Intelligent stationery automatic classification and storage system and method

Abstract

The invention discloses an intelligent stationery automatic classification and storage system and method, and belongs to the technical field of product classification, and the system comprises an image recognition module, a scoring classification module and a storage module. The image recognition module is used for acquiring an image of stationery in real time and performing primary processing on the image; constructing a stationery feature model, and generating a stationery detection sequence of the to-be-detected stationery; the scoring classification model is used for constructing a stationery comparison sequence, comparing the to-be-detected stationery with each kind of stationery in the stationery comparison sequence, classifying the to-be-detected stationery and endowing the to-be-detected stationery with labels; and the storage module is used for moving the stationery to be detected to a specified position, identifying different types of stationery which are not consistent with existing stationery types in a storage process, calculating the occurrence frequency of the different types of stationery in a certain type of stationery, judging whether the different types of stationery are corrected stationery or not, and performing real-time adjustment on the stationery comparison sequence.

Description

Technical Field

[0001] The invention relates to an intelligent stationery automatic classification and storage system and method, belonging to the technical field of product classification. Background Art

[0002] With the rapid development of Internet technology, there are a large number of product categories, manufacturers, suppliers, and brand names and product names of products under the same category. The number of products can reach millions or tens of millions, etc., and it takes a lot of time to classify product information. Therefore, the workload of classifying product information is very large.

[0003] Traditional product management methods usually rely on manual recording and management, which is inefficient and prone to errors. The classification and storage of products lack intelligent means, which makes it difficult to find and count, and increases the workload of managers. In addition, as market competition intensifies, companies need to respond to market changes more quickly and adjust product lines in a timely manner to meet consumer needs. However, traditional manual operations are difficult to respond quickly, often resulting in missed business opportunities. In order to improve competitiveness, many companies have begun to seek more efficient and intelligent product management solutions.

[0004] The invention patent with the existing announcement number CN103699523B discloses a product classification method, which includes: extracting product text features based on product text used to describe the product to be classified; extracting product image features based on the product image of the product to be classified; generating product features of the product to be classified based on the product text features and the product image features; inputting the product features of the product to be classified into a pre-trained product classification model to obtain a classification result; extracting product text features and product image features of the product to be classified, and then generating product features based on the product text features and the product image features, so as to classify using the product features to obtain a classification result.

[0005] Although the prior art comprehensively considers the text features and image features of the products to be classified, and improves the classification accuracy compared to classification based on product information alone, it does not take into account changes in product types during the classification process. For example, over time, the appearance design or manufacturing standards of the product change, resulting in a mismatch with the feature data in the original product comparison library. Therefore, the present application provides an intelligent stationery automatic classification and storage system and method, which constructs a stationery feature model, identifies features in stationery images, and compares the stationery to be detected with existing stationery images for classification, classifies and stores the stationery to be detected. During the storage process, it identifies heterogeneous stationery that does not conform to the existing stationery type, and determines whether the heterogeneous stationery is a corrected stationery, and performs real-time corrections on the existing stationery type based on the corrected stationery. Summary of the invention

[0006] In view of the deficiencies of the prior art, the purpose of the present invention is to provide an intelligent stationery automatic classification and storage system and method. By constructing a stationery feature model, the stationery to be detected is identified and classified. During the classification process, a stationery comparison library is constructed for the existing stationery types, and during the storage process, the frequency of abnormal stationery in a certain type of stationery is identified. When the frequency exceeds the set threshold, the stationery comparison library is corrected in real time.

[0007] To achieve the above purpose, the present invention provides the following technical solutions: An intelligent stationery automatic classification and storage system, comprising: an image recognition module, a scoring and classification module, and a storage module; The image recognition module is used to obtain the images of the stationery in the batch to be detected in real time, perform preliminary processing on the images; and construct a stationery feature model to generate a stationery detection sequence of the stationery to be detected; The scoring and classification model is used to construct a stationery comparison sequence, and the stationery comparison sequence records the characteristic data of each type of stationery; compare the stationery to be detected with each type of stationery in the stationery comparison sequence, classify the stationery to be detected and assign a label; The storage module is used to move the stationery to be detected to a specified position, identify the heterogeneous stationery that does not match the existing stationery types during the storage process, calculate the frequency of the heterogeneous stationery appearing in a certain type of stationery, determine whether the heterogeneous stationery is a corrected stationery, and adjust the stationery comparison sequence in real time.

