Training method and system for improving rapid memory and correcting memory pile error

Abstract

The invention belongs to the technical field of data processing, and mainly relates to a training method for improving rapid memory and memory pile error correction, which comprises the following steps of: inputting an image of a user frequently going to a place, dividing the image into at least two grid units by using an image division unit, and extracting matched feature vectors by using a convolutional neural network; dividing the content to be memorized into a plurality of text segments and matched feature vectors; the control module compares the aggregated feature vectors of the memory content and the image, and records the similarity; repeating the steps to obtain other images, and finally outputting the image feature vector with the highest similarity with the memory content feature vector. The control module performs matching and stake determination according to the output image feature vector and the text segment feature vector; and displaying the stake-fixed image to the user for memorizing, thereby solving the technical defect that the existing memory training method cannot match the image and the memory content. In addition, the invention further designs a training system for improving the rapid memory and correcting the error of the memory pile, and the defects in the prior art are further overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a training method and system for improving rapid memory and correcting memory errors. Background Art

[0002] At present, many memory methods, such as the Method of Loci, mainly rely on images in the user's daily life as memory credentials. Although these images have a certain visual impression, there are significant defects in the fit between the selected pictures and the memory content. Specifically, the existing methods of locating often do not fully consider the relationship between the semantic characteristics of the image and the content that the user needs to remember. If the fit between the image and the locating content is not high, it will cause confusion and discomfort to the user during the memory process, thereby affecting the effect and persistence of the memory. Therefore, the application effect of the existing locating method has not reached the best state and cannot give full play to its potential advantages.

[0003] In the prior art, images are usually anchored in a random manner or according to the order input by the user. Among them, the images randomly selected by this method often lack an inherent logical connection with the memory content, which makes it difficult for users to establish an effective connection when recalling. For example, when a user is memorizing a content about a "shopping list", if the anchored image is a "park scene", it may not be possible to naturally match the shopping items with the image, resulting in difficulties in memory conversion. At the same time, when images are used as anchors directly in the order of input, users may be limited by sequential memory. Long-term memory information processing, especially when switching between multiple images and content, can easily cause information interference when users are memorizing content, reducing memory efficiency. This memory method may not provide users with a clear and stable memory framework, making the stored information more vague and difficult to retrieve.

[0004] Based on this, we continue to improve the existing training methods and systems for improving rapid memory. Summary of the invention

[0005] One of the purposes of the present invention is to provide a training method for improving rapid memory and correcting memory errors, which can improve memory speed and correct errors in the pile setting process, in response to the defects of the prior art.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A training method for improving rapid memory and memory pile error correction, comprising the following steps:

[0008] S101, inputting an image of a place where the user often moves through an image input module, dividing the image into at least two grid units through an image division unit, and extracting a feature vector matching the grid unit from each grid unit through a convolutional neural network;

[0009] S201, importing the content that the user needs to remember into a text embedding model through a content input module, and dividing the memory content into multiple text segments and feature vectors matching the text segments according to the text embedding model;

[0010] S301, aggregating the feature vectors matching at least two text segments into a feature vector of memory content by weighted averaging; and at the same time aggregating the feature vectors matching at least two grid units into a feature vector of an image by weighted averaging;

[0011] S401, the control module performs a similarity comparison between the feature vector of the memory content and the feature vector of the image, and records the similarity between the feature vectors of the image and the memory content;

[0012] Repeat step S101 to obtain images of places where other users frequently move, and compare and record them with the feature vectors of the memory content respectively;

[0013] S501, the control module outputs the feature vector of the image with the highest similarity to the feature vector of the memory content;

[0014] At the same time, the control module matches and locates the feature vector of each grid unit of the output image with the feature vector of the text segment;

[0015] The fixed image is output to the display panel for the user to remember.

[0016] The technical solution achieves the following technical effects:

[0017] This application uses an image input module to obtain images of places where users often move around, and divides these images into multiple grid units. The feature vector of each grid unit is extracted through a convolutional neural network (CNN) to ensure that the existence characteristics of each sub-area are fully utilized in the memory process; the content to be remembered is imported into the text embedding model through the content input module, divided into multiple text segments, and its feature vector is extracted accordingly. Each content segment and the corresponding grid unit feature vector are weighted averaged to generate a more representative overall feature vector. In this way, it is possible to focus on features with higher importance, improve the matching degree between the image and the memory content, and thus enhance the memory effect.

