TWI801982B - Classifying system and classifying method of automatically classifying digital file - Google Patents

Abstract

A classifying method of automatically classifying a digital file is disclosed and includes following steps: performing a classification procedure to an input file in accordance with a deep-learning classification model to determine a first category that the input file probably belongs to; reading a relevant data structure to confirm at least one slave category relevant to the first category, wherein the relevant data structure records relationships of multiple pre-defined categories and one or more slave categories; performing a feature comparing procedure to compare the input file with at least one digital file of the slave category in accordance with a machine-learning comparing model; and, labelling the input file as the first category and modifying a confidence of the input file belonging the first category when a comparing result of the feature comparing procedure matches a classification condition.

Description

Automatic digital file classification system and classification method

The invention relates to a classification system and a classification method, in particular to a classification system and a classification method for classifying digital files.

With the evolution of the times and the advancement of technology, more and more analog data have been digitized in recent years for the convenience of users to collect and search in the future.

However, as the amount of digital data explodes, problems related to data processing begin to arise. For example, in order to make digital data easy to search, the data must be classified first, but the traditional method of digitizing analog data and then manually classifying it will be difficult to achieve in the case of a huge amount of data. In addition to requiring a lot of labor costs and time, it is also prone to human error in classification. Therefore, automated classification systems began to receive attention.

In addition, with the explosion of data volume, new forms of data such as commodities, objects, news, and diseases continue to appear, resulting in more and more complex data types, which also makes the classification accuracy of automated classification systems on the market continue to decline. This shows that the current single-level classification method (that is, a classification procedure that is only performed once) is no longer sufficient.

In view of this, there is an urgent need in the market for a novel and highly accurate automatic classification system to solve the classification problem of digital data.

The main purpose of the present invention is to provide an automated digital archive classification system and classification method, which can compare digital archives with other relevant ones through machine learning comparison models when the reliability of the classification results of the deep learning classification model is insufficient. Compare the features of the categories to improve the accuracy of the classification results.

In order to achieve the above-mentioned purpose, the classification method of the present invention is mainly applied to a classification system, and includes: a) receiving an input file by the classification system; the file undergoes a classification procedure, wherein the classification procedure outputs a possible first category of the input file and a corresponding confidence level; c) the classification system reads an associative data structure to identify the at least one subcategory, wherein the association data structure records an association between a predefined plurality of categories and one or more subcategories; d) the classification system is based on a pre-trained machine learning comparison model for the input performing a feature comparison procedure on the file and at least one digital file in the subordinate category; and e) marking the input file as the first category when a comparison result of the feature comparison procedure meets a classification condition, and The confidence that the input profile belongs to the first category is increased.

In order to achieve the above-mentioned purpose, the classification system of the present invention mainly includes: an input unit, receiving an input file to be classified; A storage unit storing a deep learning classification model, a machine learning comparison model, and an associative data structure, wherein the associative data structure records an association between predefined plural categories and one or more subcategories; a The processing unit is connected to the input unit and the storage unit, and executes the following procedure: performing a classification procedure on the input file based on the pre-trained deep learning classification model, wherein the classification procedure outputs the possible first category of the input file and a corresponding reliability; reading the association data structure to confirm at least one subcategory associated with the first category; comparing the input file with the subcategory based on the pre-trained machine learning comparison model at least one of the digital files in a feature comparison program; and when a comparison result of the feature comparison program meets a classification condition, the input file is marked as the first category, and the input file belongs to the first category the credibility of the category; and an output unit connected to the processing unit to output the first category to which the input file belongs and the corresponding credibility.

Compared with the related technology, the present invention classifies the input files through at least two stages of procedures, which can effectively improve the accuracy of the classification results.

1: classification system

11: Processing unit

12: Input unit

13: Network unit

14: storage unit

141: Artificial Intelligence Algorithms

142:Deep Learning Classification Model

143:Machine Learning Comparison Model

144:Associative Data Structure

15: Output unit

2: Graphic data structure

21: node

22: side

3: Tree data structure

31: root node

32: child node

33: Attributes

34:Associative link

41: Home appliance category

411: washing machine

42: Outdoor Category

421: raincoat

43: Apparel category

431: shirt

5: Clothing features

S10~S32, S40~S58: classification steps

Fig. 1 is a first embodiment of the block diagram of the classification system of the present invention.