[0008] Specifically, the image recognition module includes an image acquisition unit and a feature extraction unit; The image acquisition unit captures the high-definition images of the stationery in the batch to be recognized in real time through a high-definition camera and performs preprocessing on the images; The feature extraction unit is configured with a feature fusion strategy. The feature fusion strategy analyzes the collected images by constructing a stationery feature model, extracts the key appearance features of the stationery to be detected, and generates a stationery detection sequence.

[0009] Specifically, the feature fusion strategy includes: Using the Transformer model, construct a stationery feature model including an embedding layer, a merging layer, and a fusion layer; Collect image samples and construct an image sample set ; Divide the image into image blocks of to generate an image block set of the image ; where ; is the image sample set the th image, ; Convert the image patch into a Token vector , and generate the Token sequence of the image ; where is the th image patch in the set of image patches , ; Adopt the self-attention mechanism, take the Token sequence as the input, and generate the feature sequence of the image ; Set the number of levels of the merging layer to , and generate a series of feature sequences , , , ; Concatenate the feature sequences of different levels to generate a set of feature sequences ; Use the self-attention mechanism, take the set of feature sequences as the input, and output the fused feature sequence ; Construct a validation set, and use the validation set to validate the stationery feature model; if the validation fails, expand the training set and continue training the model; if the validation passes, the stationery feature model is successfully constructed; Real-time obtain the image of the stationery to be detected in the image acquisition unit, and use the stationery feature model to generate the stationery detection sequence .

[0010] Specifically, the scoring and classification module includes a dynamic scoring unit and a label management unit; A comparison strategy is configured in the dynamic scoring unit. The comparison strategy constructs a stationery comparison sequence, calculates the similarity between the stationery to be detected and each type of stationery in the stationery comparison sequence, and classifies and assigns labels to the stationery to be detected according to the calculation results; The label management unit is used to construct a label database and update the label database in real time during the classification storage process.

[0011] Specifically, the comparison strategy includes: Generate a stationery comparison sequence based on the existing stationery images for classification; Calculate the stationery detection sequence Similarity with stationery and construct a similarity set of the stationery to be detected ; wherein, ; is the th kind of stationery in the stationery comparison sequence ; ; Obtain the maximum similarity in the similarity set , and assign a corresponding label to the stationery to be detected; Transmit the maximum similarity and label of the stationery to be detected to the storage module.

[0012] Specifically, the storage module includes a storage management unit and an anomaly detection unit; The storage management unit is used to control the storage device to move the stationery to a specified storage location, set a bundling quantity threshold, and bundle and store the stationery; The anomaly detection unit is configured with an anomaly recognition strategy and a feedback correction strategy; the anomaly recognition strategy is used to identify abnormal stationery that does not match the existing stationery types, calculate the frequency of abnormal stationery appearing in the same type of stationery, and determine whether the abnormal stationery is corrected stationery; the feedback correction strategy is used to adjust the file comparison sequence in real time according to the corrected stationery.

[0013] Specifically, the anomaly recognition strategy includes: Set the similarity threshold to , and obtain the maximum similarity of the stationery to be detected; Judge whether the stationery to be detected is abnormal stationery; if , the stationery to be detected is not abnormal stationery; if , the stationery to be detected is abnormal stationery; Calculate the abnormal frequency of the th kind of stationery, and construct an abnormal frequency set ; wherein, ; Set the abnormal frequency threshold to , and judge whether the abnormal stationery in the th kind of stationery is corrected stationery; if , the abnormal stationery in the th kind of stationery is not corrected stationery; if , the abnormal stationery in the th kind of stationery is corrected stationery.

