WO2020191988A1 - 新类别识别方法和基于模糊理论和深度学习的机器人系统 - Google Patents
新类别识别方法和基于模糊理论和深度学习的机器人系统 Download PDFInfo
- Publication number
- WO2020191988A1 WO2020191988A1 PCT/CN2019/099963 CN2019099963W WO2020191988A1 WO 2020191988 A1 WO2020191988 A1 WO 2020191988A1 CN 2019099963 W CN2019099963 W CN 2019099963W WO 2020191988 A1 WO2020191988 A1 WO 2020191988A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- category
- new category
- new
- membership
- degree
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
Definitions
- the invention relates to the field of information technology, in particular to a new category recognition method and a robot system based on fuzzy theory and deep learning.
- the existing object recognition technology can realize the recognition of the object through the characteristics of the known category, such as the method of deep learning to recognize the object.
- the inventor found that the prior art has at least the following problems: 1.
- the existing recognition technology can only identify known categories, but cannot identify new categories; 2. Once a new category appears, the existing recognition system will Invalidate.
- an embodiment of the present invention provides an identification method, the method including:
- Recognition step Recognizing the first object to obtain the first result
- New category judging step judging whether the category of the first object belongs to the set of first known categories according to the first result: if no, the category of the first object is taken as the first new category; if yes, then The category of the first object is a known category.
- the first result includes at least one known category to which the first object belongs and the degree of membership of each known category of the first object belonging to the at least one known category;
- the step of judging the new category specifically includes: taking the maximum membership degree of the membership degrees of the first object belonging to each of the at least one known category; if the maximum membership degree is less than or equal to the preset first object A degree of membership, the category of the first object is the first new category; if the maximum degree of membership is greater than the preset first degree of membership, the known category of the first object is the one corresponding to the maximum degree of membership Known category.
- the method further includes:
- New category set judging step judging whether the first new category belongs to the second new category set according to the first result: yes, then the first new category corresponding to the second new category set The new category serves as the first new category; if not, the first new category is added to the set of the second new category.
- the step of judging the new category set includes:
- New category set membership degree calculation step calculating the membership degree of each new category in the set where the first new category belongs to the second new category;
- the new category set is added to the judging step: acquiring the largest degree of membership in the degree of membership, and judging whether the largest degree of membership is greater than the preset second degree of membership: yes, then the second new degree of membership corresponding to the maximum degree of membership
- a new category in the set of categories serves as the first new category; if not, the first new category is added to the set of second new categories.
- the judging step of adding the new category set further includes:
- the maximum degree of membership is greater than the preset second degree of membership, then determine whether the first object has a category label: yes, use the category label as the label of the first new category, and set the The label of a new category is used as the label of the new category in the set of the second new category corresponding to the maximum degree of membership; if not, the label of the second new category corresponding to the maximum degree of membership is obtained The label of the new category is used as the label of the first new category;
- the maximum degree of membership is not greater than the preset second degree of membership, it is determined whether the first object has a category label: yes, the category label is used as the label of the first new category, and the The label of the first new category is used as the label of the first new category in the set of the second new category; if not, the label of the first new category is generated, and the label of the first new category is used as the label of the first new category.
- the step of calculating the membership degree of the new category set further includes:
- the step of adding the new category set to judging further includes:
- the first result is taken as the first portrait of the first new category, and the first portrait of the first new category is merged with the maximum membership
- the second portrait of the new category in the set of the second new category corresponding to the degree of membership obtains the third portrait; the third portrait is taken as all of the set of the second new category corresponding to the maximum degree of membership. State the second portrait of the new category;
- the first portrait of the first new category is taken as the second portrait of the first new category in the set of second new categories.
- the step of judging the new category set further includes:
- the number of samples of the new category in the set of the second new category corresponding to the largest degree of membership is increased by 1, and the first object is added In the sample set of the new category in the set of the second new category corresponding to the largest degree of membership;
- the first new category does not belong to the set of the second new category, after adding the first new category to the set of the second new category, add the The sample number of the first new category is set to 1, and the first object is added to the sample set of the first new category in the set of the second new category;
- the method also includes:
- Existing recognition system obtaining step obtaining the first recognition device or system that recognizes the first object in the recognition step;
- the step of generating a new recognition system obtaining each new category whose number of samples is greater than the preset number of samples from the set of the second new category; obtaining samples in the set of samples of each new category, and dividing each new category
- the category is deleted from the set of the second new category, and each of the new categories is added to the set of the first known category to obtain the set of the second known category, through the sample and each new category
- the tag of trains the first recognition device or system, and obtains the recognition device or system capable of recognizing each new category as the second recognition device or system.
- the first recognition device or system includes a deep learning model; an input variable of the deep learning model is an object; an output variable of the deep learning model is each category in the set of the first known categories and belongs to the The degree of membership of each category;
- the step of generating the new recognition system specifically includes:
- each new category whose number of samples is greater than the preset number of samples from the set of the second new category; obtain samples in the set of samples of each new category, and change each new category from the second new category Is deleted from the set, and each new category is added to the set of the first known category to obtain the set of the second known category, and each category in the set of the second known category belongs to all
- the membership degree of each category is used as the output variable of the deep learning model, and the deep learning model is trained through the samples in the sample set of each new category and the label of each new category to obtain Recognize the deep learning model of each new category as the second recognition device or system.
- an embodiment of the present invention provides an identification device that executes the identification method according to any one of the embodiments of the first aspect.
- an embodiment of the present invention provides a robot system, the robot system including the device according to any one of the embodiments of the second aspect.
- the method and system in the embodiment of the present invention can identify new categories.
- the existing recognition technology uses methods such as machine learning or deep learning to recognize existing categories through features or samples of existing categories. If an object belonging to a new category is input, it cannot be recognized.
- the method and system in the embodiments of the present invention can automatically define objects belonging to a new category as a new category, and when there are objects belonging to a new category of the same type in the future, they can be automatically recognized as the previously defined new category. Realize the recognition of new categories.
- the method and system in the embodiments of the present invention have the ability to recognize new categories so that the robot can not only recognize based on the results of previous machine learning or deep learning, but also recognize known categories specified by humans through programs or data , And can extend the recognition ability to objects belonging to new categories, so that the robot can perform autonomous recognition, thereby laying the foundation for the robot to have autonomous consciousness.
- the method and system in the embodiments of the present invention establish an association between the new category and the existing category through the membership degree in the fuzzy theory, and calculate the membership degree of the new category belonging to the existing category and the new category belonging to the identified new category The degree of membership of, so as to identify new categories through existing categories, and achieve the ability of association recognition similar to humans.
- the new category recognition method and the robot system based on fuzzy theory and deep learning provided by this embodiment include: recognizing a first object to obtain a first result; judging whether the category of the first object belongs to the first result according to the first result A set of known categories: if no, the category of the first object is taken as the first new category; if yes, the category of the first object is a known category.
- the category of the first object is determined according to the first result. If the first object does not belong to the set of first known categories, the first object to which the first object belongs is obtained. New category, which can realize the recognition of the new category.
- FIG. 1 is a flowchart of the identification method provided by Embodiments 1 and 2 of the present invention.
- FIG. 3 is a flowchart of the identification method provided by the embodiments 4, 5, and 6 of the present invention.
- FIG. 5 is a flowchart of the identification method provided by Embodiments 8 and 9 of the present invention.
- Embodiment 10 is a flowchart of the identification method provided by Embodiment 10 of the present invention.
- FIG. 7 is a flowchart of the identification method provided by Embodiment 11 of the present invention.
- FIG. 8 is a functional block diagram of the identification system provided by Embodiments 12 and 13 of the present invention.
- FIG. 9 is a functional block diagram of an identification system provided by Embodiment 14 of the present invention.
- Fig. 10 is a functional block diagram of the recognition system provided by embodiments 15, 16, and 17 of the present invention.
- FIG. 11 is a functional block diagram of an identification system provided by Embodiment 18 of the present invention.
- FIG. 12 is a functional block diagram of the recognition system provided by Embodiments 19 and 20 of the present invention.
- FIG. 13 is a functional block diagram of an identification system provided by Embodiment 21 of the present invention.
- FIG. 14 is a functional block diagram of an identification system provided by Embodiment 22 of the present invention.
- an identification method includes:
- Recognition step S10 the first object is recognized by the first recognition device or system to obtain the first result.
- the method of recognizing the first object includes using machine learning or deep learning or other recognition methods.
- the first recognition device or system includes one or more of a machine learning system, a deep learning system, and the like.
- the objects include people, objects, scenes, events, etc.
- the formats of the objects include images, videos, texts, etc., during recognition, the people, objects, scenes, events, etc. are identified from the images, videos, or texts, etc. Category. Before this step, the first object to be recognized needs to be acquired.
- New category judging step S20 judging whether the category of the first object belongs to the set of first known categories according to the first result: if no, the category of the first object is taken as the first new category; if yes, then The category of the first object is a known category.
- the set of first known categories is a set of categories that can be recognized by the first recognition device or system.
- the first recognition device or system has already determined the category set that can be recognized by the first recognition device or system when it is constructed or trained.
- This embodiment adds some steps to distinguish the known category from the new category by the degree of membership, which is a more specific implementation scheme of the first embodiment.
- the identification method provided according to Embodiment 1 further includes:
- the first result in the identifying step S10 includes at least one known category to which the first object belongs and the degree of membership of the first object to each known category in the at least one known category to which it belongs. (The degree of membership can also be probability or probability).
- the new category judging step S20 includes: taking the maximum membership degree of the membership degrees of each known category in the at least one known category to which the first object belongs, and determining whether the maximum membership degree is greater than the preset first object.
- the degree of membership that is, whether the category of the first object belongs to the set of the first known category
- the maximum degree of membership is greater than the preset first degree of membership (that is, the category of the first object belongs to the first known Category set)
- the known category of the first object is the known category corresponding to the maximum degree of membership
- the maximum degree of membership is less than or equal to the preset first degree of membership (that is, the The category does not belong to the set of the first known category)
- the category of the first object is the first new category.
- the first object is a photo of a person, and the first result is that the person in the photo belongs to the existing category Zhang San has a membership degree (or probability or probability) of 10% and belongs to the existing category Li Si
- the degree of membership is 20%, and the membership of the existing category Wangwu is 70%.
- the maximum degree of membership at this time is 70%.