Fig. 2 is a first specific embodiment of the structure diagram of graphic data in the present invention.

FIG. 3A is a first specific embodiment of the tree structure diagram of the present invention.

FIG. 3B is a second specific embodiment of the tree structure diagram of the present invention.

Fig. 4 is a first specific embodiment of the flow chart of the classification method of the present invention.

Fig. 5 is a first specific embodiment of the feature comparison diagram of the present invention.

FIG. 6 is a first specific embodiment of a schematic diagram of feature correlation of the present invention.

FIG. 7 is a second specific embodiment of the feature correlation schematic diagram of the present invention.

Fig. 8 is a second specific embodiment of the flowchart of the classification method of the present invention.

A preferred embodiment of the present invention will be described in detail below in conjunction with the drawings.

Please refer to FIG. 1 , which is the first specific embodiment of the block diagram of the classification system of the present invention. The present invention discloses an automatic digital file classification system (hereinafter referred to as the classification system 1 for short in the specification). files (hereinafter referred to as input files), and automatically identify the content of the input files to complete the classification. In this way, the user can automatically obtain the category of the input file through the classification system 1, which is convenient for the user to file and store the input file. Human intervention is not required, so the classification speed is fast and no human error occurs, which is quite convenient.

The input files may be various digital multimedia files that need to be classified, such as digital documents, digital images, and digital audio, but not limited thereto. Specific parameters such as characters, symbols, and graphics are recorded in digital files, and the classification system 1 can automatically classify digital files based on these parameters. Specific image features are recorded in digital images, based on which the classification system 1 can automatically Classify digital images. The classification system 1 can automatically classify digital audio based on certain waveform characteristics recorded in digital audio.

However, the above are only some specific implementation examples of the types of input files in the present invention, but are not limited to the above types.

As shown in FIG. 1 , the classification system 1 of the present invention may include a processing unit 11 , and an input unit 12 connected to the processing unit 11 , a network unit 13 , a storage unit 14 and an output unit 15 . In one embodiment, the classification system 1 can be a personal computer (Personal Computer, PC), an industrial computer (Industrial PC, IPC), a cloud server (Cloud Server), a cabinet server (Cabinet Server), a tablet computer (Tablet), Notebook computer (Laptop) or any electronic device with the above-mentioned units 11-15.

The processing unit 11 can be, for example, a processor (Processor), a micro control unit (Micro Control Unit, MCU), a central processing unit (Central Processing Unit, CPU), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), and a field programmable Programmable logic gate array (Field Programmable Gate Array, FPGA), etc., to integrate and control the data of the input unit 12, the network unit 13, the storage unit 14 and the output unit 15, but not limited thereto.

The input unit 12 can be, for example, various man-machine interfaces, such as keyboards, mice, touch panels, touch screens, etc., for accepting user operations and importing input files that need to be classified into the classification system 1 middle.

The network unit 13 can be various wired or wireless network connection interfaces. The classification system 1 can be connected to the network through the network unit 13 , so that the user can transmit the input files to be classified to the classification system 1 from the remote end through the network.

The storage unit 14 can be, for example, a hard disk (Hard-Drive Disk, HDD), a solid-state disk (Solid-State Disk, SSD), a flash memory (FLASH Memory), a read-only memory (Read Only Memory, ROM), random access memory (Random Access Memory, RAM), non-volatile memory (Non-Volatile Memory), etc., but not limited thereto.

In the present invention, the storage unit 14 stores an artificial intelligence algorithm 141 , a deep learning classification model 142 , a machine learning comparison model 143 and an associated data structure 144 . The artificial intelligence algorithm 141 can be, for example, a convolutional neural network (Convolutional Neural Networks, CNN). The artificial intelligence algorithm 141 uses supervised learning (Supervised Learning) technology to perform in-depth learning according to the materials (not shown) with multiple category labels predetermined by the manager of the classification system 1, so as to generate the in-depth learning classification Model 142. As long as the classification system 1 imports the input file into the trained deep learning classification model 142, it can preliminarily determine which category the input file belongs to among the predetermined categories and the corresponding credibility.