[0018] At the same time, the control module of this application will perform a similarity comparison between the generated image feature vector and the memory content feature vector, and record the degree of similarity. This process can intelligently identify the fit between the image and the content, thereby dynamically adjusting the positioning strategy. When the system outputs the image with the highest similarity to the memory content feature vector, the user's positioning process can be more reasonable, making memory more efficient and accurate.

[0019] As a further improvement of the training method for improving rapid memory and memory pile error correction of the present invention, the working mode of the image division unit is:

[0020] Define the image of the unit where the input image is divided into units, where the height of the image is defined as H, the width is defined as W, and the channel is defined as C;

[0021] The image is divided into M*N rectangular units, and each rectangular unit is processed as follows to obtain a grid unit;

[0022]

[0023] Among them, G i，j The grid unit into which the image is divided at the i-th row and j-th column, where i and j are the row index and column index of the image respectively.

[0024] As a further improvement of the training method for improving rapid memory and memory pile error correction of the present invention, the image division unit also includes processing of dividing the image overlapping area, and the processing method is:

[0025] The image segmentation unit obtains the overlapping area by using a sliding window method, and the window size is k *W k , the step is S, and the processing method is as follows:

[0026] G k =I[x:x+H k，y :y+W k ]

[0027] Among them, G k For the Kth overlapping area, X=S*i, Y=S*j.

[0028] As a further improvement to the training method for improving rapid memory and memory pile error correction of the present invention, the text embedding model divides the memory content into multiple text segments, wherein the vector representation of each text segment is:

[0029]

[0030] Among them, V wk is the feature vector of the corresponding text segment, and D is the dimension of the feature vector.

[0031] As a further improvement to the training method for improving rapid memory and memory pile error correction of the present invention, the text embedding model divides the memory content into multiple text segments, wherein the vector representation of each text segment is:

[0032]

[0033] in, is the feature vector of the corresponding text segment, and D is the dimension of the feature vector.

[0034] As a further improvement to the training method for improving rapid memory and memory pile error correction of the present invention, a method for aggregating the feature vectors of at least two grid unit matches into a feature vector of an image by weighted averaging is as follows:

[0035] First, the feature vector of the grid unit is assigned a weight ω according to the edge feature of the grid unit in the image. i，j , the feature vector after adding weights is obtained by calculation, and g is the weight calculation function;

[0036] Secondly, the feature vectors of the grid cells are integrated using the weighted average method to obtain the feature vector V of the image overall .

[0037] As a further improvement to the training method for improving rapid memory and memory pile error correction of the present invention, the method of integrating grid units using the weighted average method is:

[0038]

[0039] Where N and M are the number of rows and columns of grid cells after image segmentation, respectively.

[0040] As a further improvement to the training method for improving rapid memory and memory pile error correction of the present invention, the control module uses cosine similarity or Euclidean distance to perform similarity comparison between the feature vector of the memory content and the feature vector of the image.

[0041] The method of using cosine similarity to compare the feature vector of the memory content with the feature vector of the image is:

[0042]

[0043] Among them, V Ia is the feature vector of the image, is the feature vector of the memory content.

[0044] The second purpose of the present invention is to provide a training system for improving rapid memory and correcting memory errors, which can improve memory speed and correct errors in the pile setting process in response to the defects of the prior art.

[0045] In order to achieve the technical purpose, this application implements the following technical solutions:

[0046] A training system for improving rapid memory and memory pile error correction, characterized in that it includes an image input module, a content input module, and a control module connected to the image input module and the content input module;

[0047] The image input module is used to input an image of a place where the user often moves and divide the image into at least two grid units through an image division unit arranged in the image input module;

[0048] The content input module is used to input the user's memory content and divide the memory content into multiple text segments and feature vectors matching the text segments through a text embedding model set in the content input module;

[0049] The control module is used to match and fix the image and memory content according to any of the above training methods for improving rapid memory and memory pile error correction, and output the fixed image to the display panel. The control module is communicatively connected with the display panel.