- the default first degree of membership is 69%. Therefore, the maximum degree of membership is greater than the preset first degree of membership.
- the category corresponding to the maximum degree of membership is Wangwu, so the category of the first object is the existing category Wangwu.
- the first object is a photo of a person, and the first result is that the person in the photo belongs to the existing category.
- the membership degree (or probability or probability) of Zhang San is 30% and belongs to the existing category.
- the membership of the fourth category is 40%, and the membership of the existing category Wangwu is 30%.
- the maximum degree of membership at this time is 40%.
- the default first degree of membership is 69%. Therefore, the maximum degree of membership is less than the preset first degree of membership. Therefore, the category of the first object is the first new category.
- This embodiment adds some steps to distinguish whether it is a new category in the second new category set by the degree of membership. This is a supplement to the steps of the first embodiment, because the embodiment only distinguishes existing categories and new categories. It does not distinguish whether the new category is a new category that has been identified in the past, that is, a new category that is already in the second new category set. This embodiment also distinguishes whether the first new category belongs to the set of the second new category based on the degree of membership.
- the third embodiment is executed only when the category of the first object is not a known category. It will not be executed when the category of an object is a known category.
- the method further includes:
- Category set judging step S30 Judge whether the first new category belongs to the second new category set according to the first result: if yes, then the first new category corresponding to the second new category set The new category serves as the first new category; if not, the first new category is added to the set of the second new category. Before this step, it is necessary to obtain the set of the second new category, and determine whether the set of the second new category exists. If it does not exist, create an empty set of the second new category. Obviously, the first new category Does not belong to the second new category collection.
- This embodiment adds some steps for judging whether it belongs to the second new category set through the degree of membership, which is a more specific implementation solution of the third embodiment.
- the category set judgment step S30 includes:
- New category set membership degree calculation step S31 Calculate the membership degree of each new category in the set where the first new category belongs to the second new category. Before this step, it is necessary to obtain the set of the second new category, and determine whether the set of the second new category exists. If it does not exist, create an empty set of the second new category. Obviously, the first new category The membership degree of each new category in the set belonging to the second new category is zero.
- the new category set is added to the judgment step S32: Obtain the largest degree of membership in the degree of membership, and judge whether the largest degree of membership is greater than the preset second degree of membership (obviously the greater the degree of membership, the first new category belongs to the second new The greater the possibility of a new category in the set of categories): yes, the new category in the set of the second new category corresponding to the maximum degree of membership is taken as the first new category; if not, the first new category A new category is added to the set of the second new category (because the first new category does not belong to the set of the second new category, it needs to be added).
- This embodiment adds some steps for generating tags of new categories.
- the new category set addition judgment step S32 further includes: if the maximum degree of membership is greater than the preset second degree of membership, then a new category in the set of the second new category corresponding to the maximum degree of membership is used as the first new category Afterwards, it is determined whether the first object has a category label, and if it is (inputting a labeled object belonging to a new category into the first recognition device or system can make the new category in the second new category set have a real label), then Use the category label as the label of the first new category, and use the label of the first new category as the label of the new category in the set of the second new category corresponding to the maximum degree of membership, if not , The label of the new category in the set of the second new category corresponding to the maximum degree of membership is obtained as the label of the first new category; if the maximum degree of membership is not greater than the preset second degree of membership, then After the first new category is added to the set of the second new category, it is determined whether the first object has a category label,
- the label of the first new category needs to be different from the label of the existing category and also from the label of the second new category.
- the label of any new category in the set (because the label of each category needs to be unique, otherwise it is difficult to distinguish each category, the specific steps of generating the label of the first new category include L: randomly generate a code, and judge the code Whether it is the same as the label of the existing category and the label of any new category in the set of the second new category: yes, go back to L and re-execute; if not, use the code as the label of the first new category.
- the representation method of the new category in the second new category set includes the use of an automatically generated unique identifier, and the unique identifier can be randomly generated by existing technology.
- the new category representation method in the second new category set is also Including the direct use of labels to represent the new category in the second new category set.
- This embodiment adds some steps for generating the first portrait of the new category.
- the new category set membership degree calculation step S31 further includes: using the first result as the first portrait of the first new category.
- Portrait refers to the portrait in the portrait technology in big data and artificial intelligence technology, and the portrait can also be called a feature
- the first new category belongs to the degree of membership of each new category in the set of the second new category.
- the certain category in the set of the second new category is calculated when calculating the similarity.
- the degree of membership in the second portrait of the new category belonging to this category is regarded as 0, or when there is a certain category in the second portrait of a new category in the set of the second new category but there is no such category in the first portrait
- the membership degree of the category in the first portrait is regarded as 0, which is convenient for calculating the similarity.
- the calculation method of similarity includes the method of calculating cosine similarity; the method of calculating cosine similarity includes formula 1:
- Xi refers to the degree of membership belonging to the i-th category in the first portrait
- Yi refers to the degree of membership belonging to the i-th category in the second portrait. If the i-th category does not exist in the first portrait, Xi is 0, if If the i-th category does not exist in the second portrait, Yi is 0.
- the new category set addition judgment step S32 further includes: if the maximum degree of membership is not greater than the preset second degree of membership, after adding the first new category to the set of the second new category (the portrait and the label are If it is associated with the category, the association can be carried out through technologies such as a database, knowledge base, or big data), and the first portrait of the first new category is used as the first portrait in the set of the second new category.
- the second portrait of the new category if the maximum degree of membership is greater than the preset second degree of membership, after taking the new category in the set of the second new category corresponding to the maximum degree of membership as the first new category, The first result is used as the first portrait of the first new category, and the first portrait of the first new category is merged with the first portrait of the new category in the set of the second new category corresponding to the maximum degree of membership.
- the second portrait gets the third portrait (that is, the first portrait and the second portrait are merged to get the third portrait.
- the composition of the second portrait is similar to the first portrait
- the first portrait and the second portrait include at least one category and the degree of membership belonging to each category in the at least one category; wherein, during the fusion, the categories in the first portrait and the second portrait are merged as the first Category in the three portraits; weighted average the membership degrees of each category in the first portrait and the second portrait to obtain the membership degrees of each category in the third portrait), and use the third portrait as the
- the second portrait of the new category in the set of the second new category corresponding to the maximum degree of membership (update the portrait in the set of the second new category, so that the portrait of the new category can better represent the new category)
- the first portrait, the second portrait, and the third portrait are all portraits, just to look clearer, so add the first, second, and third).
- the weighted average method includes ((the degree of membership belonging to each category in the first portrait) ⁇ K1+(the degree of membership belonging to each category in the second portrait) ⁇ K2)/(K1+K2).
- K1 and K2 are generally taken as 1, or K1 is taken as 1, and K2 is taken as the number of existing samples of the new category in the set of the second new category corresponding to the second portrait.
- S30 needs to be extended to: judging whether the first new category belongs to the set of the second new category according to the first result: if yes, use the new category in the set of the second new category corresponding to the maximum degree of membership as the For the first new category, add 1 to the number of samples of the new category in the set of the second new category corresponding to the maximum degree of membership; if not, add the first new category to the set of the second new category After that, the number of samples of the first new category in the set of the second new category is set to 1.
- the new category set is added to the step of fusing the first portrait and the second portrait to obtain the third portrait described in step S32. If there is a certain category in the first portrait and the second portrait does not have that category, then When calculating the membership degree of this category in the third portrait, consider the membership of this category in the first portrait as 0; or if there is a certain category in the second portrait but not in the first portrait, then When calculating the membership degree of this category in the third portrait, consider the membership degree of this category in the first portrait as 0. This processing is to facilitate weighted average, because the categories in the first portrait and the second portrait In most cases it is the same, but it can also be different.
- the membership degree of Zhang San belonging to the existing category is 30%, the membership degree of the existing category Li Si is 40%, and the membership degree of Wang Wu belonging to the existing category is 30%;
- the membership degree of the existing category Zhang San is 32%, the membership degree of the existing category Li Si is 38%, and the membership degree of the existing category Wang Wu is 30%;
- the similarity between the first portrait and the second portrait is calculated , It can be calculated by calculation methods including cosine similarity; if the first portrait and the second portrait are merged into the third portrait, the categories in the third portrait are Zhang San, Li Si, Wang Wu, and the third In the portrait, the membership degree of Zhang San belonging to the existing category is 30% plus 32% divided by 2 or 31%, and the membership degree of the existing category Li Si is 40% plus 38% divided by 2 or 39%, belonging to the existing category Wang Wu’s membership is 30% plus 30% divided by 2, or 30%.
- the membership of the existing category Zhang San is 30%, the membership of the existing category Li Si is 30%, and the membership of the existing category Wang Wu is 30%, which belongs to the existing category.
- Zhu Liu s membership degree is 10%
- the membership degree of Zhang San belonging to the existing category is 32%, the membership degree of the existing category Li Si is 38%, and the membership degree of the existing category Wang Wu is 30. %; If the similarity between the first portrait and the second portrait is calculated, it can be calculated by calculation methods including cosine similarity; if the first portrait and the second portrait are merged into the third portrait, the third portrait
- the category of is Zhang San, Li Si, Wang Wu, Zhu Liu.
- the membership degree of Zhang San belonging to the existing category in the third portrait is 30% plus 32% divided by 2, or 31%, which belongs to the existing category Li Si 30% plus 38% divided by 2 is 34%, the membership of the existing category Wangwu is 30% plus 30% divided by 2 is 30%, the membership of the existing category Zhuliu is 10% plus 0% Divide by 2 or 5%).
- This embodiment adds some steps to update the tags of the new category.
- the tags in the set of the second new category can be updated, and the tags in the set of the second new category are generated in Example 3.
- the tags have been generated in Example 5, but the tags generated at this time are not yet understandable tags by humans.
- Label acquiring step S40 acquiring the new category of the label to be updated in the set of the label and the second new category.
- Label update step S50 Use the acquired label as a label of a new category of labels to be updated in the set of the second new category.
- the first label may be assigned by the user, or obtained through other means, for example, obtained through the Internet.
- the set of the second new category has been generated through Example 3, so that the new object can also be recognized as a new category in the set of the second new category.
- the label of the new category is generated through Example 5, but it is generated in Example 5.
- the label of is still different from the label in our human concept, such as New1, New2, etc.
- the user may not know the label of the category of the new object when the new object first appears, it is possible for the user to know the true label of the new category over time.