Moreover, based on a plurality of predetermined categories, the manager of the classification system 1 can design corresponding object features for each category, such as color, shape, texture, etc., and the artificial intelligence algorithm 141 performs machine learning based on these object features , to generate the machine learning comparison model 143 . As long as the classification system 1 imports the input file into the machine learning comparison model 143 that has been trained, it can execute the feature comparison program by using parameters such as feature vectors and graphics of the input file to determine which category (or more) the input file matches. categories) are similar.

The main technical feature of the present invention is that the classification system 1 first uses the deep learning classification model 142 to perform a first-stage classification procedure on the input files, so as to preliminarily determine the category to which the input files belong. And after the classification process is completed, the machine learning comparison model 143 is used to perform a second-stage comparison process on the input file and the subordinate categories related to this category to confirm whether the classification result of the deep learning classification model 142 is credible. With at least two levels of processing procedures, the accuracy of the classification results given by the classification system 1 can be effectively improved.

As mentioned above, the deep learning classification model 142 is used to determine the category to which the input file belongs (hereinafter referred to as the first category), and the machine learning comparison model 143 is used to determine the input file and other categories associated with the first category ( Hereinafter referred to as subcategories) there is indeed similarity between. In order to achieve the above-mentioned technical means, the classification system 1 uses the association data structure 144 to record the association between each of the predefined plural categories and one or more subordinate categories (details will be described later).

The deep learning and machine learning are commonly used technical means in the field of artificial intelligence. In order to keep the content of the present invention concise, they will not be repeated here.

The processing unit 11 is used to receive the input file to be classified from the input unit 12 or the network unit 13, and classify the input file based on the models 142, 143 and the relational data structure 144 in the storage unit 14, and then pass the output unit 15 Outputting the classification result of the input file.

In one embodiment, after receiving the input file, the processing unit 11 performs a classification procedure on the input file based on the pre-trained deep learning classification model 142 . Since the deep learning classification model 142 is trained based on a plurality of predetermined category labels, after the classification program is executed, the processing unit 11 can judge the possible category of the input file from the multiple category labels (hereinafter referred to as is the first category), and the confidence that the input file belongs to the first category.

After preliminarily determining the first category to which the input file belongs, the processing unit 11 further reads the association data structure 144 to confirm whether the first category is associated with one or more subordinate categories. It is worth mentioning that the subordinate category does not refer to the subcategories included in the first category, but refers to other categories that have one or more common or similar object characteristics with the first category among the multiple category tags .

If after querying the correlation data structure 144, it is confirmed that the first category is related to at least one subcategory, the processing unit 11 further compares the input file with at least one digital file in the subcategory based on the pre-trained machine learning comparison model 143. perform a feature comparison procedure, and Get a comparison result. In one embodiment, the comparison result can show whether the input file and the subcategory really have similarity (ie, whether they have the same or similar object features), and the similarity degree (ie, the similarity of the object features).

After the feature comparison procedure ends, the processing unit 11 can determine whether the comparison result of the feature comparison sequence meets a preset classification condition, for example, whether the correlation procedure is higher than a preset threshold value. If the comparison result meets the classification conditions, the processing unit 11 can determine that the input file is highly likely to belong to the first category, so the input file is directly marked as the first category, and the output of the deep learning classification model 142 is increased. the credibility. Finally, the classification system 1 can output the first category and the revised reliability through the output unit 15, so that the user can use it as a reference when filing and managing the input file.

The present invention implements a two-level processing program through the deep learning classification model 142 and the machine learning comparison model 143, which can effectively improve the traditional single-level classification program, because the content of the digital file is complex (such as the proportion of the background part of the digital image). Large, background noise of digital audio, etc.), leading to the problem of lower classification accuracy.

Moreover, when the comparison result given by the machine learning comparison model 143 shows that the input file meets the classification conditions, the present invention will further correct the credibility of the classification result given by the deep learning classification model 142 . Therefore, the processing unit 11 can use the input file, the first category, and the revised credibility as parameters, and retrain the deep learning classification model 142 based on the artificial intelligence algorithm 141 , so as to gradually modify the deep learning classification model 142 . In this way, the deep learning classification model 142 will become more and more accurate as the classification system 1 is used for a longer period of time and the number of times of use increases. Thereby, it is helpful for the classification system 1 to establish knowledge for various new commodities and objects.

