencies; The fifth step involves retrieving text fragments containing the same or similar contextual patterns from a large-scale corpus, statistically analyzing the frequency of each candidate word within these contextual patterns, and calculating the co-occurrence probability statistics. This method ensures the accuracy of the contextual definition and the reproducibility of the co-occurrence probability calculation.

[0074] Step 4-2: Calculate the semantic relevance of each candidate word in the current context using the improved TF-IDF algorithm: in, Candidate words, For the semantic similarity between words and their context, To determine the match between vocabulary and the user's learning level, This is a probability factor for the recurrence of time-related words.

[0075] Furthermore, the probability factor for the recurrence of time-related words. The specific calculation method is as follows: ,in For vocabulary In time The memory recall probability (calculated based on the individual forgetting curve model, with a value range of [0,1]). The time weighting function (calculated using the formula is) ,in This is the initial learning time. (The maximum effective memory time window is typically set to 30 days). This factor ensures that semantic relevance calculations fully consider the impact of time on the user's memory state. Vocabulary recall is higher in the early stages of learning, gradually decreasing over time, consistent with the forgetting curve. At that time, the system will automatically Setting a minimum threshold of 0.1 avoids repair failures caused by long-term forgetting. This constraint ensures the stability and reliability of the algorithm at different time scales.

[0076] Step 4-3: Based on the memory constraints obtained in Step 3, adjust the probability distribution of candidate words to generate the final set of repair candidate solutions.

[0077] Specific implementation of the improved TF-IDF algorithm: According to the embodiments of this application, two key improvements are made to the traditional TF-IDF algorithm. First, a semantic similarity factor is optionally introduced. This factor is obtained through cosine similarity calculation of word vectors, which can capture the deep semantic relationship between words and context, rather than simply relying on word frequency statistics. Secondly, a user learning level matching factor is added. This factor is calculated based on the user's historical learning data and vocabulary difficulty rating, ensuring that the recommended candidate words match the user's current learning level.

[0078] Furthermore, the term frequency function The specific calculation method is as follows: ,in Words In the document The number of times it appears in Document The formula calculates the relative frequency of a word in a specific document by summing the occurrences of all words in the document. To avoid high-frequency words excessively influencing the calculation results, the system employs a logarithmic transformation, with the transformation formula being: ,when Time setting Finally, the values ​​are mapped to the [0,1] interval through Min-Max normalization.

[0079] Furthermore, the inverse document frequency function The specific calculation method is as follows: ,in This indicates the total number of documents in the corpus. Indicates included words The number of documents is used to calculate the rarity of a word in the entire corpus. A higher value indicates a rarer word. To prevent the denominator from being zero, the denominator is set to 1 when the word does not appear in any document. Finally, normalization is performed by dividing by the maximum IDF value in the corpus to ensure that all IDF values ​​are within a reasonable range.

[0080] Furthermore, the memory recall probability function The specific calculation method is as follows: ,in For the time calculated based on the forgetting curve model Memory retention rate For vocabulary User familiarity coefficient (calculated based on historical learning records, with a value range of [0,1]). The strength of the time-related contextual cue (reflecting the activation level of memory for the word in the current context of the damaged card, with a value range of [0,1]); the formula for calculating the user familiarity coefficient is as follows: ,in For users to access vocabulary The number of times it was correct in history The smoothing process of adding 1 and 2 to the total number of encounters avoids the problem of zero probability for new words.

[0081] Furthermore, the time-related contextual cue intensity function The specific calculation method is as follows: ,in The total number of words visible around the damaged area. For the first Contextual vocabulary, The semantic similarity between the target word and the context words is calculated using cosine similarity. For the position index of context words, This is the location index of the damaged area. This is the distance attenuation parameter (usually set to 0.2, which controls the degree of influence of distance on the alert intensity). The time decay factor (calculated using the formula is) ,in (This is the time decay coefficient, usually set to 0.01). This function is designed to ensure that contextual words that are closer to the damaged area and have a higher semantic relevance contribute more to memory activation. It also takes into account the decay effect of time on the effect of contextual cues, which is consistent with the memory cueing theory in cognitive psychology.