- Another method for updating tags is to input tagged objects belonging to the new category into the first recognition device or system, and it can also make the new category in the second new category set have real tags .
- This implementation adds some steps to convert the new category into a known category that the recognition device or system can recognize.
- the category set judging step S30 further includes: if the first new category belongs to the set of the second new category, the new category in the set of the second new category corresponding to the maximum degree of membership is used as the first new category After that, the number of samples of the new category in the set of the second new category corresponding to the maximum degree of membership is increased by 1, and the first object is added to the new category in the set of the second new category corresponding to the maximum degree of membership If the first new category does not belong to the set of the second new category, after adding the first new category to the set of the second new category, add the set of the second new category The number of samples in the first new category is set to 1, the sample set of the first new category in the set of the second new category is created (empty when created), and the first object is added to the first The sample set of the first new category in the set of two new categories.
- New recognition system generation step S70 Obtain each new category whose number of samples is greater than the preset number of samples from the set of the second new category; obtain samples in the set of samples of each new category, and compare each The new category is deleted from the set of the second new category, and each new category is added to the set of the first known category to obtain the set of the second known category.
- the category label trains the first recognition device or system, and obtains the recognition device or system capable of recognizing each new category as the second recognition device or system.
- the second recognition device or system capable of recognizing the new category can be used as the first recognition device or system in the recognition step S10, and the method in any of the above embodiments can be used again to obtain more recognition.
- the second recognition device or system of the new category enables more and more new categories to be recognized by the second recognition device or system.
- a new category is completely changed into a known category, so that the second recognition device or system can autonomously evolve and recognize the new category, not only the first recognition device or system can recognize Those known categories.
- the object of the new category is recognized next time, if the object belongs to the new category, the object is input into the second recognition device or system capable of recognizing the new category, through the second recognition device or The output of the system can directly identify the category of the object as the new category, because the new category has been added to the output variable of the second recognition device or system, and becomes a known category.
- This embodiment adds some steps to convert the new category into a known category that can be recognized by the deep learning recognition system.
- the first recognition device or system includes a deep learning model; an input variable of the deep learning model is an object; an output variable of the deep learning model is each category in the set of the first known categories and belongs to the The degree of membership of each category.
- the new recognition system generating step 70 specifically includes: obtaining each new category with a sample number greater than a preset sample number from the set of the second new category; obtaining samples in the sample set of each new category; Each new category is deleted from the second new category set, and each new category is added to the first known category set to obtain the second known category set, and the second known category Each category in the set and the degree of membership belonging to each category are used as the output variable of the deep learning model (because each new category is added to the set of the first known category, the second existing category is obtained).
- the set of known categories, and the output of the deep learning model is each category in the set of the first known category and the degree of membership belonging to each category, so the output variables of the deep learning model need to be updated) , Training the deep learning model through the samples in the sample set of each new category and the label of each new category to obtain a deep learning model capable of recognizing each new category as the second recognition Device or system.
- the step of training the deep learning model through the samples in the sample set of each new category and the label of each new category specifically includes: combining the samples in the sample set of each new category As the input data of the deep learning model, the membership degree corresponding to each new category in the expected output data of the deep learning model is set to 1, and the membership degrees corresponding to all other categories are set to 0.
- the learning model undergoes supervised training.
- This embodiment adds some steps of applying new objects based on portraits of new categories through portraits.
- the method further includes:
- Object category judging step S80 Determine whether the category of the object to be processed exists: No, use the method in any of the foregoing embodiments to perform identification to obtain the category of the object to be processed. Before this step, it also includes obtaining the object to be processed and the application corresponding to the processing. Applications corresponding to the processing include applications such as decision-making, analysis, statistics, etc., and can be used in industrial fields, military fields, and so on.
- Object information fusion step S90 Determine whether the category of the object to be processed belongs to the set of known categories: yes, then obtain the known category corresponding to the category of the object to be processed from the set of known categories ⁇ ; No, obtain the portrait of the category of the object to be processed from the set of the second new category, and obtain each known category and belong to the category from the portrait of the category of the object to be processed
- the membership degree of each known category is obtained, the information of each known category is acquired, and the information of each known category is fused according to the membership degree of each known category to obtain the Information about the category of the object to be processed.
- the information of the known category includes knowledge of the known category.
- the specific steps of the fusion include obtaining a preset function or module, taking the information of each known category and the degree of membership belonging to each known category as the input of the preset function or module, and calculating Obtain the information of the category of the object to be processed, and return the information of the category of the object to be processed to the application corresponding to the processing.
- Another simple implementation step of the fusion includes a weighted average of the information of each known category to obtain the information of the category of the object to be processed, and the weighted average of the information belonging to each The membership degree of the known category is used as the weight of the information of each known category.
- This embodiment adds some steps to identify the new category.
- the method further includes:
- Second object obtaining step SA0 obtaining a second object to be recognized
- Second object recognition step SB0 input the second object into the second recognition device or system for recognition, and obtain a second result
- Second object category judgment step SC0 Determine whether the category of the second object belongs to the set of second known categories according to the second result: No, use the category of the second object as the second new category; yes , The category of the first object is a known category. If the category of the first object is a known category, it is determined whether the category of the second object belongs to the set of the first known category: No, then the category of the second object is important to the first recognition device or system. It is the first new category; if yes, the category of the second object is a known category to the first recognition device or system.
- Embodiments 12-20 provide an identification device, which respectively correspond to Embodiments 1-11. The details will not be repeated here, and only the main modules corresponding to the main steps in Embodiments 1-11 are listed.
- an identification device includes:
- Recognition module 10 Recognize the first object through the first recognition device or system to obtain the first result.
- New category judging module 20 judging whether the category of the first object belongs to the set of first known categories according to the first result: if no, the category of the first object is taken as the first new category; if yes, then The category of the first object is a known category.
- the identification device provided according to Embodiment 12 further includes:
- the first result in the recognition module 10 includes at least one known category to which the first object belongs and the degree of membership of the first object belonging to each known category in the at least one known category.
- the new category judgment module 20 specifically includes: taking the maximum membership degree of the membership degrees of each of the at least one known category that the first object belongs to, and judging whether the maximum membership degree is greater than the preset first membership If the maximum degree of membership is greater than the preset first degree of membership, then the known category of the first object is the known category corresponding to the maximum degree of membership; if the maximum degree of membership is less than or equal to the preset first degree of membership A degree of membership, the category of the first object is the first new category.
- the method further includes:
- the set judgment module 30 of the second new category judge whether the first new category belongs to the set of the second new category according to the first result: if yes, set the second new category corresponding to the first new category A new category in the set of is used as the first new category; otherwise, the first new category is added to the set of the second new category.
- the set judgment module 30 of the second new category includes:
- New category set membership degree calculation module 31 Calculate the membership degree of each new category in the set where the first new category belongs to the second new category.
- the new category set is added to the judgment module 32: Obtain the largest degree of membership in the degree of membership, and determine whether the largest degree of membership is greater than the preset second degree of membership: yes, then the second degree of membership corresponding to the largest degree of membership A new category in the set of new categories serves as the first new category; otherwise, the first new category is added to the set of second new categories.
- the new category set addition judging module 32 further includes: if the maximum degree of membership is greater than the preset second degree of membership, judging whether the first object has a category label, and if so, using the category label as the first The label of the new category, and use the label of the first new category as the label of the new category in the set of the second new category corresponding to the maximum degree of membership; if not, obtain the maximum degree of membership The label of the new category in the corresponding set of the second new category is used as the label of the first new category; if the maximum degree of membership is not greater than the preset second degree of membership, then the first Whether the object has a category label, if yes, use the category label as the label of the first new category, and use the label of the first new category as the first new category in the set of the second new category If not, generate the label of the first new category, and use the label of the first new category as the label of the first new category in the set of second new categories.
- the new category set membership degree calculation module 31 further includes: taking the first result as the first portrait of the first new category. Calculate the similarity between the first portrait of the first new category and the second portrait of each new category in the set of the second new category, and regard the similarity as the first new category belonging to the The degree of membership of each new category in the set of the second new category.
- the calculation method of similarity includes a method of calculating cosine similarity; the method of calculating cosine similarity includes formula 1:
- Xi refers to the degree of membership belonging to the i-th category in the first portrait
- Yi refers to the degree of membership belonging to the i-th category in the second portrait.
- the new category set addition judging module 32 further includes: if the maximum degree of membership is not greater than the preset second degree of membership, taking the first portrait of the first new category as all of the sets of the second new category The second portrait of the first new category; if the maximum degree of membership is greater than the preset second degree of membership, the first result is taken as the first portrait of the first new category, and the first new category is merged
- the first portrait of the category and the second portrait of the new category in the set of the second new category corresponding to the maximum degree of membership obtain the third portrait; wherein, when merging, the first portrait and the The categories in the second portrait are merged as the categories in the third portrait; the membership degrees of each category in the first portrait and the second portrait are weighted and averaged to obtain the third portrait belonging to all categories.
- the degree of membership of each category; and the third portrait is taken as the second portrait of the new category in the set of the second new category corresponding to the largest degree of membership.
- the label acquisition module 40 acquires the new category of the label to be updated in the set of the label and the second new category.
- Label update module 50 Use the acquired label as the label of the new category of the label to be updated in the set of the second new category.
- the set judgment module 30 of the second new category further includes: if the first new category belongs to the set of the second new category, then the largest membership degree corresponds to the new category in the set of the second new category. The number of samples is increased by 1, and the first object is added to the sample set of the new category in the set of the second new category corresponding to the largest degree of membership; if the first new category does not belong to the second A new category set, after adding the first new category to the second new category set, the number of samples of the first new category in the second new category set is set to 1, and all The first object is added to the sample set of the first new category in the set of the second new category.
- Existing recognition system acquiring module 60 acquiring the first recognition device or system that recognizes the first object in the recognition module;
- New recognition system generation module 70 Obtain each new category whose number of samples is greater than the preset number of samples from the set of the second new category; obtain samples in the set of samples of each new category, and compare each The new category is deleted from the set of the second new category, and each new category is added to the set of the first known category to obtain the set of the second known category. Through the sample and each new category The category label trains the first recognition device or system, and obtains the recognition device or system capable of recognizing each new category as the second recognition device or system.
- the second identification device or system capable of identifying the new category can be used as the first identification device or system in the identification module 10, and the method in any of the above embodiments can be used to obtain more identification.