Please refer to FIG. 2 , which is a first specific embodiment of the graphic data structure diagram of the present invention. In one embodiment, the associative data structure 144 of the present invention can be a graph data structure 2 . As shown in Figure 2, the graphic data structure 2 is mainly composed of multiple nodes (node) 21 and multiple edges (edge) 22, and the two ends of each edge 22 are respectively connected to two nodes 21, thereby describing the relationship between the two nodes 21 sex.

Specifically, each node 21 may correspond to a category label predefined by the administrator, such as home appliance, outdoor, clothing, television, washing machine, and the like. Each edge 22 can respectively describe what is the relationship between the two nodes 21 (eg, primary relationship and secondary relationship), and the degree of relationship between the two nodes 21 (eg, high, low, or a specific relationship score).

For example, a first node representing a washing machine may be connected to a second node representing an electrical appliance through a first edge, and the first edge records that the first node and the second node are mainly associated with a high degree of association. And, since the digital files (such as photos) related to the washing machine often contain the characteristics of the clothes in addition to the characteristics of the washing machine, when establishing the graphic data structure 2, the manager can make the first node representing the washing machine pass through the second node at the same time. The edge connection represents the third node of the clothes, and the second edge records that the first node and the third node belong to the secondary association, and the degree of association is general.

In the present invention, the deep learning classification model 142 can preliminarily classify the input files through the edges 22 representing the primary associations, and compare the input files with relevant subordinate categories through the edges 22 representing the secondary associations. In this way, the accuracy of classification results can be improved based on the graphic data structure 2 .

Please refer to FIG. 3A again, which is a first specific embodiment of the tree structure diagram of the present invention. In another embodiment, the relational data structure 144 of the present invention can be a tree data structure 3 . As shown in Figure 3A, the tree-like data structure 3 has a plurality of root nodes (Root) 31 and a plurality of child nodes 32 (or subtrees (subtree)), wherein each child node 32 is connected to a root node 31 respectively, and each child node Festival Point 32 can be set with an attribute. The attributes are used to describe the relationship between the child node 32 and another child node 32 , but the present invention does not require all child nodes 32 to be set with corresponding attributes.

Specifically, each root node 31 can correspond to a generalized category label predetermined by the manager, such as home appliances, outdoor, clothing machine, etc., and each child node 32 can correspond to multiple narrow category labels under the broad category label. , such as TV washing machines, climbing equipment, raincoats, shirts, skirts, socks, etc. The attribute can be used to describe the relevance and degree of association between this child node 32 and other child nodes 32 under the same root node 31, or the association and relationship between this child node 32 and other child nodes 32 under different root nodes 31 degree of association.

Please also refer to FIG. 3B , which is a second specific embodiment of the tree structure diagram of the present invention. In the embodiment of FIG. 3B , the root node 31 represents category tags such as electrical appliances, kitchen utensils, outdoors, and clothing in a broad sense, and the sub-nodes 32 represent televisions, washing machines, dishwashers, storage cabinets, dishes, mountaineering equipment, raincoats, etc. in a narrow sense. Category tags such as shirts, skirts, socks, etc., and each child node 32 belongs to a root node 31 respectively.

As mentioned above, since there is a high probability of clothes appearing in the digital files related to washing machines (such as documents, photos, etc.) besides washing machines, but clothes do not belong to the category of electrical appliances, nor do they belong to the category of washing machines. Therefore, the administrator can set an attribute 33 in the child node 32 representing the washing machine in the tree data structure 3, and describe in the attribute 33 that the child node 32 of the washing machine is related to clothes. That is, consider the category of clothes as a subordinate category of the washing machine.

Similarly, although the child node 32 of the raincoat is connected to the outdoor root node 31, it is different from the washing machine, but the raincoat is related to clothes, and the probability of clothes appearing in the digital files related to raincoats is very high. Therefore, the manager can also use the attribute 33 to describe that the child node 32 of the raincoat is related to clothes. That is, consider the category of clothes as a subordinate category of raincoats. And, the child node 32 of the shirt and the child node 32 of the skirt Although connected to the root node 31 of clothing, it is different from washing machines and raincoats, but shirts and skirts are related to clothing, and the probability of clothing appearing in their digital files is very high. Therefore, the manager can also use the attribute 33 to describe that the child nodes 32 of shirts and skirts are related to clothes. That is, the category of clothes is regarded as a subordinate category of shirts and skirts.