[0082] Data preprocessing for each factor in the algorithm: Due to the different numerical ranges of the terms in the TF-IDF formula, the system uses the Min-Max normalization method to standardize each factor. The specific process is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The values ​​are normalized to the 0-1 range after logarithmic transformation. Normalization is performed by dividing the value by the maximum IDF value, resulting in semantic similarity. The value is linearly mapped from the original range of -1 to 1 to the range of 0 to 1, indicating a user match degree. No further processing is needed since the values ​​are already within the 0-1 range. This normalization preprocessing ensures that the four factors have the same weight contribution in the multiplication operation, avoiding the bias caused by different numerical ranges in the semantic relevance calculation results.

[0083] In some implementations, the calculation process for the user's learning level matching degree is as follows: First, obtain the difficulty level value of the candidate words (usually an integer between 1 and 10, representing the learning difficulty level of the words) and the user's current learning level value (also a value between 1 and 10, representing the user's English proficiency). Calculate the absolute difference between the two values, divide this difference by the smoothing parameter value (usually set to 2.0, used to adjust the sensitivity of the matching degree function), take a negative sign for the division result, and then calculate the natural exponential function value of the negative value. The result of the exponential function calculation is the matching degree value, which ranges from 0 to 1. The closer the value is to 1, the better the difficulty of the candidate words matches the user's level; the closer it is to 0, the lower the matching degree. This matching degree reflects the degree of adaptation between the recommended content and the user's cognitive ability.

[0084] Furthermore, the smoothing parameter has strict constraints: the range of the smoothing parameter is limited to [1.0, 5.0]. When set to 1.0, the matching degree function is highly sensitive to differences in difficulty, which is suitable for precise matching scenarios; when set to 2.0, it provides standard matching sensitivity, which is suitable for most application scenarios; when set to 5.0, the matching degree function is more tolerant of differences in difficulty, which is suitable for scenarios where the user's level assessment is uncertain; values ​​outside this range will cause the matching degree function to lose its discriminative ability or become overly sensitive, affecting the accuracy of candidate word selection.

[0085] It should be noted that the semantic similarity function is implemented as follows: First, a pre-trained word vector model (such as Word2Vec or GloVe) is used to convert candidate words and context words into high-dimensional vector representations. Then, the semantic association is calculated using cosine similarity. The specific calculation process is as follows: Obtain the vector representation of the candidate words (an N-dimensional floating-point array, usually N is 100-300 dimensions) and the vector representation of the context (obtained by weighted averaging of the vectors of all words in the context, with weights determined based on the TF-IDF values ​​of the words, also an N-dimensional floating-point array). Calculate the dot product of the two vectors (i.e., multiplying the corresponding elements one by one and then summing them). At the same time, calculate the magnitude of each vector (i.e., the square root of the sum of the squares of the elements in each dimension). Divide the dot product result by the product of the two magnitudes. The result of this division operation is the semantic similarity value, which is a dimensionless value ranging from -1 to +1. The closer the value is to +1, the higher the semantic association between the two words; the closer it is to -1, the stronger the semantic opposition; and the closer it is to 0, the less semantically related the two words are.

[0086] Application example in a specific scenario: Continuing with the previous example of "sophisticated", the system retrieved candidate words related to "soph" including "sophisticated", "philosophy", and "sophomore". For "sophisticated", the TF value is 0.8 (high frequency in the user's learning records), the IDF value is 6.2 (relatively rare in a general corpus), the SIM value is 0.92 (highly relevant to the context of "advanced vocabulary"), and the Diff value is 0.78 (matches the user's intermediate to advanced level). The final semantic relevance calculation process is as follows: the TF value of 0.8, the IDF value of 6.2, the semantic similarity value of 0.92, and the user matching value of 0.78 are multiplied sequentially. The result of multiplying these four values ​​consecutively is 3.55, which quantifies the comprehensive score of the candidate word across multiple dimensions. In comparison, “philosophy” and “sophomore” obtained comprehensive scores of 1.23 and 0.89 respectively using the same calculation method. By comparing the scores of the three candidate solutions, “sophisticated”, which had the highest score, was identified as the highest probability repair candidate solution.

[0087] Step 5: Temporal memory verification and optimal selection According to an embodiment of this application, a temporal memory verification algorithm is used to analyze the matching degree between candidate repair solutions and the user's forgetting time points, and to select the optimal repair content.

[0088] It should be understood that by analyzing the temporal characteristics of user memories, the temporal consistency of repair candidate solutions can be verified.