[0050] The technical solution produces the following technical effects:

[0051] The control module obtains the image and text content through the image input module and the content input module and extracts the feature vectors of the two; the similarity between the image feature vector and the text feature vector is obtained by similarity test of the feature vectors, and matching and anchoring are performed according to the similarity results. The display control unit is responsible for receiving the anchored image output by the control module, and controlling the display panel to display the image in an appropriate manner so that the user can perform memory training. Through the technical solution, the training system of the present invention can effectively improve the user's rapid memory and provide an error correction function during the memory process, thereby helping the user to memorize information more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a process flow chart of the technical solution of Embodiment 1 of the present invention;

[0053] Figure 2 This is a schematic diagram of the structure of Embodiment 4 of the present invention;

[0054] In the figure:

[0055] 1- Image input module;

[0056] 11 - Image division unit;

[0057] 2- Content input module;

[0058] 21-Text Embedding Model;

[0059] 3- Control module;

[0060] 4- Display panel. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application. The terms used herein in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0062] In the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected" and "fixed" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the terms in the present invention can be understood according to specific circumstances.

[0063] Although the present application discloses the preferred embodiments as follows, it is not intended to limit the claims. Any person skilled in the art can make several possible changes and modifications without departing from the concept of the present application. Therefore, the protection scope of the present application shall be based on the scope defined by the claims of the present application. The present invention is further described in detail below in conjunction with specific implementation methods, but the embodiments of the present invention are not limited thereto.

[0064] The present invention is further described in detail below in conjunction with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0065] Example 1

[0066] like Figure 1-2 As shown, it is known that the existing memory training methods and systems do not achieve the link of matching the user's daily images with the content that the user needs to remember. However, some of the content that the user needs to remember can be further improved by locating the content under the matching pictures. Therefore, the present application makes improvements to the existing memory training methods. Specifically, the present application provides a training method for improving rapid memory and memory pile error correction, which includes the following steps:

[0067] S101, inputting an image of a place where a user often moves through the image input module 1, dividing the image into at least two grid units through the image division unit 11, and extracting a feature vector matching the grid unit from each grid unit through a convolutional neural network;

[0068] S201, importing the content that the user needs to remember into the text embedding model 21 through the content input module 2, and dividing the memory content into multiple text segments and feature vectors matching the text segments according to the text embedding model 21;

[0069] S301, aggregating the feature vectors matching at least two text segments into a feature vector of memory content by weighted averaging; and at the same time aggregating the feature vectors matching at least two grid units into a feature vector of an image by weighted averaging;

[0070] S401, the control module 3 performs a similarity comparison between the feature vector of the memory content and the feature vector of the image, and records the similarity between the feature vectors of the image and the memory content;

[0071] Repeat step S101 to obtain images of places where other users frequently move, and compare and record them with the feature vectors of the memory content respectively;

[0072] S501, the control module 3 outputs the feature vector of the image with the highest similarity to the feature vector of the memory content;

[0073] At the same time, the control module 3 matches and locates the feature vector of each grid unit of the output image with the feature vector of the text segment;

[0074] The anchored image is output to the display panel 4 for the user to memorize.

[0075] The working principle of the technical solution can be summarized as image input and division, feature vector extraction, input and processing of memory content, feature vector aggregation and similarity comparison and selection mechanism and parts. Among them, the image input and division mainly splices the image input module 1 and the image division unit 11. The image input module 1 collects images of places where users often move around (such as home environment, workplace, etc.). These images provide a real background and perspective for subsequent memory. The image division unit 11 divides the input image into at least two grid units. This process breaks down the complex image information and makes it more refined and manageable. Each grid unit will extract features separately.

[0076] The feature vector extraction part mainly processes each grid unit through a convolutional neural network (CNN) to extract its feature vector. The feature vector contains the key visual information of the grid unit and lays the foundation for subsequent feature matching. The input and processing of memory content mainly rely on the text embedding model 21, through which the memory content input by the content input module 2 is divided into multiple text segments, and the feature vector of each text segment is extracted accordingly. This process ensures the structuring of the memory content and also provides preparation for similarity analysis.

[0077] The feature vector aggregation part relies on the weighted average method to perform weighted average aggregation on the feature vectors of the memory content and the image respectively. After this step, a more representative overall feature vector is formed, which can effectively reflect the user's memory needs and the required visual cues. Finally, through similarity comparison and selection: the feature vector of the memory content is similarly compared with the feature vector of the image, and the degree of similarity is recorded. In this way, the system can identify which image is most suitable for associating with the given memory content. And the image feature vector with the highest similarity to the memory content feature vector is output to ensure that users can use the most helpful images when memorizing.

[0078] In addition, the above technical solution of the present application also includes a stake correction process: the feature vector of the image grid unit outputted by the special image is matched with the feature vector of the text segment, and a deeper memory connection is established by the stake method. The content of the user's memory and the corresponding image form a stable and interconnected memory structure in the mind. Finally, the image processed and optimized by the stake is output to the display panel 4 for the user to memorize. The user can improve the efficiency and effect of memory by learning in this context.