- the second recognition device or system of the new category enables more and more new categories to be recognized by the second recognition device or system.
- the first recognition device or system includes a deep learning model; an input variable of the deep learning model is an object; an output variable of the deep learning model is each category in the set of the first known categories and belongs to the The degree of membership of each category.
- the new recognition system generation module 70 specifically includes: acquiring each new category whose number of samples is greater than a preset number of samples from the set of the second new category; acquiring samples in the set of samples of each new category, and dividing the Each new category is deleted from the second new category set, and each new category is added to the first known category set to obtain the second known category set, and the second known category
- Each category in the set and the degree of membership belonging to each category are used as the output variables of the deep learning model.
- the sample set of each new category is compared with the label of each new category.
- the deep learning model is trained to obtain a deep learning model capable of recognizing each new category as the second recognition device or system.
- the step of training the deep learning model through the samples in the sample set of each new category and the label of each new category specifically includes: combining the samples in the sample set of each new category As the input data of the deep learning model, the membership degree corresponding to each new category in the expected output data of the deep learning model is set to 1, and the membership degrees corresponding to all other existing categories are set to 0.
- the described deep learning model performs supervised training.
- the device further includes:
- the object category judgment module 80 Determine whether the category of the object to be processed exists: No, use the method in any of the foregoing embodiments to identify the category of the object to be processed.
- Object information fusion module 90 Determine whether the category of the object to be processed belongs to the set of known categories: if yes, obtain the known category corresponding to the category of the object to be processed from the set of known categories ⁇ ; No, obtain the portrait of the category of the object to be processed from the set of the second new category, and obtain each known category and belong to the category from the portrait of the category of the object to be processed
- the membership degree of each known category is obtained, the information of each known category is acquired, and the information of each known category is fused according to the membership degree of each known category to obtain the Information about the category of the object to be processed.
- the device further includes:
- the second object acquiring module A0 acquiring the second object to be identified
- the second object recognition module B0 input the second object into the second recognition device or system for recognition, and obtain a second result;
- the second object category judgment module C0 judge whether the category of the second object belongs to the set of second known categories according to the second result: if no, use the category of the second object as the second new category; yes , The category of the first object is a known category. If the category of the first object is a known category, it is determined whether the category of the second object belongs to the set of the first known category: No, then the category of the second object is important to the first recognition device or system. It is the first new category; if yes, the category of the second object is a known category to the first recognition device or system.
- Embodiment 23 provides a system including the device described in any one of Embodiments 12-22.
- Embodiment 24 provides a robot system that includes the device described in any one of Embodiments 12-22.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
新类别识别方法和基于模糊理论和深度学习的机器人系统，包括：对第一对象进行识别得到第一结果；根据所述第一结果判断所述第一对象的类别是否属于第一已知类别的集合：否，则将所述第一对象的类别作为第一新类别；是，则所述第一对象的类别为已知类别。上述方法和机器人系统，通过根据所述第一结果来判断所述第一对象的类别，若所述第一对象不属于第一已知类别的集合，则得到所述第一对象所属的第一新类别，从而可以实现对新类别的识别。
Description
本发明涉及信息技术领域，特别是涉及一种新类别识别方法和基于模糊理论和深度学习的机器人系统。
已有对象的识别技术可以通过已知类别的特征来实现对对象的识别，例如深度学习对对象进行识别的办法。
在实现本发明过程中，发明人发现现有技术中至少存在如下问题：1、现有识别技术只能识别已知类别，无法识别新类别；2、一旦新类别出现，现有识别系统就会失效。
因此，现有技术还有待于改进和发展。
发明内容
基于此，有必要针对现有技术的缺陷或不足，提供新类别识别方法和基于模糊理论和深度学习的机器人系统，以解决现有技术无法对新类别进行识别的缺点。
第一方面，本发明实施例提供一种识别方法，所述方法包括：
识别步骤：对第一对象进行识别得到第一结果；
新类别判断步骤：根据所述第一结果判断所述第一对象的类别是否属于第一已知类别的集合：否，则将所述第一对象的类别作为第一新类别；是，则所述第一对象的类别为已知类别。
优选地，
所述第一结果包括所述第一对象所属的至少一个已知类别以及所述第一对象属于所述至少一个已知类别中每一个已知类别的隶属度；
所述新类别判断步骤具体包括：取所述第一对象属于所述至少一个已知类别中每一个已知类别的隶属度中的最大隶属度；若所述最大隶属度小于或等于预设第一隶属度，则所述第一对象的类别为第一新类别；若所述最大隶属度大于预设第一隶属度，则所述第一对象的已知类别为所述最大隶属度对应的已知类别。
优选地，所述方法还包括：
新类别集合判断步骤：根据所述第一结果判断所述第一新类别是否属于第二新类别的集合：是，则将所述第一新类别对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合。
优选地，所述新类别集合判断步骤包括：
新类别集合隶属度计算步骤:计算所述第一新类别属于所述第二新类别的集合中每一个新类别的隶属度；
新类别集合加入判断步骤:获取所述隶属度中最大的隶属度，判断所述最大的隶属度是否大于预设第二隶属度：是，则将所述最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合。
优选地，
所述新类别集合加入判断步骤还包括:
若所述最大的隶属度大于预设第二隶属度，则判断所述第一对象是否有类别标签：是， 则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述最大隶属度对应的所述第二新类别的集合中的所述新类别的标签；否，则获取所述最大隶属度对应的所述第二新类别的集合中的所述新类别的标签作为所述第一新类别的标签；