[0089] Step 5-1: Establish a user memory time series model to analyze the trajectory of changes in the user's mastery of the word at different time points.

[0090] Step 5-2: Calculate the temporal consistency score between each repair candidate and the user's current memory state: in, For the first One candidate solution, As time weight, Candidate solutions and time points The degree of matching of memory states, For time step.

[0091] Furthermore, the boundary conditions of the summation interval have explicit definitions and constraints: start time Defined as the time point at which a user initially learns the vocabulary. ), End Time Defined as the current time point when a user submits a damaged card for repair, it must meet the following conditions: Time order constraints; time step Setting it to 0.5 days ensures that the time-series consistency score has sufficient temporal resolution when evaluating the time window. The system automatically adjusts the step size to 1.0 day to improve computational efficiency. During the summation process, only the time points when the user actually generates memory data are considered, i.e., the time points when there are learning or review records. For time points without data, an interpolation method is used to estimate the memory state. To avoid excessive computational complexity, when the time window exceeds 180 days, the system adopts a segmented weighted summation method, dividing the time axis into several sub-intervals. The average value is used to represent the memory state of each sub-interval. This constraint ensures the computational feasibility and numerical stability of the algorithm over long time spans.

[0092] Step 5-3: Based on the combined semantic probability and temporal consistency score, select the optimal repair content.

[0093] The specific implementation of the temporal memory verification algorithm: According to the embodiments of this application, the algorithm realizes temporal consistency verification by constructing a time trajectory diagram of the user's memory state. The specific process includes: First, the system collects data on the user's mastery of the target vocabulary at different time points, such as learning test results, review performance, and forgetting test records. Then, optionally, a continuous memory state curve is constructed using cubic spline interpolation, which reflects the change law of the user's memory intensity over time. In some implementations, the time weight is calculated using an exponential decay function, and the calculation process is as follows: First, obtain the time difference between the current time and the historical time point (a positive value in days), multiply the time difference by the time decay parameter (usually set to a positive value between 0.1 and 0.3 to control the decay rate), take the negative sign of the product, and then calculate the natural exponential function value of the negative value. The result of the exponential function calculation is the time weight value, which ranges from 0 to 1. The closer the value is to 1, the higher the importance of the corresponding time point; the closer it is to 0, the lower the importance due to the long time. This weight reflects the influence of time distance on the importance of memory verification. In addition, the matching degree function... The similarity between candidate schemes and historical memory states is obtained by comparing the similarity between them. The higher the similarity, the closer the matching degree is to 1.

[0094] Furthermore, the specific implementation method for constructing continuous memory state curves using cubic spline interpolation is as follows: A cubic spline interpolation algorithm with natural boundary conditions is employed, and the specific calculation steps are as follows: The first step is to arrange the collected discrete memory data points in chronological order to form a data sequence. ,in As a time node, This corresponds to the memory strength value; The second step is to construct a tridiagonal system of linear equations to solve for the second derivative values ​​at each node. The system of equations is in the form of: ,in It is a tridiagonal matrix. Let be the vector of the second derivative to be found. The vector of terms on the right-hand side; The third step is to use the chasing method (Thomas algorithm) to solve the tridiagonal linear equation system and obtain the second derivative values ​​of each node. The fourth step is to use the cubic spline interpolation formula. Calculate each interval spline function coefficients within , , , ; Step 5: For any query time point The method determines the interval in which the memory trajectory lies and applies the corresponding spline function to calculate the interpolation result, thus obtaining a continuous and smooth memory state curve. This method ensures the continuity and smoothness of the memory trajectory, providing a reliable mathematical basis for temporal consistency verification.

[0095] Furthermore, the specific calculation method for spline function coefficients is as follows: for the interval The formula for calculating the coefficients of the spline function within the range is as follows: (Node function value) (Terms related to the first derivative) (Second derivative term) (Coefficients of the third-order term), where The node spacing (time step, in days). For nodes The second derivative value at point (obtained by solving a system of tridiagonal equations); the natural boundary conditions are set as follows. That is, the second derivative at the first and last nodes is zero, ensuring the natural extension of the spline curve at the boundary; tridiagonal matrix The element composition is: main diagonal elements Top and bottom diagonal elements Right-hand term vector The elements are Solving this system of linear equations ensures the continuity and smoothness of the spline function at all nodes, providing a mathematical guarantee for the time-continuous modeling of memory states.