[0014] Specifically, the feedback correction strategy includes: Obtain the image of the corrected stationery among the th types of stationery, and construct a set of heterogeneous images ; Define the stationery in the image as the th type of stationery, and update the stationery comparison sequence to ; Use the stationery feature model to sequentially generate the anomaly detection sequences of the images , , , and re-perform automatic classification and labeling of stationery based on the updated stationery comparison sequence, and perform anomaly recognition. , ,

[0015] An intelligent stationery automatic classification and storage method, including: Step S1: Real-time capture high-definition images of stationery within the batch to be recognized, and preprocess the images; Step S2: Use the Transformer model to construct a stationery feature model, and generate a stationery detection sequence of the stationery to be detected ; Step S3: Construct a stationery comparison sequence, compare the stationery to be detected with each type of stationery in the stationery comparison sequence one by one, classify the stationery to be detected and assign labels; Step S4: Control the storage device to perform homogeneous storage, and set a bundling quantity threshold to bundle the stationery; Step S5: Identify heterogeneous stationery that does not match the existing stationery categories, and adjust the stationery comparison sequence in real time.

[0016] Specifically, the step S5 includes: S5.1: Set the similarity threshold to , and determine whether the stationery to be detected is heterogeneous stationery; if , the stationery to be detected is not heterogeneous stationery; if , the stationery to be detected is heterogeneous stationery; S5.2: Calculate the heterogeneous frequency of the th type of stationery , and construct a set of heterogeneous frequencies ; S5.3: Set the heterogeneous frequency threshold to , and determine whether the heterogeneous stationery among the th type of stationery is corrected stationery; if , the heterogeneous stationery among the th type of stationery is not corrected stationery; if , the Among the stationery, the different type of stationery is the correction stationery; S5.4: Obtain the image of the correction stationery among the types of stationery, and construct a set of different-type images ; S5.5: Define the stationery in the image as the type of stationery, and update the stationery comparison sequence as ; S5.6: Use the stationery feature model to generate the anomaly detection sequences of the images , , in turn, and re-perform the automatic classification and labeling of the stationery based on the updated stationery comparison sequence. , , , respectively.

[0017] Advantages of the present invention: 1. Using the stationery feature model, an elaborate stationery comparison sequence is constructed, which records the unique feature data of each type of stationery, and the stationery to be detected is carefully compared with each type of stationery in the stationery comparison sequence in multiple dimensions such as appearance features, dimensions, and materials, so as to judge the category of the stationery to be detected and assign a label, improving the accuracy of classification, ensuring the consistency and traceability of the label, and avoiding classification errors caused by label confusion.

[0018] 2. Monitor the situation of the stationery in the storage area in real time. Once a different-type stationery that does not match the existing stationery categories is found, analyze the frequency of the different-type stationery appearing in a certain stationery category, judge whether it is a correction stationery, automatically update the stationery comparison sequence, add the newly identified stationery feature data to it, and re-identify and classify the different-type stationery, ensuring the accuracy and timeliness of the system, and at the same time enhancing the flexibility and adaptability of the system. Description of the Drawings

[0019] Figure 1 is a structural diagram of an intelligent stationery automatic classification and storage system; Figure 2 is a flowchart of feature extraction of an intelligent stationery automatic classification and storage system; Figure 3 is a flowchart of a comparison strategy of an intelligent stationery automatic classification and storage system; Figure 4 is a flowchart of anomaly recognition and correction of an intelligent stationery automatic classification and storage system; Figure 5 is a flowchart of an intelligent stationery automatic classification and storage method. Detailed Embodiment

[0020] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other. Embodiment

[0021] Reference Figures 1 to 4 As shown, this embodiment introduces an intelligent stationery automatic classification and storage system, including: an image recognition module, a scoring and classification module, and a storage module; The image recognition module is used to capture the images of stationery in the batch to be detected in real time, perform preliminary processing on the images, such as cropping, scaling, and denoising, to improve the image quality; and construct a stationery feature model, extract the appearance features from the images of the stationery to be detected, and integrate the appearance feature information into a stationery detection sequence; The scoring and classification module is used to obtain the original stationery images for classification, and construct a stationery comparison sequence using the stationery feature model. The stationery comparison sequence is used to record the feature data of each type of stationery, compare the stationery to be detected with each type of stationery in the stationery comparison sequence, judge the category of the stationery to be detected, and assign corresponding labels; The storage module is used to control the storage device to move the classified and labeled stationery to the specified storage area according to the corresponding label information, and identify the heterogeneous stationery that does not match the existing stationery categories during the storage process. By the frequency of heterogeneous stationery appearing in a certain type of stationery category, judge whether the heterogeneous stationery is a corrected stationery, so that the system can timely respond to the situation where the appearance design or manufacturing standard of stationery changes or a new type of stationery is introduced, improving the accuracy, timeliness, and flexibility of the system. Once the corrected stationery is identified, immediately update the stationery comparison sequence, and re-identify and classify the heterogeneous stationery.