Through the attributes 33 set in the tree data structure 3 , when the processing unit 11 inquires the tree data structure 3 , it can know that the categories of washing machine, raincoat, shirt and skirt have associated links 34 . Therefore, when the deep learning classification model 142 judges that an input file belongs to the category of washing machine, raincoat, shirt or skirt, but the reliability is not enough, it can make the machine learning comparison The model 143 is used to compare the input file with associated subordinate categories to confirm the category to which the input file belongs.

Please also refer to FIG. 4 , which is a first specific embodiment of the flow chart of the classification method of the present invention. The present invention further discloses an automatic digital file classification method (hereinafter referred to simply as a classification method in the description), the classification method is applied to the classification system 1 shown in Figure 1, and uses the method shown in Figure 2, Figure 3A and Figure 3B The data structure shown to automatically classify the input files.

As shown in Figure 4, to implement the classification method of the present invention, first the classification system 1 receives the input files to be classified through the input unit 12 or the network unit 13 (step S10), and the input files are digitized digital multimedia file. Next, the classification system 1 performs a classification procedure on the input file based on the pre-trained deep learning classification model 142 (step S12). After the classification program is executed, a preliminary classification result will be output, which shows the possible category of the input file ( hereinafter referred to as the first category), and the corresponding credibility.

For example, when the input file is a photo related to a dishwasher, the classification system 1 can use the deep learning classification model 142 to determine that the input file belongs to the category of dishwasher in step S12, And the reliability is 99%. For another example, the input file is a photo related to the dishwasher, but in addition to the dishwasher, the photo also includes object features such as plates, cups, and counters. In this embodiment, the deep learning classification model 142 may determine that the input file belongs to the dishwasher category in the classification procedure, and the confidence level is 70%.

In order to solve the problem of inaccurate classification results caused by insufficient credibility, the classification system 1 can read the association data structure 144 from the storage unit 14 after step S12 to confirm whether there is any At least one subordinate category associated with the first category. If the association data structure 144 shows that the first category is associated with at least one subordinate category (for example, as shown in FIG. 3B, the category of the dishwasher is associated with the category of the dishes), then the classification system 1 is based on the pre-trained machine. The learning comparison model 143 performs a feature comparison procedure on the input file and at least one digital file in the subordinate category. If the comparison result of the feature comparison program shows that the input file meets a preset classification condition (for example, the similarity with the subordinate category exceeds 70%), the classification system 1 can mark the input file as the first category, and The confidence that the input file belongs to the first category can be increased.

More specifically, after step S12, the classification system 1 can first confirm whether the reliability of the classification result of the deep learning classification model 142 is higher than or equal to a preset threshold value (step S14), such as 70%, 80%, etc., Not limited. If the reliability is higher than or equal to the threshold value, it means that the classification result of the deep learning classification model 142 is accurate, so the classification system 1 can directly mark the input file as the first category (step S16), and end the entire classification procedure .

If it is determined in step S14 that the reliability is lower than the threshold value, it means that the classification result of the deep learning classification model 142 is not accurate enough. At this time, the classification system 1 reads the relational data structure 144 (step S18), and confirms whether there is a first category-related at least one subordinate category of the link (step S20).

If it is determined in step S20 that the first category is related to at least one subordinate category, the classification system 1 can execute a feature comparison procedure through the machine learning comparison model 143 . Specifically, the machine learning comparison model 143 can extract at least one object feature corresponding to the input file based on the subcategory (step S22), whereby at least one of the input file and the subcategory can be compared based on the object feature. The digital file is subjected to a feature comparison program (step S24). That is, it is determined whether the input file has the same or similar object characteristics as the digital files in this subordinate category. After the feature comparison program is completed, the classification system 1 judges whether the comparison result of the feature comparison program meets the preset classification conditions (step S26 ).

If the comparison result of the feature comparison program shows that the input files do not meet the preset classification conditions, the classification system 1 can display a classification failure message through the output unit 15 (step S28 ). Conversely, if the comparison result shows that the input file meets the classification condition, the classification system 1 may mark the input file as the first category (step S30), and correct the credibility of the input file belonging to the first category ( Step S32). For example, the classification system 1 may revise the credibility so that the credibility is higher than or equal to the threshold value used in step S14.