若所述最大的隶属度不大于预设第二隶属度，则判断所述第一对象是否有类别标签：是，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述第二新类别的集合中所述第一新类别的标签；否，则生成所述第一新类别的标签，将所述第一新类别的标签作为所述第二新类别的集合中所述第一新类别的标签。
优选地，
所述新类别集合隶属度计算步骤还包括：
将所述第一结果作为所述第一新类别的第一画像；计算所述第一新类别的所述第一画像与所述第二新类别的集合中每一个新类别的第二画像的相似度；将所述相似度作为所述第一新类别属于所述第二新类别的集合中所述每一个新类别的隶属度；
所述新类别集合加入判断步骤还包括:
若所述最大的隶属度大于预设第二隶属度，则将所述第一结果作为所述第一新类别的第一画像，融合所述第一新类别的第一画像与所述最大隶属度对应的所述第二新类别的集合中的新类别的第二画像得到第三画像；将所述第三画像作为所述最大的隶属度对应的所述第二新类别的集合中的所述新类别的第二画像；
若所述最大的隶属度不大于预设第二隶属度，则将所述第一新类别的第一画像作为所述第二新类别的集合中的所述第一新类别的第二画像。
优选地，
所述新类别集合判断步骤还包括：
若所述第一新类别属于第二新类别的集合，则将所述最大的隶属度对应的所述第二新类别的集合中的新类别的样本数加1，将所述第一对象加入所述最大的隶属度对应的所述第二新类别的集合中的新类别的样本集合中；
若所述第一新类别不属于所述第二新类别的集合，则在将所述第一新类别加入所述第二新类别的集合之后，将所述第二新类别的集合中所述第一新类别的样本数设置为1，将所述第一对象加入所述第二新类别的集合中所述第一新类别的样本集合；
所述方法还包括：
已有识别系统获取步骤：获取所述识别步骤中对所述第一对象进行识别的第一识别装置或系统；
新识别系统生成步骤：从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，通过所述样本和所述每一个新类别的标签对所述第一识别装置或系统进行训练，得到能够识别所述每一个新类别的识别装置或系统，作为第二识别装置或系统。
优选地，
所述第一识别装置或系统包括深度学习模型；所述深度学习模型的输入变量为对象；所述深度学习模型的输出变量为所述第一已知类别的集合中每一个类别及属于所述每一个类别的隶属度；
所述新识别系统生成步骤具体包括：
从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，将所述第二已知类别的集合中每一个类别及属于所述每一个类别的隶属度作为所述深度学习模型的输出变量，通过所述每一个新类别的样本集合中的样本和所述每一个新类别的标签对所述深度学习模型进行训练，得到能够识别所述每一个新类别的深度学习模型，作为第二识别装置或系统。
第二方面，本发明实施例提供一种识别装置，所述装置执行如第一方面的实施例中任一项所述的识别方法。
第三方面，本发明实施例提供一种机器人系统，所述机器人系统含有如第二方面的实施例中任一项所述的装置。
本发明实施例的有益效果：
1、本发明实施例中的方法和系统能识别新类别。已有的识别技术是通过已有类别的特征或样本进行机器学习或深度学习等方法来识别已有类别，如果输入属于新类别的对象就无法识别。而本发明实施例中的方法和系统可以将属于新类别的对象自动定义为一种新的类别，并当将来有属于同类新类别的对象时可以自动识别为之前所定义的新类别，从而可以实现对新类别的识别。
2、本发明实施例中的方法和系统对新类别的识别能力可以使得机器人不仅仅是根据之前机器学习或深度学习的结果进行识别，不仅仅能识别人类通过程序或数据所指定的已知类别，而且能够将识别能力延伸到属于新类别的对象，从而使得机器人能够进行自主识别，从而为机器人拥有自主意识打下基础。
3、本发明实施例中的方法和系统通过模糊理论中的隶属度在新类别与已有类别之间建立关联，计算新类别属于已有类别的隶属度以及新类别属于已识别出来的新类别的隶属度，从而通过已有类别来识别新类别，达到了与人类类似的联想识别的能力。
本实施例提供的新类别识别方法和基于模糊理论和深度学习的机器人系统，包括：对第一对象进行识别得到第一结果；根据所述第一结果判断所述第一对象的类别是否属于第一已知类别的集合：否，则将所述第一对象的类别作为第一新类别；是，则所述第一对象的类别为已知类别。上述方法和机器人系统，通过根据所述第一结果来判断所述第一对象的类别，若所述第一对象不属于第一已知类别的集合，则得到所述第一对象所属的第一新类别，从而可以实现对新类别的识别。
图1为本发明的实施例1、2提供的识别方法的流程图；
图2为本发明的实施例3提供的识别方法的流程图；
图3为本发明的实施例4、5、6提供的识别方法的流程图；
图4为本发明的实施例7提供的识别方法的流程图；
图5为本发明的实施例8、9提供的识别方法的流程图；
图6为本发明的实施例10提供的识别方法的流程图；
图7为本发明的实施例11提供的识别方法的流程图；
图8为本发明的实施例12、13提供的识别系统的原理框图；
图9为本发明的实施例14提供的识别系统的原理框图；
图10为本发明的实施例15、16、17提供的识别系统的原理框图；
图11为本发明的实施例18提供的识别系统的原理框图；
图12为本发明的实施例19、20提供的识别系统的原理框图；
图13为本发明的实施例21提供的识别系统的原理框图；
图14为本发明的实施例22提供的识别系统的原理框图。
下面结合本发明实施方式，对本发明实施例中的技术方案进行详细地描述。
实施例1：
如图1所示，一种识别方法，包括：
识别步骤S10：通过第一识别装置或系统对第一对象进行识别得到第一结果。对第一对象进行识别的方法包括采用机器学习或深度学习或其他识别方法。第一识别装置或系统包括机器学习系统、深度学习系统、等中的一种或几种。所述对象包括人、物、场景、事件等等，对象的格式包括图像、视频、文本等等，识别时是从图像或视频或文本等中识别出其中的人或物或场景或事件等等的类别。在此步骤之前还需要获取待识别的第一对象。
新类别判断步骤S20：根据所述第一结果判断所述第一对象的类别是否属于第一已知类别的集合：否，则将所述第一对象的类别作为第一新类别；是，则所述第一对象的类别为已知类别。所述第一已知类别的集合是所述第一识别装置或系统能够识别的类别集合。第一识别装置或系统在进行构建或训练时就已经确定了所述第一识别装置或系统能够识别的类别集合。
实施例2：
本实施例增加了通过隶属度来区分已知类别与新类别的一些步骤，是实施例1的一种更具体的实现方案
如图1所示，根据实施例1提供的识别方法，还包括：
识别步骤S10中所述的第一结果包括第一对象所属的至少一个已知类别以及所述第一对象属于所述所属的至少一个已知类别中每一个已知类别的隶属度。(所述隶属度也可以为可能性或概率)。
新类别判断步骤S20包括：取所述第一对象属于所述所属的至少一个已知类别中每一个已知类别的隶属度中的最大隶属度，判断所述最大隶属度是否大于预设第一隶属度(即判断所述第一对象的类别是否属于第一已知类别的集合)，若所述最大隶属度大于预设第一隶属度(即所述第一对象的类别属于第一已知类别的集合)，则所述第一对象的已知类别为所述最大隶属度对应的已知类别；若所述最大隶属度小于或等于预设第一隶属度(即所述第一对象的类别不属于第一已知类别的集合)，则所述第一对象的类别为第一新类别。
例如，所述第一对象为一个人的照片，第一结果为该照片中的人属于已有类别张三的隶属度(或概率或可能性)为10％、属于已有类别李四的隶属度为20％、属于已有类别王五的隶属度为70％。此时最大的隶属度是70％。预设第一隶属度为69％。所以最大隶属度大于预设第一隶属度。最大隶属度对应的类别是王五，所以所述第一对象的类别为已有类别王五。又例如，所述第一对象为一个人的照片，第一结果为该照片中的人属于已有类别张三的隶属度(或称概率或称可能性)为30％、属于已有类别李四的隶属度为40％、属于已有类别王五的隶属度为30％。此时最大的隶属度是40％。预设第一隶属度为69％。所以最大隶属度小于预设第一隶属度。因此所述第一对象的类别为第一新类别。
实施例3：
本实施例增加了通过隶属度来区分是否为第二新类别的集合中的新类别的一些步骤，这是对实施例1的步骤的补充，因为实施例只区分了已有类别和新类别，而没有区分新类别是否是过去已经识别到的新类别，也就是已在第二新类别的集合中的新类别。此实施例也是通过隶属度来区分所述第一新类别是否属于第二新类别的集合，实施例3只有在所述第一对象的类别不为已知类别时才会执行，在所述第一对象的类别为已知类别时则不会执行。
如图2所示，根据实施例1提供的识别方法，若所述第一对象的类别不为已知类别，则还包括：
类别集合判断步骤S30：根据所述第一结果判断所述第一新类别是否属于第二新类别的集合：是，则将所述第一新类别对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合。在此步骤前还要获取第二新类别的集合，并判断所述第二新类别的集合是否存在，若不存在，则新建一个空的第二新类别的集合，显然所述第一新类别不属于第二新类别的集合。
实施例4：
本实施例增加了通过隶属度来判断是否属于第二新类别的集合的一些步骤，是实施例3的一种更具体的实现方案。
如图3所示，根据实施例3提供的识别方法，类别集合判断步骤S30包括：
新类别集合隶属度计算步骤S31:计算所述第一新类别属于第二新类别的集合中每一个新类别的隶属度。在此步骤前还要获取第二新类别的集合，并判断所述第二新类别的集合是否存在，若不存在，则新建一个空的第二新类别的集合，显然所述第一新类别属于第二新类别的集合中每一个新类别的隶属度为0。
新类别集合加入判断步骤S32:获取所述隶属度中最大的隶属度，判断最大的隶属度是否大于预设第二隶属度(显然隶属度越大，则所述第一新类别属于第二新类别的集合中新类别的可能性就越大)：是，则将最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合(因为所述第一新类别不属于所述第二新类别的集合，所以需要加入)。
实施例5：
本实施例增加了对新类别的标签的生成的一些步骤。
如图3所示，根据实施例4提供的识别方法，
新类别集合加入判断步骤S32还包括:若最大的隶属度大于预设第二隶属度，则在将最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别之后，判断所述第一对象是否有类别标签，若是(将有标签的属于新类别的对象输入第一识别装置或系统可以使得第二新类别的集合中的新类别具有真实的标签)，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述最大隶属度对应的所述第二新类别的集合中的新类别的标签，若否，则获取所述最大隶属度对应的所述第二新类别的集合中的新类别的标签作为所述第一新类别的标签；若最大的隶属度不大于预设第二隶属度，则在将所述第一新类别加入所述第二新类别的集合之后，判断所述第一对象是否有类别标签，若是(将有标签的属于新类别的对象输入第一识别装置或系统可以使得第二新类别的集合中的新类别具有真实的标签)，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述第二新类别的集合中所述第一新类别的标签，若否，则生成所述第一新类别的标签，所述第一新类别的标签需要区别于已有类别的标签，也要区别于第二新类别的集合中任一新类别 的标签(因为每一个类别的标签需要唯一，否则难以进行各类别的区别，生成所述第一新类别的标签的具体步骤包括L：随机生成一个编码，并判断该编码与已有类别的标签、第二新类别的集合中任一新类别的标签是否相同：是，则回到L处重新执行；否，则将所述编码作为所述第一新类别的标签。例如，随机生成一个标签“新人三”，比较生成的标签与已有类别的标签“张三”、“李四”、“王二”、第二新类别的集合中每一个新类别的标签“新人一”、“新人二”是否相同，通过比较发现“新人三”可以作为所述第一新类别的标签。需要说明的是：初始时第二新类别的集合为空，所以初始时每一个新类别的标签也不存在)，将所述第一新类别的标签作为第二新类别的集合中所述第一新类别的标签。第二新类别的集合中新类别的表示方法包括用一个自动生成的唯一标识符来表示，而唯一标识符可以通过现有技术随机生成，当然第二新类别的集合中新类别的表示方法也包括直接用标签来表示第二新类别的集合中的新类别。
实施例6：
本实施例增加了对新类别的第一画像的生成的一些步骤。
如图3所示，根据实施例4提供的识别方法，
新类别集合隶属度计算步骤S31还包括：将所述第一结果作为所述第一新类别的第一画像。(画像指的是大数据和人工智能技术里的画像技术中的画像，画像也可以称之为特征)；获取第二新类别的集合中每一个新类别的画像作为第二新类别的集合中所述每一个新类别的第二画像，计算所述第一新类别的第一画像与第二新类别的集合中每一个新类别的第二画像的相似度，将所述相似度作为所述第一新类别属于第二新类别的集合中每一个新类别的隶属度。