[0022] Specifically, the image recognition module includes an image acquisition unit and a feature extraction unit; The image acquisition unit captures the high-definition images of stationery in the batch to be recognized in real time through a high-definition camera, and performs preprocessing on the images to ensure that the image quality meets the requirements of subsequent feature extraction; among them, the preprocessing includes: automatically identifying and cropping the image based on edge detection, so as to remove the unnecessary background or redundant parts in the image, making the stationery the main body of the image, reducing the interference of irrelevant information, using interpolation algorithms, such as bilinear interpolation or bicubic interpolation, to scale the cropped image according to the preset image size standard, while ensuring that the main features of the stationery are not distorted, and using filters to eliminate the noise points or miscellaneous points in the image, improving the clarity and purity of the image, which is beneficial to the subsequent feature extraction; A feature fusion strategy is configured in the feature extraction unit. By constructing a stationery feature model, the feature fusion strategy analyzes the collected images, extracts the key appearance features of the stationery to be detected, such as shape, color, and texture, and generates a stationery detection sequence to facilitate the identification and classification of the stationery batches to be recognized in the subsequent process.

[0023] Specifically, the specific steps of the feature fusion strategy include: Use the Transformer model to construct a stationery feature model, including an embedding layer, a merging layer, and a fusion layer; the embedding layer is used to divide the image into several image patches and map the image patches to Tokens. Each Token contains the information in the image patch and is processed by the Transformer. The merging layer is used to merge the Tokens to generate Token sequences of different numbers. The fusion layer is used to extract features from the Token sequences of different numbers in different stages; Collect a large number of image samples and construct an image sample set ; where is the number of samples collected, is the image sample set in the th image, ; Use the image sample set as the training set, and divide the image into image patches of to generate an image patch set of the image . Through the division of the image, each image patch can be processed independently, reducing the computational complexity of a single processing, thereby improving the ability to capture image detail information; where is the number of image patches, is the size of the image patch, is the set of image patches after division in the th image patch, ; Convert the image patch into a Token vector , thereby generating a Token sequence of the image ; where , is the vector sequence corresponding to the image patch , is the weight matrix of the linear transformation, is the bias matrix. Both the weight matrix and the bias matrix are to be solved. At this time, set and initial values, and train the parameters to be solved through the training set; Adopt the self-attention mechanism, and use the Token sequence as the input, merge the Token vectors in the Token sequence and generate the feature sequence of the image ; The expression of the self-attention mechanism is as follows: ; , ,

[0024]

[0025]

[0026] where, is the dimension of the feature sequence , , , , are the query vector, key vector, and value vector respectively, , , are the weight matrices to be solved respectively, is the dimension of the key vector; Perform hierarchical merging by designing multiple merging layers, and extract features in each layer to generate feature sequences of different dimensions, so that the model can capture the global and local information of the stationery image; Set the number of levels of the merging layer to , and in the subsequent levels, use the output of the previous level as the input of the current level, and use the self-attention mechanism to generate a series of feature sequences , , , ; where, , , , have dimensions of , , , , and ; Concatenate the feature sequences of different levels to generate a set of feature sequences ; represents the vector concatenation operation, and since the feature sequences of different levels have different dimensions, the dimensions need to be adjusted to ensure correct concatenation, such as padding with zero vectors; Use the self-attention mechanism to process the set of feature sequences As the input, output the fused feature sequence ; where is the fused feature sequence is the dimension of, and ; Construct a validation set and use the validation set to validate the stationery feature model; if the validation fails, expand the training set and continue training the model; if the validation passes, the stationery feature model is successfully constructed; Real-time obtain the image of the stationery to be detected in the image acquisition unit, and use the stationery feature model to generate the stationery detection sequence and transmit the stationery detection sequence to the scoring and classification module; where is the dimension of the stationery detection sequence .