[0096] Furthermore, the value of the time decay parameter is subject to strict theoretical constraints: the range of the time decay parameter is limited to [0.05, 0.5]. When set to 0.05, it corresponds to slow decay, which is suitable for users with good long-term memory retention; when set to 0.1, it corresponds to standard decay rate, which is suitable for users with general memory characteristics; when set to 0.3, it corresponds to relatively fast decay, which is suitable for users with predominantly short-term memory; when set to 0.5, it corresponds to rapid decay, which is suitable for users with weak memory retention. Exceeding this range will cause the time weighting function to lose its discriminative power or produce numerical calculation anomalies, affecting the accuracy of the temporal consistency score.

[0097] It should be understood that the matching degree function The specific calculation method is as follows: First, the candidate solutions are... Feature vectors and time points The memory state vectors are compared and analyzed. The specific implementation is as follows: ,in The first candidate solution Each feature component For time points The first memory state Each feature component For the first Importance weights of dimensional features For the feature dimension, Adjust parameters for matching degree.

[0098] Furthermore, the dimension of a feature vector is defined and its specific meaning is: feature dimension. The dimension is fixed at 12 to ensure a balance between the completeness of feature representation and computational efficiency; candidate scheme feature vectors The specific components are: The word length feature is (number of characters, normalized to [0,1]). This is a syllable complexity feature (the logarithmic value of the number of syllables, after normalization). The word frequency level feature (based on corpus statistics, with values ​​[0,1]) is used. This is a semantic abstraction feature (a quantitative score of concrete-to-abstract representation, with values ​​[0,1]). This is a one-hot encoding of part-of-speech category features (nouns, verbs, adjectives, etc.). The context matching feature (semantic relevance with surrounding words, with values ​​[0,1]) This is a user familiarity feature (calculated based on historical learning records, with values ​​[0,1]). The learning difficulty feature (standardized difficulty level, with values ​​[0,1]) This represents the memory strength feature (the degree of memory retention at the current time point, with values ​​[0,1]). The frequency feature of review (normalized value of historical review count, taking values ​​[0,1]). The time interval feature (the length of time since the first learning, after normalization) is used. For language model confidence features (predicted probabilities of models such as BERT, with values ​​[0,1]); memory state vector The meanings of each dimension correspond to the feature vectors of the candidate solutions, but the values ​​are based on time points. The specific memory state is calculated; feature importance weights. The allocation is as follows: semantically related features ( The weight is 0.15, and the memory-related features ( The weight is 0.12, and the lexical attribute features ( The weight is 0.08, and the time and context features ( The weight is set to 0.10 to ensure that each feature makes a reasonable contribution to the matching degree calculation.

[0099] Furthermore, the specific data preprocessing methods for each dimension of the 12-dimensional feature vector are as follows: The vocabulary length feature is normalized using Min-Max, and the formula is as follows: ,in (Single-letter word) (Extremely long compound words); The syllable complexity feature first calculates the number of syllables, then takes the logarithm. Finally, by dividing by the largest logarithm Perform normalization; Word frequency ranking features are based on corpus word frequency ranking and use a formula. Directly mapped to the range [0,1]; The part-of-speech category features are encoded using one-hot encoding. The main parts of speech include nouns (1,0,0,0), verbs (0,1,0,0), adjectives (0,0,1,0), and adverbs (0,0,0,1). Then, L2 normalization is used to ensure that the vector magnitude is 1. The frequency characteristics of review are expressed by formula Normalize the saturation function to avoid excessively large values ​​caused by over-review; The time interval feature is normalized using exponential decay, as shown in the formula: The time interval is in days, and 30 days is the decay constant; other characteristics Since the calculation results are already within the range of [0,1], no additional normalization is required. This preprocessing scheme ensures that all features have the same numerical scale and reasonable contribution weight in the matching degree calculation.

[0100] Furthermore, matching degree adjustment parameters The value of has a clearly defined range of constraints: The value range is limited to [0.1, 2.0]. When set to 0.1, the Gaussian function exhibits extremely high sensitivity, allowing only nearly perfect matches to pass verification; when... Setting it to 0.5 provides standard matching sensitivity, balancing matching accuracy and fault tolerance; when... Setting it to 1.0 provides relatively lenient matching conditions, suitable for users whose memory state is not stable; when Setting it to 2.0 provides the highest matching tolerance, making it suitable for scenarios with high uncertainty in the information being remembered. Values ​​outside this range can cause the matching function to be too strict or too lenient, affecting the effectiveness of candidate selection.