It is worth mentioning that the classification system 1 judges in step S20 that the first category is not associated with any subordinate category, because the classification procedure of the first stage (that is, the classification procedure executed by the deep learning classification model 142) The reliability of the classification results is insufficient, and there is no comparison object that can carry out the comparison program of the second stage (that is, the comparison program performed by the machine learning comparison model 143), so the classification system 1 can directly pass the output The unit 15 displays a message of classification failure (step S28).

Please refer to FIG. 5 , which is the first specific embodiment of the feature comparison diagram of the present invention. Figure 5 takes the product photo used in the online store as an example to illustrate.

As shown in FIG. 5 , the appliance category 41 includes the category of the washing machine 411 , and the photo of the washing machine 411 includes the clothes feature 5 . Outdoor class 42 includes classes for rainwear 421 No, and the photo of the raincoat 421 contains the clothing feature 5. The clothing category 43 includes the category of the shirt 431 , and the photo of the shirt 431 includes the clothing feature 5 . Therefore, when designing the relational data structure 144 , the manager of the classification system 1 can establish the relation among the categories of the washing machine 411 , the category of the raincoat 421 and the category of the shirt 431 .

For example, the manager can connect the nodes 21 representing the washing machine 411, the raincoat 421 and the shirt 431 through the edge 22 in the graph data structure 2, or describe the nodes 21 representing the washing machine 411, the raincoat 421 and the shirt 431 through the attribute 33 in the tree data structure 3. The relationship between the child nodes 32 of the shirt 431.

With the above correlation, the classification system 1 can compare the input file with the raincoat 421 and/or the shirt 431 through the machine learning comparison model 143 when the deep learning classification model 142 judges that the input file belongs to the category of the washing machine 411, but the reliability is not enough. The feature comparison is performed on the digital files in the category of , so as to correct the credibility that the input file is the category of the washing machine 411 . Similarly, when the deep learning classification model 142 judges that the input file belongs to the category of the raincoat 421 but the reliability is not enough, the classification system 1 can also use the machine learning comparison model 143 to compare the input file with the category of the washing machine 411 and/or the shirt 431 Compare the digital files of , to correct the credibility of the input file as the category of raincoat 421, and so on.

In one embodiment, after the classification process is completed, the classification system 1 can output the classification result and the corresponding reliability through the output unit 15, so that the user can perform subsequent processing on the input file. In another embodiment, after the classification process is completed, the classification system 1 can move and store the input file to the corresponding database of the category according to the classification result. In this way, the input file can be archived automatically, and it is beneficial to retrain the deep learning classification model 142 .

Please refer to FIG. 6 and FIG. 7 , which are respectively the first specific embodiment and the second specific embodiment of the feature correlation schematic diagram of the present invention.

The embodiment in FIG. 6 takes the data structure used for classifying digital multimedia files in the medical field as an example. In the embodiment of Figure 6, the category of ultrasound includes the category of benign tumors and the category of malignant tumors, the category of computerized tomography includes the category of benign tumors and the category of malignant tumors, and the category of nuclear magnetic resonance also includes categories of benign tumors and malignant tumors. Moreover, the deep learning classification model 142 has been trained by the above category labels.

When a tumor image is input into the classification system 1, the deep learning classification model 142 preliminarily judges that the tumor image belongs to a benign tumor in the ultrasound category, but the reliability is not high. In this case, the classification system 1 can refer to the association described in the association data structure 144 to compare the tumor image with other benign tumor images in the ultrasound category through the machine learning comparison model 143, and compare This tumor image is subjected to a feature comparison procedure with images of benign tumors in the MRI category. Whether the tumor image is indeed a benign tumor can be confirmed through a feature comparison program (ie, judging whether the two have the same or similar object features).

Specifically, since the ultrasonic image and the nuclear magnetic resonance image may have the same object feature, the object feature may be less obvious in the ultrasonic image but more obvious in the nuclear magnetic resonance image. Therefore, the manager of the classification system 1 can establish the correlation between the benign tumor in the ultrasound category and the benign tumor in the MRI category in the correlation data structure 144 . The reliability of the classification results of the deep learning classification model 142 can be revised by comparing the tumor images with images of benign tumors in other related categories.

The embodiment of FIG. 7 takes the data structure used for classifying products on the production line as an example. In the embodiment of Fig. 7, the category of the first part includes the category of good products and the category of defective products. The category of products, and the category of defective products includes defective categories such as improper assembly, foreign objects, and machine aging. The category of the second part also includes the category of good products and the category of defective products, and the category of defective products includes the categories of defects such as improper assembly, foreign objects, and machine aging. Moreover, the deep learning classification model 142 has been trained by the above category labels.