[0027] Specifically, the scoring and classification module includes a dynamic scoring unit and a label management unit; The dynamic scoring unit is configured with a comparison strategy. The comparison strategy constructs a stationery comparison sequence by obtaining the original stationery images for classification, compares the features of the stationery to be detected with the features of each stationery in the stationery comparison sequence one by one, calculates the similarity score to determine whether it belongs to the same or the same series of stationery, and automatically classifies the stationery to be detected into the corresponding category according to the calculation result and assigns an accurate label; The label management module is used to construct a label database based on the feature vectors of the existing images. The label database is used to record the label information of each stationery; as new stationery is continuously added, the label database is updated in real time to ensure the integrity and timeliness of the database.

[0028] Specifically, the specific steps of the comparison strategy include: Obtain the existing stationery images for classification in the system and use the images as the input of the stationery feature model to generate the stationery comparison sequence ; where is the number of types of stationery, is the th stationery in the stationery comparison sequence , and , and , is the dimension of the stationery ; Calculate the similarity between the stationery detection sequence and the stationery , and construct the similarity set of the stationery to be detected; The expression is as follows:

[0029] In the formula, Detection sequence for stationery With stationery The matching score of each feature in Stationery included Features in , ; If the stationery detection sequence Stationery is not included Features in , ; Get similarity set The maximum similarity in , at this time, the type of stationery corresponding to the maximum similarity is the type of stationery to be detected, and the corresponding label is assigned. For example, the maximum similarity in the similarity set is , the corresponding type of stationery to be tested is Stationery, labeled ; The maximum similarity and label of the stationery to be detected are transmitted to the storage module for subsequent use or analysis.

[0030] Specifically, the storage module includes a storage management unit and an anomaly detection unit; The storage management unit receives the classified stationery and label information, and controls the storage devices according to the categories, such as automated shelves or robot arms, to move the stationery to the designated storage area, ensuring that the same type of stationery is stored together, which is convenient for subsequent management and retrieval, and improves work efficiency. During the storage process, the stationery bundling quantity threshold is set to When the accumulated quantity of a certain type of stationery reaches the bundling quantity threshold, the stationery is bundled to improve storage efficiency and space utilization, and the quantity information of the same type of stationery is updated, such as subtracting the bundling quantity threshold from the original quantity to ensure the accuracy and real-time nature of the data; The anomaly detection unit is configured with anomaly identification strategy and feedback correction strategy; The anomaly identification strategy is used to identify heterogeneous stationery that does not conform to the existing stationery category during the storage process, and calculate the frequency of heterogeneous stationery in the same type of stationery, and determine whether the heterogeneous stationery is a revised stationery by setting a threshold; The feedback correction strategy is used to obtain the correction stationery in the abnormal identification strategy and immediately adjust the stationery comparison sequence in real time, which improves the adaptability and robustness of stationery classification and storage and ensures the stable operation of the system.

[0031] Specifically, the specific steps of the anomaly identification strategy include: Set the similarity threshold to , and obtain the maximum similarity of the stationery to be detected ; Compare the maximum similarity with the similarity threshold , and determine whether the stationery to be detected is abnormal stationery; if , the stationery to be detected is not abnormal stationery; if , the stationery to be detected is abnormal stationery; During the storage process, obtain the number of abnormal stationery among the th stationery in the storage area and the total number of stationery , and calculate the abnormal frequency of the th stationery , and construct an abnormal frequency set ;

[0032] In the formula, ; Set the abnormal frequency threshold to , and determine whether the abnormal stationery among the th stationery is corrected stationery; if , the abnormal stationery among the th stationery is not corrected stationery; if , the abnormal stationery among the th stationery is corrected stationery.

[0033] Specifically, the specific steps of the feedback correction strategy include: Obtain the image of the corrected stationery among the th stationery, and construct an abnormal image set ; Define the stationery in the image as the th stationery, and use the stationery feature model to generate the abnormal detection sequence of the image , and update the stationery comparison sequence to ; is the dimension of the abnormal detection sequence ; Use the stationery feature model to sequentially generate the abnormal detection sequences , , of the images , , , and based on the updated stationery comparison sequence, execute the comparison strategy again, re - perform the automatic classification and labeling of stationery, and perform abnormal identification.