[0101] Application example in a specific scenario: In the "sophisticated" repair case, the system analyzed user Zhang's memory trajectory of the word over the past 24 days. The timeline shows: Day 1 memory strength 1.0, Day 3 0.95, Day 7 0.88, Day 15 0.82, Day 24 (current) 0.72. For the candidate solution "sophisticated", the system calculates the matching degree at each time point: , , , , The time weights are as follows: , , , , The final temporal consistency score calculation process is as follows: The time weights at each time point are multiplied by their corresponding matching values ​​to obtain five product results: 0.35 multiplied by 0.98 equals 0.343, 0.47 multiplied by 0.95 equals 0.447, 0.64 multiplied by 0.90 equals 0.576, 0.82 multiplied by 0.85 equals 0.697, and 1.0 multiplied by 0.78 equals 0.78. These five product results are then added together, and the sum is 3.68. This value serves as the temporal consistency score, quantifying the degree of matching between the candidate solution and the user's historical memory trajectory.

[0102] Step 6: Language Model Validation and Result Optimization According to an embodiment of this application, the semantic rationality of the repaired content is verified based on a language model, and the repair result of memory-guided content reconstruction is output.

[0103] In some implementations, a pre-trained language model (such as BERT) may be used to verify the grammatical correctness and semantic coherence of the repaired content. It should be noted that this step is a technique well-known to those skilled in the art, assessing the reasonableness of the repaired content by calculating its perplexity in the language model. Therefore, when the perplexity exceeds a preset threshold, the system returns to step 4 to regenerate candidate solutions.

[0104] Description of the pre-trained language model: BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model developed by Google based on the Transformer architecture. It possesses bidirectional encoding capabilities, allowing it to consider both preceding and following contextual information. This model has been pre-trained on a large-scale text corpus and exhibits powerful language understanding and representation capabilities. This application utilizes the Masked Language Model (MLM) feature of the BERT model, taking the repaired content as the input sequence and evaluating semantic reasonableness by calculating the predicted probability at each word position. The perplexity threshold is set to 15.

[0105] Furthermore, the setting of the perplexity threshold is based on strict theoretical grounds: the perplexity threshold is limited to the range of [5, 50]. When the threshold is set to 5, the semantic rationality requirement is extremely high, and only repair content that highly conforms to language rules is accepted; when the threshold is set to 15, a standard semantic verification strength is provided, balancing semantic accuracy and repair tolerance; when the threshold is set to 30, the semantic verification is relatively lenient, suitable for scenarios with many professional terms or special expressions; when the threshold is set to 50, the semantic tolerance is the greatest, suitable for situations where the repair content may contain rare words or neologisms; a threshold below 5 will lead to excessive strictness and rejection of reasonable repair solutions, while a threshold above 50 will lose the effectiveness of semantic verification.

[0106] Decoding the BERT model output: The system uses a perplexity conversion algorithm to convert the original output of the BERT model into a semantic reasonableness judgment result. The specific process is as follows: First, the repaired text content is input into the BERT model to obtain the predicted probability distribution of each word position; then, the perplexity value of the entire sequence is calculated, with the perplexity calculation formula being the exponent of the average negative log probability of all word positions; next, the perplexity value is compared with a preset threshold of 15. When the perplexity is less than the threshold, a "semantic reasonable" label is output; when the perplexity is greater than or equal to the threshold, a "semantic abnormal" label is output, and a regeneration process is triggered; finally, the binarized semantic reasonableness judgment result is passed to the subsequent content recognition module, ensuring that only repaired content that passes semantic verification enters the final classification processing stage.

[0107] Step 7: Content Recognition and Classification Result Generation According to an embodiment of this application, the restored complete image is subjected to content region recognition, and combined with memory reinforcement timing prediction, a classification result that integrates damage repair and temporal optimization is generated.

[0108] In some implementations, optical character recognition (OCR) technology may be optionally used to identify the text content in the restored image, which is a well-established technology in the field.