当第一画像中有某个类别而第二新类别的集合中某个新类别的第二画像中没有该个类别时，则在计算相似度时将第二新类别的集合中所述某个新类别的第二画像中属于该个类别的隶属度视为0，或者第二新类别的集合中某个新类别的第二画像中有某个类别而第一画像中没有该个类别时，则在计算相似度时将所述第一画像中属于该个类别的隶属度视为0，这样便于计算相似度。其中，相似度的计算方法包括采用余弦相似度进行计算的方法；余弦相似度进行计算的方法包括公式1：
其中Xi指的是第一画像中属于第i类的隶属度，Yi指的是第二画像中属于第i类的隶属度，如果第一画像中不存在第i类，则Xi为0，如果第二画像中不存在第i类，则Yi为0。
新类别集合加入判断步骤S32还包括:若最大的隶属度不大于预设第二隶属度，则在将所述第一新类别加入所述第二新类别的集合之后(所述画像、标签是与所述类别关联的，可以通过数据库或知识库或大数据等技术进行这种关联)，将所述第一新类别的第一画像作为所述第二新类别的集合中的所述第一新类别的第二画像；若最大的隶属度大于预设第二隶属度，则在将最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别之后，将所述第一结果作为所述第一新类别的第一画像，融合所述第一新类别的第一画像与所述最大隶 属度对应的所述第二新类别的集合中的新类别的第二画像得到第三画像(即将第一画像和第二画像进行融合得到第三画像。因为第二新类别的集合是由新类别的加入而形成的，所以第二画像与第一画像的构成相似，第一画像和第二画像中包括至少一个类别以及属于所述至少一个类别中每个类别的隶属度；其中，在进行融合时，将第一画像和第二画像中的类别归并后作为第三画像中的类别；对第一画像和第二画像中属于每一类别的隶属度进行加权平均得到第三画像中属于所述每一类别的隶属度)，将所述第三画像作为所述最大隶属度对应的所述第二新类别的集合中的所述新类别的第二画像(进行第二新类别的集合中画像的更新，使得新类别的画像能更能代表该新类别)(第一画像、第二画像、第三画像都是画像，只是为了看起来更清晰，所以加上第一、第二、第三)。所述加权平均的方法包括((第一画像中属于每一类别的隶属度)×K1+(第二画像中属于所述每一类别的隶属度)×K2)/(K1+K2)。其中，K1、K2一般都取1，或者K1取1，K2取第二画像对应所述第二新类别的集合中的新类别的已有样本数。如果K2取第二画像对应所述第二新类别的集合中的新类别的已有样本数，则需要在S30中记录所述第二新类别的集合中的新类别的已有样本数，则S30需要扩展为：根据所述第一结果判断所述第一新类别是否属于第二新类别的集合：是，则将最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别，将最大隶属度对应的所述第二新类别的集合中的新类别的样本数加1；否，则在将所述第一新类别加入所述第二新类别的集合之后，将所述第二新类别的集合中所述第一新类别的样本数设置为1。
新类别集合加入判断步骤S32中所述将第一画像和第二画像进行融合得到第三画像的步骤中，如果第一画像中有某个类别、而第二画像中没有该个类别，则在计算第三画像中属于该个类别的隶属度时将第一画像中属于该个类别的隶属度视为0；或者第二画像中有某个类别、而第一画像中没有该个类别，则在计算第三画像中属于该个类别的隶属度时将第一画像中属于该个类别的隶属度视为0，这样处理是为了便于进行加权平均，因为第一画像和第二画像中的类别在大部分情况下是相同的，但也有可能不同。
例如，第一画像中属于已有类别张三的隶属度为30％、属于已有类别李四的隶属度为40％、属于已有类别王五的隶属度为30％；第二画像中属于已有类别张三的隶属度为32％、属于已有类别李四的隶属度为38％、属于已有类别王五的隶属度为30％；如果计算第一画像与第二画像的相似度，则可以通过包括余弦相似度在内的计算方法进行计算；如果将第一画像与第二画像融合为第三画像，则第三画像中的类别为张三、李四、王五，第三画像中属于已有类别张三的隶属度为30％加32％除以2即31％、属于已有类别李四的隶属度为40％加38％除以2即39％、属于已有类别王五的隶属度为30％加30％除以2即30％。
又例如，第一画像中属于已有类别张三的隶属度为30％、属于已有类别李四的隶属度为30％、属于已有类别王五的隶属度为30％、属于已有类别朱六的隶属度为10％；第二画像中属于已有类别张三的隶属度为32％、属于已有类别李四的隶属度为38％、属于已有类别王五的隶属度为30％；如果计算第一画像与第二画像的相似度，则可以通过包括余弦相似度在内的计算方法进行计算；如果将第一画像与第二画像融合为第三画像，则第三画像中的类别为张三、李四、王五、朱六，第三画像中属于已有类别张三的隶属度为30％加32％除以2即31％、属于已有类别李四的隶属度为30％加38％除以2即34％、属于已有类别王五的隶属度为30％加30％除以2即30％、属于已有类别朱六的隶属度为10％加0％除以2即5％)。
实施例7：
本实施例增加了对新类别的标签的更新的一些步骤，可以对第二新类别的集合里的标签 进行更新，而第二新类别的集合里的标签是在实施例3中产生的，在实施例5中已对标签进行了生成，但此时生成的标签还不是人类可以理解的标签。
如图4所示，根据实施例3提供的识别方法，
标签获取步骤S40：获取标签和所述第二新类别的集合中的待更新标签的新类别。
标签更新步骤S50:将获取的所述标签作为所述第二新类别的集合中的待更新标签的新类别的标签。
所述第一标签可以是用户赋予的，也可以通过其他方式得到的，例如通过互联网获得的。通过实施例3已经生成了第二新类别的集合，使得新对象也可以被识别为第二新类别的集合中的新类别，通过实施例5生成了新类别的标签，但实施例5中生成的标签与我们人类观念中的标签还是不一样的，例如New1、New2等。虽然新对象刚出现时用户可能不知道这个新对象所属类别的标签，但随着时间的推移，用户是有可能知道新类别的真实标签，例如知道New2其实是郑七，那此时就可以把New2对应的新类别的标签改为郑七，那么以后如果又有对象属于该新类别，就知道该新对象为郑七了。
另一种标签的更新方法(在实施例5中提供)是将有标签的属于新类别的对象输入第一识别装置或系统，也可以使得第二新类别的集合中的新类别具有真实的标签。
实施例8：
本实施增加了将新类别转换为识别装置或系统能够识别的已知类别的一些步骤。
如图5所示，根据实施例3提供的识别方法，
类别集合判断步骤S30还包括：若所述第一新类别属于第二新类别的集合，则在将最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别之后，将最大隶属度对应的所述第二新类别的集合中的新类别的样本数加1，将所述第一对象加入最大隶属度对应的所述第二新类别的集合中的新类别的样本集合中；若所述第一新类别不属于第二新类别的集合，则在将所述第一新类别加入所述第二新类别的集合之后，将所述第二新类别的集合中所述第一新类别的样本数设置为1，新建所述第二新类别的集合中所述第一新类别的样本集合(新建时为空)，将所述第一对象加入所述第二新类别的集合中所述第一新类别的样本集合。
已有识别系统获取步骤S60：获取所述识别步骤中对所述第一对象进行识别的第一识别装置或系统；
新识别系统生成步骤S70:从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，通过所述样本和所述每一个新类别的标签对所述第一识别装置或系统进行训练，得到能够识别所述每一个新类别的识别装置或系统，作为第二识别装置或系统。
然后又可以将所述能够识别所述新类别的第二识别装置或系统作为识别步骤S10中的所述第一识别装置或系统，重新通过上述任一实施例中的方法，得到能够识别更多新类别的第二识别装置或系统，使得越来越多的新类别能够被所述第二识别装置或系统所识别。
到新识别系统生成步骤S70这一步就完全将一个新类别变成了已知类别，从而使得第二识别装置或系统能自主地进化识别新类别，不仅仅是第一识别装置或系统能够识别的那些已知类别。
下次再进行所述新类别的对象的识别时，如果该对象属于所述新类别，则将该对象输入所述能够识别所述新类别的第二识别装置或系统，通过第二识别装置或系统的输出就能够直 接识别出该对象的类别为所述新类别，因为新类别已经加入到了第二识别装置或系统的输出变量中，成了已知类别。
实施例9：
本实施例增加了将新类别转换为深度学习识别系统能够识别的已知类别的一些步骤。
如图5所示，根据实施例8提供的识别方法，
所述第一识别装置或系统包括深度学习模型；所述深度学习模型的输入变量为对象；所述深度学习模型的输出变量为所述第一已知类别的集合中每一个类别及属于所述每一个类别的隶属度。
新识别系统生成步骤70具体包括:从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，将所述第二已知类别的集合中每一个类别及属于所述每一个类别的隶属度作为所述深度学习模型的输出变量(因为在第一已知类别的集合中增加了所述每一新类别，得到了第二已知类别的集合，而所述深度学习模型的输出为第一已知类别的集合中每一个类别及属于所述每一个类别的隶属度，所以需要对所述深度学习模型的输出变量进行更新)，通过所述每一个新类别的样本集合中的样本和所述每一个新类别的标签对所述深度学习模型进行训练，得到能够识别所述每一个新类别的深度学习模型，作为第二识别装置或系统。
其中，通过所述每一个新类别的样本集合中的样本和所述每一个新类别的标签对所述深度学习模型进行训练的步骤具体包括：将所述每一个新类别的样本集合中的样本作为所述深度学习模型的输入数据，将所述深度学习模型的预期输出数据中所述每一个新类别对应的隶属度设置为1、其他所有类别对应的隶属度设置为0,对所述深度学习模型进行有监督训练。当然在进行有监督训练之前还可以用所述第二新类别的集合中所有新类别的样本对所述深度学习模型进行无监督训练。
下次再进行所述新类别的对象的识别时，如果该对象属于所述新类别，则将该对象输入所述能够识别所述新类别的深度学习模型，通过该深度学习模型的输出就能够直接识别出该对象的类别为所述新类别，因为新类别已经加入到了该深度学习模型的输出变量中，成了已知类别。
实施例10：
本实施例增加了通过画像进行基于新类别画像的新对象应用的一些步骤。
如图6所示，根据实施例6提供的识别方法，所述方法还包括：
对象类别判断步骤S80：判断待处理的对象的类别是否存在：否，则使用上述任一实施例中的方法进行识别得到所述待处理的对象的类别。在此步骤之前还包括获取待处理的对象和所述处理对应的应用。所述处理对应的应用包括决策、分析、统计等等应用，可以用于工业领域、军事领域等等。
对象信息融合步骤S90：判断待处理的对象的类别是否属于所述已知类别的集合：是，则从所述已知类别的集合中获取与所述待处理的对象的类别对应的已知类别的信息；否，则从所述第二新类别的集合中获取所述待处理的对象的类别的画像，从所述待处理的对象的类别的画像中获取每一个已知类别及属于所述每一个已知类别的隶属度，获取所述每一个已知类别的信息，根据所述属于所述每一个已知类别的隶属度对所述每一个已知类别的信息进行融合，得到所述待处理的对象的类别的信息。所述已知类别的信息包括已知类别的知识。所 述融合的具体步骤包括获取预设的函数或模块，将所述每一个已知类别的信息及属于所述每一个已知类别的隶属度作为所述预设的函数或模块的输入，计算得到所述待处理的对象的类别的信息，将所述待处理的对象的类别的信息返回给所述处理对应的应用。所述融合的另一种简单的实现步骤包括对所述每一个已知类别的信息进行加权平均，得到所述待处理的对象的类别的信息，在加权平均时将所述属于所述每一个已知类别的隶属度作为所述每一个已知类别的信息的权值。
实施例11：
本实施例增加了识别新类别的一些步骤。
如图7所示，根据实施例8提供的识别方法，所述方法还包括：
第二对象获取步骤SA0：获取待识别的第二对象；
第二对象识别步骤SB0：将所述第二对象输入所述第二识别装置或系统进行识别，得到第二结果；
第二对象类别判断步骤SC0：根据所述第二结果判断所述第二对象的类别是否属于第二已知类别的集合：否，则将所述第二对象的类别作为第二新类别；是，则所述第一对象的类别为已知类别。若所述第一对象的类别为已知类别时，则判断所述第二对象的类别是否属于第一已知类别的集合：否，则所述第二对象的类别对第一识别装置或系统而言为第一新类别；是，则所述第二对象的类别对第一识别装置或系统而言为已知类别。
此后的针对所述第二对象的步骤与针对所述第一对象的步骤类似，请参见实施例1至10，在此不再赘述。
实施例12～20提供一种识别装置，分别与实施例1～11对应，其中细节，在此不再赘述，只列出其中与实施例1～11中各主要步骤对应的主要模块。
实施例12：
如图8所示，一种识别装置，包括：
识别模块10：通过第一识别装置或系统对第一对象进行识别得到第一结果。