[0034] Example 2 Please refer to Figure 5, Another embodiment provided by the present invention: An intelligent stationery automatic classification and storage method, comprising the following steps: Step S1: A high-definition camera captures high-definition images of stationery within the batch to be recognized in real time, and preprocesses the images to ensure that the image quality meets the requirements of subsequent feature extraction; Step S2: Use a Transformer model to construct a stationery feature model, analyze the collected images, extract the key appearance features of the stationery to be detected, and generate a stationery detection sequence , facilitating the recognition and classification of stationery in the batch to be recognized subsequently; Step S3: Construct a stationery comparison sequence, compare the stationery to be detected with each type of stationery in the stationery comparison sequence one by one, determine whether it belongs to the same or the same series of stationery, automatically classify the stationery to be detected into the corresponding category, and assign an accurate label; Step S4: Control the storage device to perform storage of the same category according to the label, and set a bundling quantity threshold. When the quantity of a certain type of stationery accumulates to the bundling quantity threshold, bundle the stationery to achieve efficient management; Step S5: Identify different types of stationery that do not match the existing stationery categories during the storage process, and adjust the stationery comparison sequence in real time.

[0035] Specifically, the specific steps of constructing the stationery feature model in Step S2 include: S2.1: Collect a large number of image samples and construct an image sample set ; where is the number of samples collected, is the image sample set is the th image in ; S2.2: Use the image sample set as the training set, and divide the image into image patches of , generating an image patch set of image ; where is the number of image patches, is the size of the image patch, is the th image patch after division, ; S2.3: Convert the image patch into a Token vector , thereby generating a Token sequence of image ; where , is an image block corresponding vector sequence, is the weight matrix of the linear transformation, is the bias matrix; S2.4: Adopt the self-attention mechanism, take the Token sequence as the input, and generate the feature sequence of the image ; The expression of the self-attention mechanism is as follows: ; Among them, , ,

[0036]

[0037]

[0038] where, is the dimension of the feature sequence , , , , are the query vector, key vector, and value vector respectively, , , are the weight matrices to be solved respectively, is the dimension of the key vector; S2.5: Set the number of levels of the merging layer to , in the subsequent levels, use the output of the previous level as the input of the current level, and use the self-attention mechanism to generate a series of feature sequences , , , ; Among them, , , , of dimensions are , , , , and ; S2.6: Concatenate the feature sequences of different levels to generate the feature sequence set ; represents the vector concatenation operation, and since the feature sequences of different levels have different dimensions, the dimensions need to be adjusted to ensure correct concatenation, such as padding with zero vectors; S2.7: Use the self-attention mechanism, take the feature sequence set as the input, and output the fused feature sequence ; Among them, For the fused feature sequence of dimensions, and ; S2.8: Construct a validation set and use the validation set to validate the stationery feature model; if the validation fails, expand the training set and return to S2.2 to continue training the model; if the validation passes, the stationery feature model is successfully constructed.

[0039] Specifically, the specific steps of step S3 include: S3.1: Obtain existing stationery images for classification and use the images as the input of the stationery feature model to generate a stationery comparison sequence ; where is the number of types of stationery, is the th type of stationery, , and , is the dimension of the stationery ; S3.2: Calculate the similarity between the stationery detection sequence and the stationery , and construct a similarity set of the stationery to be detected; the expression is as follows:

[0040] In the formula, is the matching score of each feature in the stationery detection sequence and the stationery ; if the stationery detection sequence contains the feature in the stationery , ; if the stationery detection sequence does not contain the feature in the stationery , ; S3.3: Obtain the maximum similarity in the similarity set , and at this time, the type of stationery corresponding to the maximum similarity is the type of the stationery to be detected, and assign the corresponding label. For example, if the maximum similarity in the similarity set is , at this time, the corresponding type of the stationery to be detected is the th type of stationery, and the label is .