[0109] Post-processing and classification scheme generation of OCR results: The system uses a multi-level decoding algorithm to convert the OCR recognition results into structured classification information. The specific process is as follows: First, the original character sequence output by the OCR is subjected to spell checking and vocabulary verification, filtering out recognition errors and supplementing missing characters; then, the cleaned text content is input into the vocabulary difficulty assessment module, which calculates the difficulty level (1-10 levels) based on features such as vocabulary length, syllable complexity, and usage frequency; next, combined with the user's memory reinforcement timing prediction results, the optimal review time point is calculated using a forgetting curve model, generating a review priority ranking (high, medium, and low levels); finally, the difficulty level, review priority, and recommended learning time are integrated into a structured classification scheme, outputting a complete classification result in JSON format containing fields such as "vocabulary content," "difficulty level," "review priority," "recommended review time," and "memory strength assessment."

[0110] Furthermore, the specific calculation method for vocabulary difficulty levels is as follows: a multi-feature linear regression model is used for difficulty assessment, and the specific implementation steps are as follows: The first step is to extract the basic features of the vocabulary, including letter length (number of characters), number of syllables (calculated by syllable segmentation algorithm), word frequency level (ranking of usage frequency in standard corpus), part of speech type (noun, verb, adjective, etc.), and semantic abstraction level (quantitative score of concrete words vs. abstract concepts). The second step is to standardize each feature, normalizing the feature values ​​of different dimensions to the range of [0,1]. The third step is to calculate the difficulty score using a pre-trained linear regression model. The regression equation is as follows: The method involves several steps: First, the coefficients are obtained through training on large-scale labeled data. Second, the continuous difficulty scores are mapped to discrete levels from 1 to 10, with the mapping rule being: scores of 0-1 correspond to levels 1-2, scores of 1-2 correspond to levels 3-4, and so on. Third, personalized adjustments are made based on the user's historical learning data, fine-tuning the final level assessment according to the user's mastery of words of similar difficulty. This method ensures both objectivity and personalized adaptability in difficulty assessment.

[0111] As can be seen, based on the identified complete content and the user's memory reinforcement timing prediction results, the system generates a personalized English flashcard classification scheme, such as difficulty level, review priority, and recommended study time.

[0112] Experimental verification and performance analysis To verify the effectiveness of the technical solution proposed in this application, systematic experimental tests were conducted, including a large-scale verification experiment involving 500 users and 2000 damaged English translation cards. The following charts illustrate the key technical indicators and performance comparison results.

[0113] Figure 2 The presentation includes line graphs validating the personalized forgetting curve model, demonstrating the improved Ebbinghaus forgetting curve model's accuracy in predicting memory retention rates across different user types, and comparing it to the traditional forgetting curve model.

[0114] Personalized models, on average, improved prediction accuracy by 16.8 percentage points compared to traditional models, effectively validating the individual difference parameters. , and review effect coefficient Its scientific nature.

[0115] Figure 3 Grouped bar charts show the performance comparison of image restoration methods; comparing the restoration accuracy of traditional pixel restoration, deep learning-based restoration, and the memory-guided restoration method of this application under different degrees of damage.

[0116] The method described in this application is significantly superior to the comparative method under various degrees of damage, especially under severe damage, where the accuracy is improved by more than 25 percentage points.

[0117] Figure 4 A heatmap showing the vocabulary difficulty-user level repair success rate matrix was displayed; the distribution pattern of repair success rate under different combinations of vocabulary difficulty levels (levels 1-10) and user English proficiency (beginner, intermediate, advanced, expert) was also shown.

[0118] The user learning level matching factor was validated. The design is reasonable, and the success rate of repair at different user levels is in line with the laws of cognitive science.

[0119] Figure 5 The radar charts show the multi-dimensional performance radar comparison analysis of the system; the comprehensive performance of traditional methods and the method of this application is compared from five dimensions: repair accuracy, processing speed, user satisfaction, semantic consistency, and time stability.

[0120] The method proposed in this application is significantly superior to the traditional method in four core indicators, and is only slightly lower in processing speed by 7 points, but the overall performance is improved by more than 60%.

[0121] Figure 6 The stacked area plot shows the evolution trend of temporal consistency scores; it also shows the distribution of memory matching weights and the trend of cumulative consistency scores with verification time at different historical time points in the temporal memory verification algorithm.

[0122] The temporal memory verification algorithm effectively identified repair solutions that conformed to the user's cognitive patterns, thus verifying... The scientific nature of function design.