新类别判断模块20：根据所述第一结果判断所述第一对象的类别是否属于第一已知类别的集合：否，则将所述第一对象的类别作为第一新类别；是，则所述第一对象的类别为已知类别。
实施例13：
如图8所示，根据实施例12提供的识别装置，还包括：
识别模块10中所述第一结果包括所述第一对象所属的至少一个已知类别以及所述第一对象属于所述至少一个已知类别中每一个已知类别的隶属度。
新类别判断模块20具体包括：取所述第一对象属于所述至少一个已知类别中每一个已知类别的隶属度中的最大隶属度，判断所述最大隶属度是否大于预设第一隶属度，若所述最大隶属度大于预设第一隶属度，则所述第一对象的已知类别为所述最大隶属度对应的已知类别；若所述最大隶属度小于或等于预设第一隶属度，则所述第一对象的类别为第一新类别。
实施例14：
如图9所示，根据实施例12提供的识别装置，若所述第一对象的类别不为已知类别，则还包括：
第二新类别的集合判断模块30：根据所述第一结果判断所述第一新类别是否属于第二新类别的集合：是，则将所述第一新类别对应的所述第二新类别的集合中的新类别作为所述第 一新类别；否，则将所述第一新类别加入所述第二新类别的集合。
实施例15：
如图10所示，根据实施例14提供的识别装置，
第二新类别的集合判断模块30包括：
新类别集合隶属度计算模块31:计算所述第一新类别属于所述第二新类别的集合中每一个新类别的隶属度。
新类别集合加入判断模块32:获取所述隶属度中最大的隶属度，判断所述最大的隶属度是否大于预设第二隶属度：是，则将所述最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合。
实施例16：
如图10所示，根据实施例15提供的识别装置，
新类别集合加入判断模块32还包括:若所述最大的隶属度大于预设第二隶属度，则判断所述第一对象是否有类别标签，若是，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述最大隶属度对应的所述第二新类别的集合中的所述新类别的标签，若否，则获取所述最大隶属度对应的所述第二新类别的集合中的所述新类别的标签作为所述第一新类别的标签；若所述最大的隶属度不大于预设第二隶属度，则判断所述第一对象是否有类别标签，若是，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述第二新类别的集合中所述第一新类别的标签，若否，则生成所述第一新类别的标签，将所述第一新类别的标签作为所述第二新类别的集合中所述第一新类别的标签。
实施例17：
如图10所示，根据实施例15提供的识别装置，
新类别集合隶属度计算模块31还包括：将所述第一结果作为所述第一新类别的第一画像。计算所述第一新类别的所述第一画像与所述第二新类别的集合中每一个新类别的第二画像的相似度，将所述相似度作为所述第一新类别属于所述第二新类别的集合中所述每一个新类别的隶属度。
其中，所述相似度的计算方法包括采用余弦相似度进行计算的方法；所述余弦相似度进行计算的方法包括公式1：
其中Xi指的是所述第一画像中属于第i类的隶属度，Yi指的是所述第二画像中属于第i类的隶属度。
新类别集合加入判断模块32还包括:若所述最大的隶属度不大于预设第二隶属度，则将所述第一新类别的第一画像作为所述第二新类别的集合中的所述第一新类别的第二画像；若所述最大的隶属度大于预设第二隶属度，则将所述第一结果作为所述第一新类别的第一画像，融合所述第一新类别的第一画像与所述最大隶属度对应的所述第二新类别的集合中的新类别的第二画像得到第三画像；其中，在进行融合时，将所述第一画像和所述第二画像中的类别 归并后作为所述第三画像中的类别；对所述第一画像和所述第二画像中属于每一类别的隶属度进行加权平均得到所述第三画像中属于所述每一类别的隶属度；将所述第三画像作为所述最大的隶属度对应的所述第二新类别的集合中的所述新类别的第二画像。
实施例18：
如图11所示，根据实施例14提供的识别装置，
标签获取模块40：获取标签和所述第二新类别的集合中的待更新标签的新类别。
标签更新模块50:将获取的所述标签作为所述第二新类别的集合中的待更新标签的所述新类别的标签。
实施例19：
如图12所示，根据实施例14提供的识别装置，
第二新类别的集合判断模块30还包括：若所述第一新类别属于第二新类别的集合，则将所述最大的隶属度对应的所述第二新类别的集合中的新类别的样本数加1，将所述第一对象加入所述最大的隶属度对应的所述第二新类别的集合中的新类别的样本集合中；若所述第一新类别不属于所述第二新类别的集合，则在将所述第一新类别加入所述第二新类别的集合之后，将所述第二新类别的集合中所述第一新类别的样本数设置为1，将所述第一对象加入所述第二新类别的集合中所述第一新类别的样本集合。
已有识别系统获取模块60：获取所述识别模块中对所述第一对象进行识别的第一识别装置或系统；
新识别系统生成模块70：从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，通过所述样本和所述每一个新类别的标签对所述第一识别装置或系统进行训练，得到能够识别所述每一个新类别的识别装置或系统，作为第二识别装置或系统。
然后又可以将所述能够识别所述新类别的第二识别装置或系统作为识别模块10中的所述第一识别装置或系统，重新通过上述任一实施例中的方法，得到能够识别更多新类别的第二识别装置或系统，使得越来越多的新类别能够被所述第二识别装置或系统所识别。
实施例20：
如图12所示，根据实施例19提供的识别装置，
所述第一识别装置或系统包括深度学习模型；所述深度学习模型的输入变量为对象；所述深度学习模型的输出变量为所述第一已知类别的集合中每一个类别及属于所述每一个类别的隶属度。
新识别系统生成模块70具体包括:从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，将所述第二已知类别的集合中每一个类别及属于所述每一个类别的隶属度作为所述深度学习模型的输出变量，通过所述每一个新类别的样本集合中的样本和所述每一个新类别的标签对所述深度学习模型进行训练，得到能够识别所述每一个新类别的深度学习模型，作为第二识别装置或系统。
其中，通过所述每一个新类别的样本集合中的样本和所述每一个新类别的标签对所述深度学习模型进行训练的步骤具体包括：将所述每一个新类别的样本集合中的样本作为所述深 度学习模型的输入数据，将所述深度学习模型的预期输出数据中所述每一个新类别对应的隶属度设置为1、其他所有已有类别对应的隶属度设置为0,对所述深度学习模型进行有监督训练。
实施例21：
如图13所示，根据实施例17提供的识别装置，所述装置还包括：
对象类别判断模块80：判断待处理的对象的类别是否存在：否，则使用上述任一实施例中的方法进行识别得到所述待处理的对象的类别。
对象信息融合模块90：判断待处理的对象的类别是否属于所述已知类别的集合：是，则从所述已知类别的集合中获取与所述待处理的对象的类别对应的已知类别的信息；否，则从所述第二新类别的集合中获取所述待处理的对象的类别的画像，从所述待处理的对象的类别的画像中获取每一个已知类别及属于所述每一个已知类别的隶属度，获取所述每一个已知类别的信息，根据所述属于所述每一个已知类别的隶属度对所述每一个已知类别的信息进行融合，得到所述待处理的对象的类别的信息。
实施例22：
如图14所示，根据实施例19提供的识别装置，所述装置还包括：
第二对象获取模块A0：获取待识别的第二对象；
第二对象识别模块B0：将所述第二对象输入所述第二识别装置或系统进行识别，得到第二结果；
第二对象类别判断模块C0：根据所述第二结果判断所述第二对象的类别是否属于第二已知类别的集合：否，则将所述第二对象的类别作为第二新类别；是，则所述第一对象的类别为已知类别。若所述第一对象的类别为已知类别时，则判断所述第二对象的类别是否属于第一已知类别的集合：否，则所述第二对象的类别对第一识别装置或系统而言为第一新类别；是，则所述第二对象的类别对第一识别装置或系统而言为已知类别。
此后的针对所述第二对象的步骤与针对所述第一对象的步骤类似，请参见实施例12至21，在此不再赘述。
实施例23提供一种系统，所述系统含有如实施例12～22任一实施例所述的装置。
实施例24提供一种机器人系统，所述机器人系统含有如实施例12～22任一实施例所述的装置。
以上所述实施例仅表达了本发明的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对本发明专利范围的限制。应当指出的是，则对于本领域的普通技术人员来说，在不脱离本发明构思的前提下，还可以做出若干变形和改进，这些都属于本发明的保护范围。因此，本发明专利的保护范围应以所附权利要求为准。
Claims (10)
- 一种识别方法，其特征在于，所述方法包括：识别步骤：对第一对象进行识别得到第一结果；新类别判断步骤：根据所述第一结果判断所述第一对象的类别是否属于第一已知类别的集合：否，则将所述第一对象的类别作为第一新类别；是，则所述第一对象的类别为已知类别。
- 根据权利要求1所述的识别方法，其特征在于，所述第一结果包括所述第一对象所属的至少一个已知类别以及所述第一对象属于所述至少一个已知类别中每一个已知类别的隶属度；所述新类别判断步骤具体包括：取所述第一对象属于所述至少一个已知类别中每一个已知类别的隶属度中的最大隶属度；若所述最大隶属度小于或等于预设第一隶属度，则所述第一对象的类别为第一新类别；若所述最大隶属度大于预设第一隶属度，则所述第一对象的已知类别为所述最大隶属度对应的已知类别。
- 根据权利要求1所述的识别方法，其特征在于，所述方法还包括：新类别集合判断步骤：根据所述第一结果判断所述第一新类别是否属于第二新类别的集合：是，则将所述第一新类别对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合。
- 根据权利要求3所述的识别方法，其特征在于，所述新类别集合判断步骤包括：新类别集合隶属度计算步骤:计算所述第一新类别属于所述第二新类别的集合中每一个新类别的隶属度；新类别集合加入判断步骤:获取所述隶属度中最大的隶属度，判断所述最大的隶属度是否大于预设第二隶属度：是，则将所述最大隶属度对应的所述第二新类别的集合中的新类别作为所述第一新类别；否，则将所述第一新类别加入所述第二新类别的集合。
- 根据权利要求4所述的识别方法，其特征在于，所述新类别集合加入判断步骤还包括:若所述最大的隶属度大于预设第二隶属度，则判断所述第一对象是否有类别标签：是，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述最大隶属度对应的所述第二新类别的集合中的所述新类别的标签；否，则获取所述最大隶属度对应的所述第二新类别的集合中的所述新类别的标签作为所述第一新类别的标签；若所述最大的隶属度不大于预设第二隶属度，则判断所述第一对象是否有类别标签：是，则将所述类别标签作为所述第一新类别的标签，并将所述第一新类别的标签作为所述第二新 类别的集合中所述第一新类别的标签；否，则生成所述第一新类别的标签，将所述第一新类别的标签作为所述第二新类别的集合中所述第一新类别的标签。
- 根据权利要求4所述的识别方法，其特征在于，所述新类别集合隶属度计算步骤还包括：将所述第一结果作为所述第一新类别的第一画像；计算所述第一新类别的所述第一画像与所述第二新类别的集合中每一个新类别的第二画像的相似度；将所述相似度作为所述第一新类别属于所述第二新类别的集合中所述每一个新类别的隶属度；所述新类别集合加入判断步骤还包括:若所述最大的隶属度大于预设第二隶属度，则将所述第一结果作为所述第一新类别的第一画像，融合所述第一新类别的第一画像与所述最大隶属度对应的所述第二新类别的集合中的新类别的第二画像得到第三画像；将所述第三画像作为所述最大的隶属度对应的所述第二新类别的集合中的所述新类别的第二画像；若所述最大的隶属度不大于预设第二隶属度，则将所述第一新类别的第一画像作为所述第二新类别的集合中的所述第一新类别的第二画像。
- 根据权利要求3所述的识别方法，其特征在于，所述新类别集合判断步骤还包括：若所述第一新类别属于第二新类别的集合，则将所述最大的隶属度对应的所述第二新类别的集合中的新类别的样本数加1，将所述第一对象加入所述最大的隶属度对应的所述第二新类别的集合中的新类别的样本集合中；若所述第一新类别不属于所述第二新类别的集合，则在将所述第一新类别加入所述第二新类别的集合之后，将所述第二新类别的集合中所述第一新类别的样本数设置为1，将所述第一对象加入所述第二新类别的集合中所述第一新类别的样本集合；所述方法还包括：已有识别系统获取步骤：获取所述识别步骤中对所述第一对象进行识别的第一识别装置或系统；新识别系统生成步骤：从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，通过所述样本和所述每一个新类别的标签对所述第一识别装置或系统进行训练，得到能够识别所述每一个新类别的识别装置或系统，作为第二识别装置或系统。