When a defect image is input into the classification system 1, the deep learning classification model 142 preliminarily judges that the defect image belongs to the defective product in the first part category, and belongs to the aging defect of the machine, but the reliability is not high. In this case, the classification system 1 can refer to the association described in the association data structure 144, and use the machine learning comparison model 143 to compare the defective image with other machines under the defective product category in the first component category A feature comparison program is performed on the aging image, or a feature comparison program is performed on the defective image and the machine aging image under the defective product category in the second part category. Through the feature comparison program, it can be confirmed whether the defect image really belongs to the defect category of machine aging.

As mentioned above, the classification system 1 of the present invention will use the machine learning comparison model 143 to compare the input file with the associated subordinate category when the reliability of the classification result of the deep learning classification model 142 is insufficient. right. When the category to which the input file belongs is associated with multiple subcategories (that is, has multiple subcategories), the classification system 1 can perform multiple feature comparison procedures to avoid misjudgment by a single feature comparison procedure.

As shown in FIG. 8 , in step S20 of FIG. 4 , the classification system 1 may determine based on the content of the association data structure 144 that the first category to which the input file belongs is associated with a plurality of subordinate categories at the same time (step S40 ). At this point, the machine learning comparison model 143 extracts at least one object feature of the input file based on one of the plurality of subcategories (step S42), and at least one digit of the input file and the subcategory is compared based on at least one object feature. The profile undergoes a feature comparison process (step S44) and obtains a similarity score (step S46).

After step S46, the classification system 1 completes the comparison process between the input file and one of the subcategories through the machine learning comparison model 143, and judges whether the multiple subcategories related to the first category are all compared (step S48).

Before the comparison of the plurality of subcategories is completed, the classification system 1 repeatedly executes the steps S42, S44 and S46 to perform a feature comparison procedure on the input files and other subcategories through the machine learning comparison model 143, and obtain Corresponding similarity scores. After the comparison of the plurality of subordinate categories is completed, the classification system 1 can calculate the comparison result of the feature comparison program according to the similarity scores of each subordinate category and a preset weight value (step S50 ).

The preset weight value can be described by, for example, each edge 22 in the graph data structure 2 , or described by each attribute 33 in the tree data structure 3 , but not limited thereto. Moreover, in one embodiment, the comparison result may be a numerical value, such as a percentage, but it is not limited thereto. In other words, the comparison result is calculated based on the similarity scores between the first category and each subordinate category and their importance.

After step S50, the classification system 1 can judge whether the comparison result meets the preset classification condition (step S52), that is, confirm whether the classification result of the deep learning classification model 142 is credible according to the comparison result. If it is determined in step S52 that the comparison result does not meet the classification condition, the classification system 1 may output a classification failure message through the output unit 15 (step S54 ). On the contrary, if it is judged in step S52 that the comparison result meets the classification condition, the classification system 1 can directly mark the input file as the first category (step S56), and correct the credibility of the input file belonging to the first category (step S58) .

The present invention realizes the classification program of the first stage and the comparison program of the second stage through deep learning and machine learning, and can effectively improve the accuracy rate of the classification result of the single-stage classification program.

The above descriptions are only preferred specific examples of the present invention, and are not intended to limit the patent scope of the present invention. Therefore, all equivalent changes made by using the content of the present invention are all included in the scope of the present invention. bright.

S10~S32: classification steps

Claims (9)