[0123] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for classifying direct-reading English flashcards based on image recognition processing, characterized in that, Includes the following steps: Acquire images of damaged English flashcards and users' historical learning records, identify damaged areas, and extract individual forgetting pattern features of users; The degree of memory decay and the reliability coefficient of memory are calculated based on the forgetting curve fitting algorithm; The memory-guided content reconstruction algorithm is used to adjust the image restoration parameters by using the memory credibility coefficient as a constraint weight; Based on the semantic probability distribution calculated from lexical features, and combined with memory constraints, candidate repair solutions are generated. The temporal memory verification algorithm is used to analyze the matching degree between candidate solutions and the user's memory state, and the optimal repair content is selected. The repair results are output after verifying the semantic rationality of the repaired content based on a language model. Content recognition is performed on the restored image to generate a classification result that integrates damage restoration and temporal optimization; Among them, the memory-guided content reconstruction algorithm modifies the image repair objective function by establishing a memory constraint matrix, the forgetting curve fitting algorithm adopts an improved Ebbinghaus forgetting curve model combined with individual difference parameters, and the temporal memory verification algorithm calculates the temporal consistency score by constructing a memory state time trajectory diagram.

2. The method according to claim 1, characterized in that, The extraction of individual user forgetting pattern features includes: extracting learning records from the user's historical learning database and constructing a forgetting pattern feature vector that includes learning time interval, review frequency, historical accuracy rate, and vocabulary difficulty level.

3. The method according to claim 1, characterized in that, The calculation process of the memory reliability coefficient includes: obtaining the memory retention rate value and the context weight adjustment factor value, multiplying the two values ​​to obtain the memory reliability coefficient, which comprehensively reflects the combined impact of time decay effect and individual differences on the accuracy of user memory.

4. The method according to claim 1, characterized in that, The establishment of the memory constraint matrix includes: multiplying the memory credibility coefficient with the semantic constraint vector constructed based on the user's memory content to form a numerical matrix used to constrain the restoration content to conform to the user's memory during the image restoration process.

5. The method according to claim 1, characterized in that, The calculation of the semantic probability distribution includes: using an improved TF-IDF algorithm to calculate the semantic relevance of candidate words in the current context. This algorithm introduces a semantic similarity factor and a user learning level matching factor to ensure that the recommended candidate words match the user's current learning level.

6. The method according to claim 1, characterized in that, The calculation of the temporal consistency score includes: constructing a continuous memory state curve using cubic spline interpolation, calculating the time weight using an exponential decay function, and obtaining the matching degree by comparing the similarity between the candidate scheme and the historical memory state.

7. The method according to claim 1, characterized in that, The language model verification uses a pre-trained BERT model. The semantic rationality is evaluated by calculating the perplexity of the repaired content in the language model. When the perplexity exceeds a preset threshold, the candidate solution is regenerated.

8. The method according to claim 1, characterized in that, The generation of the classification results includes: using optical character recognition technology to identify the text content in the restored image, and based on the identified complete content and the user's memory reinforcement timing prediction results, generating a personalized English flip card classification scheme that includes difficulty level, review priority, and recommended study time.

9. The method according to claim 1, characterized in that, The forgetting curve fitting algorithm uses the least squares method to fit the user's individualized forgetting parameters and determines the optimal value of the individualized parameters through nonlinear regression analysis. The individualized parameters reflect the user's basic forgetting rate and the shape characteristics of the forgetting curve.

10. A direct-input English translation card classification system based on image recognition processing, characterized in that, include: The image acquisition module is used to acquire digital images of damaged English translation cards; The database management module is used to store and manage user learning records, vocabulary knowledge base, image sample library, and model parameter library; The image processing module is used to identify the boundaries of damaged areas and extract forgetting pattern features; The memory modeling module is used to calculate the degree of memory decay and the memory reliability coefficient; The content reconstruction module is used to execute the memory-guided content reconstruction algorithm, which transforms memory information into image restoration constraints. The semantic analysis module is used to calculate the semantic probability distribution and generate candidate repair solutions; The timing verification module is used to execute the timing memory verification algorithm to select the optimal repair content; The results output module is used to generate the final repair results and classification scheme; The modules interact with each other through data interfaces to collaboratively complete the intelligent repair and classification tasks of damaged English translation cards.

Images