[0079] The following are several possible specific implementation scenarios of the technical solution of this application, which can show how the technical solution of this application works in actual scenarios:

[0080] Scenario 1: Shopping list memory

[0081] The user wants to remember the items on the shopping list (such as milk, fruit, bread, etc.). Images of places where the user often moves (such as supermarkets, kitchens, etc.) are input into the system, and the system divides these images into multiple grid cells and extracts features. For example:

[0082] Image: An image of a supermarket shelf divided into four grids: beverage section, vegetable section, snack section, and bread section.

[0083] Memorizing content: The shopping list content is divided into multiple text segments and feature vectors are generated.

[0084] The control module 3 compares the image features with the memory content features, and finally outputs the grid (beverage area) most relevant to "milk" to help users quickly locate when shopping.

[0085] Scenario 2: Review of exam knowledge points

[0086] Students need to remember the key concepts of a course. Images of the user's usual school environment are analyzed and features are extracted.

[0087] Image: Different areas of a classroom (lectern, blackboard, desks).

[0088] Memorize content: including course definitions, formulas, and important concepts, etc., which are split through the text embedding model21.

[0089] After comparison, the system outputs the image area (such as a blackboard) that is closest to the feature vector of a specific concept (such as "speed of light"), helping students enhance their memory when reviewing.

[0090] Scene 3: Travel route memory

[0091] Users want to remember the various attractions they visit during their travels. The images of places taken by users during their travels are used as input, and the system analyzes the features of these images.

[0092] Image: Famous landmarks (e.g., Eiffel Tower, Statue of Liberty) divided into grid cells.

[0093] Memory content: Text processing of the user's travel attractions and precautions.

[0094] After matching the features, the system helps users record image features related to the “Eiffel Tower”, thereby improving the depth and accuracy of their memory of the attraction.

[0095] It can be seen that the technical solution of this application optimizes the storage and retrieval process of information by combining the user's daily activity images and memory content. By enhancing the matching and staking strategies, it can significantly improve the user's memory effect and efficiency, and is suitable for various memory training and education fields.

[0096] Example 2

[0097] The difference from the first embodiment is that, in order to further improve the data processing effect of the present application, the image division unit 11 works as follows:

[0098] The image of the input image division unit 11 is defined, wherein the height of the image is defined as H, the width is defined as W, and the channel is defined as C;

[0099] The image is divided into M*N rectangular units, and each rectangular unit is processed as follows to obtain a grid unit;

[0100]

[0101] Among them, G i，j The grid cells of the image at the i-th row and j-th column, i and j are the row index and column index of the image respectively. In this way, the feature extraction process of the image is further refined, ensuring that the feature vector of each grid cell can more accurately reflect the local information of the image. In this way, the system can capture the details in the image more meticulously, thereby providing richer visual cues for the memory content.

[0102] Furthermore, the image division unit 11 also includes processing for dividing the image overlapping area, and the processing method is as follows:

[0103] The image division unit 11 obtains the overlapping area by using a sliding window method, and the window size is k *W k , the step is S, and the processing method is as follows:

[0104] G k =I[x:x+H k , y:y+W k ]

[0105] Among them, G k For the Kth overlapping area, X=S*i, Y=S*j.

[0106] The reason for dividing the image into overlapping areas is that in some images, targets or features may appear at any position or near the boundary of the image. By dividing the overlapping areas, it can be ensured that these key information are not missed. For example, when dealing with medium-sized objects, some parts may be blocked by adjacent areas. The overlapping area can capture all possible details. If only non-overlapping areas are used for feature extraction, features located at the edge of the grid may not be fully captured. These edge effects can be effectively avoided by overlapping areas, ensuring that the features extracted from each grid are complete and meaningful.

[0107] Furthermore, the text embedding model 21 divides the memory content into multiple text segments, where the vector representation of each text segment is:

[0108]

[0109] in, is the feature vector of the corresponding text segment, and D is the dimension of the feature vector. It is worth noting that any data (whether image, text or other forms) can be expressed as a feature vector after processing and conversion. The feature vector is an abstract representation of the original data, expressing the key information of the data in the form of numerical values.