An automatic digital file classification method, applied to a classification system, and comprising: a) receiving an input file by the classification system; b) the classification system performing a classification procedure on the input file based on a pre-trained deep learning classification model , wherein the classification program outputs a possible first category of the input file and a corresponding reliability; b1) judging whether the reliability is higher than or equal to a threshold; b2) when the reliability is higher than or equal to When the threshold value is reached, the input file is directly marked as the first category; c) when the reliability is lower than the threshold value, the classification system reads a relational data structure to confirm that it is associated with the first category , wherein the association data structure records an association between a predefined plurality of classes and one or more of the subordinate classes, wherein the subordinate class is one of the plurality of class tags with the first class or other categories of common or similar object characteristics; d) the classification system performs a feature comparison process on the input file and at least one digital file in the subordinate category based on a machine learning comparison model that has been pre-trained; and e) marking the input file as the first category when a comparison result of the feature comparison program meets a classification condition, and revising the confidence that the input file belongs to the first category. The automatic digital file classification method as described in claim item 1, wherein the associated data structure is a tree-like data structure with multiple root nodes and multiple child nodes, wherein one or more of the child nodes are respectively set with an attribute, and the attribute describes The association between this child node and another child node. The automated digital file classification method as described in Claim 1, wherein the association data structure is a graph data structure with multiple nodes and multiple edges, each of which connects two nodes to describe the relationship between the two nodes sex. The automatic digital file classification method as described in claim 2 or 3, wherein the step d) is to extract at least one object feature of the input file based on the subordinate category by the machine learning comparison model, and based on the at least one object The feature comparison procedure is performed on the input file and at least one digital file in the subcategory to determine whether the input file is similar to the subcategory. The automatic digital archive classification method as described in claim 2 or 3, wherein the step c) reads the association data structure and confirms the plurality of subordinate categories associated with the first category, and the step d) includes: d1) Extracting at least one object feature of the input file based on one of the plurality of subcategories by the machine learning comparison model; d2) performing an operation on the input file and at least one digital file in the subcategory based on the at least one object feature The feature comparison program, and obtain a similarity score; d3) determine whether the plurality of subcategories are compared; d4) before the plurality of subcategories are compared, repeat step d1) to step d2 ) to obtain the similarity score between the input file and other subcategories; and d5) after the comparison of the plurality of subcategories is completed, calculate the The result of this alignment for the feature alignment program. An automatic digital file classification system, comprising: an input unit for receiving an input file to be classified; A storage unit storing a deep learning classification model, a machine learning comparison model, and an associative data structure, wherein the associative data structure records an association between predefined plural categories and one or more subcategories; a The processing unit is connected to the input unit and the storage unit, and executes the following procedure: performing a classification procedure on the input file based on the pre-trained deep learning classification model, wherein the classification procedure outputs the possible first category of the input file and a corresponding credibility; after executing the classification program, it is judged whether the credibility is higher than or equal to a threshold; when it is judged that the credibility is higher than or equal to the threshold, the input file is directly marked as the first category; when it is determined that the reliability is lower than the threshold value, reading the association data structure to confirm at least one subcategory associated with the first category, wherein the subcategory is in a plurality of categories Other categories having one or more common or similar object characteristics with the first category in the label; performing a feature comparison between the input file and at least one digital file in the subordinate category based on the pre-trained machine learning comparison model a matching program; and marking the input file as the first category when a comparison result of the feature matching program meets a classification condition, and revising the confidence that the input file belongs to the first category; and an output A unit, connected to the processing unit, outputs the first category to which the input file belongs and the corresponding reliability. The automatic digital file classification system as described in claim 6, wherein the associative data structure is a tree-like data structure or a graph data structure, wherein the tree-like data structure has multiple root nodes and multiple child nodes, and one or more of the The sub-nodes are respectively set with an attribute, which describes the sub-section A relationship between a point and another child node; wherein the graph data structure has a plurality of nodes and a plurality of edges, and each of the edges connects two nodes to describe the relationship between the two nodes. The automated digital file classification system as described in claim 7, wherein the machine learning comparison model is configured to extract at least one object feature of the input file based on the subordinate category, and to extract the input file based on the at least one object feature and at least one digital file in the subordinate category to perform the feature comparison procedure. The automated digital archive classification system as described in claim 7, wherein the processing unit is configured to confirm the plurality of subordinate categories associated with the first category after reading the association data structure, and based on the pre-trained completed The machine learning comparison model performs a feature comparison procedure on the input file and at least one digital file in the subcategory, performing the following steps: a) the machine learning comparison model based on one of the plurality of subcategories - extracting at least one object feature of the input file; b) performing the feature comparison procedure on the input file and at least one digital file in the subordinate category based on the at least one object feature, and obtaining a similarity score; c) Judging whether the comparison of the plurality of subcategories is completed; d) before the completion of the comparison of the plurality of subcategories, repeat step a) and step b) to obtain the similarity between the input file and other subcategories and e) after the comparison of the plurality of subcategories is completed, the comparison result of the feature comparison program is calculated according to the relevance score of each of the subcategories and a preset weight value.

Images