- 根据权利要求7所述的识别方法，其特征在于，所述第一识别装置或系统包括深度学习模型；所述深度学习模型的输入变量为对象；所述深度学习模型的输出变量为所述第一已知类别的集合中每一个类别及属于所述每一个类别的隶属度；所述新识别系统生成步骤具体包括：从所述第二新类别的集合中获取样本数大于预设样本数的每一个新类别；获取所述每一个新类别的样本集合中的样本，将所述每一个新类别从第二新类别的集合中删除，并将所述每一个新类别加入所述第一已知类别的集合，得到第二已知类别的集合，将所述第二已知类别的集合中每一个类别及属于所述每一个类别的隶属度作为所述深度学习模型的输出变量，通过所述每一个新类别的样本集合中的样本和所述每一个新类别的标签对所述深度学习模型进行训练，得到能够识别所述每一个新类别的深度学习模型，作为第二识别装置或系统。
- 一种识别装置，其特征在于，所述装置执行如权利要求1至8任一项所述的识别方法。
- 一种机器人系统，其特征在于，所述机器人系统含有如权利要求9所述的装置。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910224601.3A CN109934302B (zh) | 2019-03-23 | 2019-03-23 | 新类别识别方法和基于模糊理论和深度学习的机器人系统 |
| CN201910224601.3 | 2019-03-23 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020191988A1 true WO2020191988A1 (zh) | 2020-10-01 |
Family
ID=66988040
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2019/099963 WO2020191988A1 (zh) | 2019-03-23 | 2019-08-09 | 新类别识别方法和基于模糊理论和深度学习的机器人系统 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109934302B (zh) |
| WO (1) | WO2020191988A1 (zh) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109934302B (zh) * | 2019-03-23 | 2021-12-07 | 大国创新智能科技(东莞)有限公司 | 新类别识别方法和基于模糊理论和深度学习的机器人系统 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778804A (zh) * | 2016-11-18 | 2017-05-31 | 天津大学 | 基于类别属性迁移学习的零样本图像分类方法 |
| US20180268298A1 (en) * | 2017-03-15 | 2018-09-20 | Salesforce.Com, Inc. | Deep Neural Network-Based Decision Network |
| CN109376786A (zh) * | 2018-10-31 | 2019-02-22 | 中国科学院深圳先进技术研究院 | 一种图像分类方法、装置、终端设备及可读存储介质 |
| CN109460471A (zh) * | 2018-11-01 | 2019-03-12 | 信融源大数据科技(北京)有限公司 | 一种基于自学习的方式建立纤维种类图谱库的方法 |
| CN109934302A (zh) * | 2019-03-23 | 2019-06-25 | 大国创新智能科技(东莞)有限公司 | 新类别识别方法和基于模糊理论和深度学习的机器人系统 |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050187001A1 (en) * | 2004-02-25 | 2005-08-25 | Fishel Alan G. | Geography based card game and method of play |
| CN100382411C (zh) * | 2006-03-10 | 2008-04-16 | 天津大学 | 基于概率的电力系统电压稳定在线监视和预防控制方法 |
| US8661148B2 (en) * | 2009-12-17 | 2014-02-25 | American Express Travel Related Services Company, Inc. | System and method for enabling industry based channels in an IP marketplace |
| CN106021377A (zh) * | 2016-05-11 | 2016-10-12 | 上海点荣金融信息服务有限责任公司 | 计算机实现的信息处理方法及装置 |
| CN107067045A (zh) * | 2017-05-31 | 2017-08-18 | 北京京东尚科信息技术有限公司 | 数据聚类方法、装置、计算机可读介质和电子设备 |
| CN107194617B (zh) * | 2017-07-06 | 2020-11-24 | 北京航空航天大学 | 一种app软件工程师软技能分类系统及方法 |
| CN107688822B (zh) * | 2017-07-18 | 2021-07-20 | 中国科学院计算技术研究所 | 基于深度学习的新增类别识别方法 |
-
2019
- 2019-03-23 CN CN201910224601.3A patent/CN109934302B/zh active Active
- 2019-08-09 WO PCT/CN2019/099963 patent/WO2020191988A1/zh active Application Filing
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778804A (zh) * | 2016-11-18 | 2017-05-31 | 天津大学 | 基于类别属性迁移学习的零样本图像分类方法 |
| US20180268298A1 (en) * | 2017-03-15 | 2018-09-20 | Salesforce.Com, Inc. | Deep Neural Network-Based Decision Network |
| CN109376786A (zh) * | 2018-10-31 | 2019-02-22 | 中国科学院深圳先进技术研究院 | 一种图像分类方法、装置、终端设备及可读存储介质 |
| CN109460471A (zh) * | 2018-11-01 | 2019-03-12 | 信融源大数据科技(北京)有限公司 | 一种基于自学习的方式建立纤维种类图谱库的方法 |
| CN109934302A (zh) * | 2019-03-23 | 2019-06-25 | 大国创新智能科技(东莞)有限公司 | 新类别识别方法和基于模糊理论和深度学习的机器人系统 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109934302B (zh) | 2021-12-07 |
| CN109934302A (zh) | 2019-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110021051B (zh) | 一种基于生成对抗网络通过文本指导的人物图像生成方法 | |
| WO2020038136A1 (zh) | 面部识别方法、装置、电子设备及计算机可读介质 | |
| Yuan et al. | Discriminative video pattern search for efficient action detection | |
| CN108288051B (zh) | 行人再识别模型训练方法及装置、电子设备和存储介质 | |
| CN104915351B (zh) | 图片排序方法及终端 | |
| CN111079646A (zh) | 基于深度学习的弱监督视频时序动作定位的方法及系统 | |
| CN110503076B (zh) | 基于人工智能的视频分类方法、装置、设备和介质 | |
| CN109344285A (zh) | 一种面向监控的视频图谱构建和挖掘方法、设备 | |
| CN113569627B (zh) | 人体姿态预测模型训练方法、人体姿态预测方法及装置 | |
| CN111680550B (zh) | 情感信息识别方法、装置、存储介质及计算机设备 | |
| CN105975932A (zh) | 基于时间序列shapelet的步态识别分类方法 | |
| Rabiee et al. | Crowd behavior representation: an attribute-based approach | |
| Cai et al. | Probabilistic attribute tree structured convolutional neural networks for facial expression recognition in the wild | |
| Yang et al. | Sampling agnostic feature representation for long-term person re-identification | |
| Venkat et al. | Recognizing occluded faces by exploiting psychophysically inspired similarity maps | |
| CN113128526B (zh) | 图像识别方法、装置、电子设备和计算机可读存储介质 | |
| WO2020191988A1 (zh) | 新类别识别方法和基于模糊理论和深度学习的机器人系统 | |
| CN114358202A (zh) | 基于药物分子图像分类的信息推送方法及装置 | |
| Batool et al. | Fundamental recognition of ADL assessments using machine learning engineering | |
| WO2023093241A1 (zh) | 行人重识别方法及装置、存储介质 | |
| CN113627380B (zh) | 一种用于智能安防及预警的跨视域行人重识别方法及系统 | |
| CN109460485A (zh) | 一种图像库建立方法、装置和存储介质 | |
| Joshi et al. | Real-time object detection and identification for visually challenged people using mobile platform | |
| Kadar et al. | Scenenet: A perceptual ontology for scene understanding | |
| Zhao et al. | Affective video classification based on spatio-temporal feature fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19922032 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19922032 Country of ref document: EP Kind code of ref document: A1 |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19922032 Country of ref document: EP Kind code of ref document: A1 |
|
| 32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 24/03/2022) |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19922032 Country of ref document: EP Kind code of ref document: A1 |