[0041] Specifically, the specific steps of step S5 include: S5.1: Set the similarity threshold to , determine whether the stationery to be detected is abnormal stationery; if , the stationery to be detected is not abnormal stationery; if , the stationery to be detected is abnormal stationery; S5.2: Obtain the number of abnormal stationery in the th type of stationery in the storage area in real time and the total number of stationery , and calculate the abnormal frequency of the th type of stationery , and construct an abnormal frequency set ;

[0042] In the formula, ; S5.3: Set the abnormal frequency threshold to , and determine whether the abnormal stationery in the th type of stationery is corrected stationery; if , the abnormal stationery in the th type of stationery is not corrected stationery; if , the abnormal stationery in the th type of stationery is corrected stationery; S5.4: Obtain the images of the corrected stationery in the th type of stationery, and construct an abnormal image set ; S5.5: Define the stationery in the image as the th type of stationery, and use the stationery feature model to generate the abnormal detection sequence of the image , and update the stationery comparison sequence to ; is the dimension of the abnormal detection sequence ; S5.6: Use the stationery feature model to sequentially generate the abnormal detection sequences , , of the images , , , and based on the updated stationery comparison sequence, execute the comparison strategy again to re - classify and label the stationery automatically.

[0043] In summary of the above embodiments, the present invention constructs a stationery feature model to obtain the features of the stationery to be detected, and constructs a stationery comparison library through the images of existing stationery types in the system. The features of the stationery to be detected are compared with each kind of stationery in the stationery comparison library in turn to obtain the category of the stationery to be detected and assign a clear label. The storage device moves the stationery to be detected to the designated position according to the label information, and identifies the non-conforming stationery that does not match the existing stationery types during the storage process. By the frequency of the non-conforming stationery appearing in a certain kind of stationery, it is judged whether it is a corrected stationery, and the stationery comparison library is adjusted and corrected in real time.

[0044] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. An automatic classification and storage method for intelligent stationery, characterized in that include: Image recognition module, scoring and classification module and storage module; The image recognition module is used to obtain images of stationery in the batch to be detected in real time, perform preliminary processing on the images, and construct a stationery feature model to generate a stationery detection sequence for the stationery to be detected; The scoring classification model is used to construct a stationery comparison sequence, which records the characteristic data of each stationery; the stationery to be detected is compared with each stationery in the stationery comparison sequence, and the stationery to be detected is classified and labeled; The storage module is used to move the stationery to be detected to a designated position, identify heterogeneous stationery that does not conform to the existing stationery type during the storage process, calculate the frequency of heterogeneous stationery appearing in a certain type of stationery, determine whether the heterogeneous stationery is a corrected stationery, and adjust the stationery comparison sequence in real time.

2. The intelligent stationery automatic classification and storage method according to claim 1, wherein: The image recognition module includes an image acquisition unit and a feature extraction unit; The image acquisition unit captures high-definition images of the stationery in the batch to be identified in real time through a high-definition camera, and pre-processes the images; The feature extraction unit is configured with a feature fusion strategy, which constructs a stationery feature model, analyzes the collected image, extracts key appearance features of the stationery to be detected, and generates a stationery detection sequence.

3. The automatic classification and storage method of intelligent stationery according to claim 2, characterized in that, The feature fusion strategy includes: Using the Transformer model, a stationery feature model including embedding layer, merging layer and fusion layer is constructed; Collect image samples and construct an image sample set ; Divide the image into image blocks of to generate a set of image blocks of the image ; where ; is the nth image in the image sample set ; Convert the image patch into a Token vector and generate the Token sequence of the image ; where is the th image patch in the image patch set ; Using the self-attention mechanism, take the said Token sequence as the input to generate the feature sequence of the said image ; ; Set the number of levels of the merged layer to , generating a series of feature sequences , , , ; Concatenate the feature sequences at different levels to generate a set of feature sequences ; Using the self-attention mechanism, the set of feature sequences is used as the input to output a fused feature sequence ; Constructing a validation set, and using the validation set to validate the stationery feature model; if the validation fails, expanding the training set, and continuing to train the model; if the validation passes, the stationery feature model is successfully constructed; Obtain the image of the stationery to be detected in the image acquisition unit in real time, and generate a stationery detection sequence using the stationery feature model .

4. The automatic classification and storage method of intelligent stationery according to claim 3, characterized in that: The scoring and classification module includes a dynamic scoring unit and a tag management unit; The dynamic scoring unit is configured with a comparison strategy, which constructs a stationery comparison sequence, calculates the similarity between the stationery to be detected and each stationery in the stationery comparison sequence, and classifies and labels the stationery to be detected according to the calculation result; The tag management unit is used to construct a tag database and update the tag database in real time during the classification storage process.