[0110] In natural language processing (NLP), feature vectors are usually embedded representations of words, sentences, or documents, aiming to capture their semantic information. The dimension D represents the number of values ​​contained in the feature vector. These values ​​can be regarded as the components of the feature or the control of the feature. In the context of word embedding, common dimensions are 50, 100, 200, 300, etc.

[0111] Furthermore, the method of aggregating the feature vectors matching at least two text segments into the feature vector of the memory content by weighted averaging is as follows:

[0112]

[0113] Where N is the number of text segments, V C is the feature vector of the memory content, and k is the index of the text segment. Other details are the same as those in Embodiment 1, and will not be described in detail in this embodiment.

[0114] Example 3

[0115] Different from Example 1, in order to further improve the matching completeness of the feature vector in the present application, further, a method of aggregating the feature vectors matched by at least two grid units into the feature vector of the image by weighted averaging is as follows:

[0116] First, the weight w is assigned to the feature vector of the grid unit according to the edge feature of the grid unit in the image. i，j , the feature vector after adding weights is obtained by calculation, and g is the weight calculation function;

[0117] Secondly, the feature vectors of the grid cells are integrated using the weighted average method to obtain the feature vector V of the image overall .

[0118] Furthermore, the method of integrating grid cells using the weighted average method is:

[0119]

[0120] Where N and M are the number of rows and columns of grid cells after image segmentation, respectively.

[0121] Assuming the image is divided into 2*2 grid cells, the feature vectors and weights are as follows:

[0122]

[0123] The corresponding weights are set according to the edge features of the grid cells:

[0124] w 1,1 =0.5

[0125] w 1,2 =0.2

[0126] w 2,1 =0.1

[0127] w 2,2 =0.2

[0128] The calculation result of the overall image feature vector is: V overall =[0.24, 0.45, 0.32].

[0129] Based on this, through the improved scheme, the feature vector of the entire image is obtained, and the features of multiple sub-regions are effectively integrated with the weighted average method. In practical applications, the weight calculation method can be flexibly adjusted according to specific tasks. This technical solution can be effectively applied in tasks such as image classification and target detection.

[0130] Furthermore, the control module 3 uses cosine similarity or Euclidean distance to compare the feature vector of the memory content with the feature vector of the image. The method of using cosine similarity to compare the feature vector of the memory content with the feature vector of the image is:

[0131]

[0132] in, is the feature vector of the image, is the feature vector of the memory content.

[0133] Other details that are the same as those in Example 1 are not described in detail in this application.

[0134] Example 4

[0135] like Figure 1-2 As shown, the difference from Example 1 is that: in order to further improve the technical defects of the existing memory training system, the present application designs a training system for improving rapid memory and memory pile error correction, including an image input module 1, a content input module 2 and a control module 3 connected to the image input module 1 and the content input module 2;

[0136] The image input module 1 is used to input an image of a place where the user often moves and divide the image into at least two grid units through an image division unit 11 provided in the image input module 1;

[0137] The content input module 2 is used to input the user's memory content and divide the memory content into multiple text segments and feature vectors matching the text segments through the text embedding model 21 set in the content input module 2;

[0138] The control module 3 is used to match and fix the image and memory content according to any of the above training methods for improving rapid memory and memory pile error correction, and output the fixed image to the display panel 4. The control module 3 is communicatively connected to the display panel 4.

[0139] Through the technical solution, the system of the present application realizes the automated processing of rapid memory training and memory pile error correction. The system collects images of the user's daily activity places through the image input module 1, and receives the content that the user needs to remember through the content input module 2. The control module 3 is responsible for processing the image and memory content, and selects the image feature vector that best matches the memory content through feature vector extraction, aggregation and similarity comparison, and outputs these images to the display panel for the user to remember.

[0140] In addition, the system also includes a user interaction interface that allows the user to input memory content and provides a feedback mechanism to adjust and optimize the memory training process. The user can input memory content, and the system generates a feature vector based on the content and matches it with the image feature vector through the control module 3. After the match is successful, the system will output the image most relevant to the memory content to help the user perform memory training.

[0141] The system also has learning and adaptation functions, which can automatically adjust the training strategy and difficulty according to the user's memory effect and feedback to adapt to the learning progress and memory ability of different users. In this way, the system can provide users with personalized memory training programs, thereby improving the efficiency and effect of memory training.

[0142] In summary, this system provides an efficient, personalized and easy-to-operate memory training method and system by combining feature vector matching of images and memory content, as well as user interaction and feedback mechanism. It can not only help users improve their memory ability, but also adjust the training plan according to the specific situation of the user, so as to achieve the best memory training effect.

[0143] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0144] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A training method for improving rapid memory and memory pile error correction, characterized in that: The following steps are included: S101, inputting an image of a place where a user often moves through an image input module, dividing the image into at least two grid units through an image division unit, and extracting a feature vector matching the grid unit from each grid unit through a convolutional neural network; S201, importing the content that the user needs to remember into a text embedding model through a content input module, and dividing the memory content into multiple text segments and feature vectors matching the text segments according to the text embedding model; S301, aggregating at least two feature vectors matched by the text segments into a feature vector of the memory content by weighted averaging; and at the same time aggregating at least two feature vectors matched by the grid units into a feature vector of the image by weighted averaging; S401, the control module performs a similarity comparison between the feature vector of the memory content and the feature vector of the image, and records the degree of similarity between the feature vector of the image and the feature vector of the memory content; Repeat step S101 to obtain the images of places where other users frequently move, and compare and record them with the feature vectors of the memory content respectively; S501, the control module outputs the feature vector of the image that has the highest similarity to the feature vector of the memory content; At the same time, the control module matches and locates the feature vector of each grid unit of the output image with the feature vector of the text segment; The image after anchoring is output to a display panel for the user to memorize.

2. A training method for improving rapid memory and memory pile error correction according to claim 1, characterized in that: The working mode of the image division unit is: The image input to the image division unit is defined, wherein the height of the image is defined as H, the width is defined as W, and the channel is defined as C; Divide the image into M*N rectangular units, and perform the following processing on each of the rectangular units to obtain the grid units; Among them, G i，j The grid unit into which the image is divided at the i-th row and j-th column, where i and j are the row index and column index of the image respectively.

3. A training method for improving rapid memory and memory pile error correction according to claim 2, characterized in that: The image division unit also includes processing for dividing the image overlapping area, and the processing method is: The image division unit obtains the overlapping area by using a sliding window method, and the window size is H. k *W k , the step is S, and the processing method is as follows: G k =I[x:x+H k ,y:y+W k ] Among them, G k For the Kth overlapping area, X=S*i, Y=S*j.

4. A training method for improving rapid memory and memory pile error correction according to claim 1, characterized in that: The text embedding model divides the memory content into multiple text segments, where the vector representation of each text segment is: in, is the feature vector of the corresponding text segment, and D is the dimension of the feature vector.

5. A training method for improving rapid memory and memory pile error correction according to claim 4, characterized in that: The method of aggregating the feature vectors of at least two text segment matches into the feature vector of the memory content by weighted average is: Where N is the number of text segments, V C is the feature vector of the memory content, and k is the index of the text segment.

6. A training method for improving rapid memory and memory pile error correction according to claim 1, characterized in that: The method of aggregating the feature vectors of at least two grid unit matches into the feature vector of the image by weighted average is: First, a weight w is assigned to the feature vector of the grid unit according to the edge feature of the grid unit located in the image. i，j , the feature vector after adding weights is obtained by calculation, and g is the weight calculation function; Secondly, the feature vectors of the grid cells are integrated using the weighted average method to obtain the feature vector V of the image overall .

7. A training method for improving rapid memory and memory pile error correction according to claim 6, characterized in that: The method of integrating the grid cells using the weighted average method is: Wherein, N and M are respectively the number of rows and the number of columns of the grid units after the image is segmented.

8. A training method for improving rapid memory and memory pile error correction according to claim 1, characterized in that: The control module uses cosine similarity or Euclidean distance to perform similarity comparison between the feature vector of the memory content and the feature vector of the image.

9. A training method for improving rapid memory and memory pile error correction according to claim 8, characterized in that: The method of using cosine similarity to compare the feature vector of the memory content with the feature vector of the image is as follows: in, is the feature vector of the image, is the feature vector of the memory content.

10. A training system for improving rapid memory and memory pile error correction, characterized in that: It includes an image input module, a content input module and a control module connected with the image input module and the content input module; The image input module is used to input an image of a place where the user often moves and divide the image into at least two grid units through an image division unit arranged in the image input module; The content input module is used to input the user's memory content and divide the memory content into multiple text segments and feature vectors matching the text segments through a text embedding model set in the content input module; The control module is used to match and fix the image and memory content according to the training method for improving rapid memory and memory error correction as described in any one of claims 1-9, and output the fixed image to the display panel, and the control module is communicatively connected to the display panel.

Images