5. The automatic classification and storage method of intelligent stationery according to claim 4, characterized in that, The comparison strategy includes: Generate a stationery comparison sequence based on existing stationery images for classification ; Calculate the stationery detection sequence and the stationery similarity , and construct the similarity set of the stationery to be detected ; where is the nth stationery in the stationery comparison sequence ; Obtain the similarity set with the maximum similarity , and assign a corresponding label to the stationery to be detected; The maximum similarity and label of the stationery to be detected are transmitted to the storage module.

6. The automatic classification and storage method of intelligent stationery according to claim 5, characterized in that: The storage module includes a storage management unit and an abnormality detection unit; The storage management unit is used to control the storage device to move the stationery to a designated storage location, and to set a bundle quantity threshold to bundle and store the stationery; The anomaly detection unit is configured with an anomaly identification strategy and a feedback correction strategy; The anomaly identification strategy is used to identify heterogeneous stationery that does not conform to the existing stationery type, and calculate the frequency of occurrence of heterogeneous stationery in the same type of stationery to determine whether the heterogeneous stationery is a corrected stationery; the feedback correction strategy is used to adjust the file comparison sequence in real time according to the corrected stationery.

7. The automatic classification and storage method of intelligent stationery according to claim 6, characterized in that The anomaly identification strategy includes: Set the similarity threshold to and obtain the maximum similarity of the stationery to be detected ; Determine whether the stationery to be detected is abnormal stationery; if , the stationery to be detected is not abnormal stationery; if , the stationery to be detected is abnormal stationery; Calculate the outlier frequency of the th stationery item , and construct an outlier frequency set ; among them, ; Set the outlier frequency threshold to , and determine whether the outlier stationery among the th types of stationery is a corrected stationery; if , the outlier stationery among the th types of stationery is not a corrected stationery; if , the outlier stationery among the th types of stationery is a corrected stationery.

8. The automatic classification and storage method of intelligent stationery according to claim 7, characterized in that, The feedback correction strategy includes: Obtain the image of the corrected stationery among the th kinds of stationery, and construct a collection of heterogeneous images ; Define the image The stationery in is the th type of stationery, and update the stationery comparison sequence to Generate images in sequence using the stationery feature model , , abnormal detection sequences , , , and re-perform automatic classification and labeling of stationery based on the updated stationery comparison sequence, and perform anomaly recognition.

9. An intelligent stationery automatic classification and storage system, which is used to implement an intelligent stationery automatic classification and storage method as described in any one of claims 1-8, and is characterized in that, include: Step S1: capturing high-definition images of stationery in the batch to be identified in real time and preprocessing the images; Step S2: Construct a stationery feature model using a Transformer model and generate a stationery detection sequence for the stationery to be detected ; Step S3: Construct a stationery comparison sequence, compare the stationery to be detected with each type of stationery in the stationery comparison sequence one by one, classify the stationery to be detected and assign labels; Step S4: Control the storage device to perform storage of the same type, set a bundling quantity threshold, and bundle the stationery; Step S5: Identify abnormal stationery that does not match the existing stationery categories, and make real-time adjustments to the stationery comparison sequence.

10. The intelligent stationery automatic classification and storage system according to claim 9, wherein The said Step S5 includes: S5.1: Set the similarity threshold to , and determine whether the stationery to be detected is abnormal stationery; if , the stationery to be detected is not abnormal stationery; if , the stationery to be detected is abnormal stationery; S5.2: Calculate the outlier frequency of the th stationery , and construct an outlier frequency set ; S5.3: Set the heterogeneous frequency threshold to , and determine whether the heterogeneous stationery among the th types of stationery is the corrected stationery; if , the heterogeneous stationery among the th types of stationery is not the corrected stationery; if , the heterogeneous stationery among the th types of stationery is the corrected stationery; S5.4: Obtain the image of the corrected stationery among the th stationery items and construct a set of heterogeneous images ; S5.5: Define the image The stationery in is the th type of stationery, and update the stationery comparison sequence to S5.6: Generate images in sequence using the stationery feature model , , abnormal detection sequences , , , and re - perform automatic classification and labeling of stationery based on the updated stationery comparison sequence.

Images