Incremental classification learning method and system for target objects in dynamic service scenarios

Abstract

The application discloses a method and system for incremental classification learning of target objects in a dynamic service scene, and new class data sets and old class memory network sets at T time are input into a classification network, the new class data sets are used for training the classification network to realize incremental learning, and the old class memory network sets are used for a replay process of the classification network at T time to prevent catastrophic forgetting; through training, parameters of the classification network are updated to obtain a classification network capable of identifying new class and old class objects; during training, a prototype selection strategy based on density peaks is used to select a prototype for each new class in the new class data sets at T time to obtain a prototype set corresponding to each new class; based on the new class data sets and the prototype set corresponding to each new class, a self-organizing incremental learning neural network is constructed for each new class to obtain a new class memory network set; and the new class memory network set and the old class memory network set are merged to obtain an updated memory network set at T time.

Description

Technical Field

[0001] This invention relates to the field of continuous learning technology, and in particular to an incremental classification learning method and system for target objects in dynamic service scenarios. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] Currently, many countries face the challenge of an aging population. The number of elderly people requiring care is also increasing. However, labor shortages and rising labor costs make it difficult for traditional elderly care services to meet the enormous demand. With the rapid development of artificial intelligence and robotics, elderly care robots are expected to become one of the most effective technologies for improving elderly care services.

[0004] Elderly care robots interact with people through sensor perception, image recognition, and voice recognition, providing accurate and high-quality services and alleviating the shortage of human caregivers. In open, everyday life scenarios, elderly people exhibit significant individual differences, service tasks are complex and diverse, and service scenarios are dynamically changing. Therefore, elderly care robots need to learn from continuous data streams online and should possess the ability to continuously learn, accumulate new knowledge, and retain what they have learned.

[0005] Existing elderly care robots need to collect and identify objects in the service scene in real time during the care process. However, existing technologies have problems such as low accuracy in recognizing dynamically changing objects in the environment due to offline training models, lack of continuous learning ability, and a decrease in the accuracy of recognizing known knowledge after learning new objects, which may lead to catastrophic forgetting. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an incremental classification learning method and system for target objects in dynamic service scenarios;

[0007] On the one hand, it provides an incremental classification learning method for target objects in dynamic service scenarios;

[0008] Incremental classification learning methods for target objects in dynamic service scenarios include:

[0009] Obtain the new class dataset at time T, which includes: images and labels of the new class target objects; obtain the complete set of old class memory networks that have been pre-learned before time T;

[0010] The new class dataset and the old class memory network set at time T are both input into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used to replay the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can recognize both new and old class objects.

[0011] During training, a prototype selection strategy based on density peaks selects a prototype for each new class in the new class dataset at time T, resulting in a prototype set for each new class.

[0012] During training, a self-organizing incremental learning neural network is constructed for each new category based on the new category dataset and the prototype set corresponding to each new category, thereby obtaining a new category memory network set. The new category memory network set is merged with the old category memory network set to obtain the memory network set updated at time T. The memory network set updated at time T is used for the replay process at time T+1.

[0013] The system acquires the image to be identified in the dynamic service scenario, inputs the image into the updated classification network, and outputs the recognition result of the target object.

[0014] On the other hand, it provides an incremental classification learning system for target objects in dynamic service scenarios;

[0015] An incremental classification learning system for target objects in dynamic service scenarios includes:

[0016] The acquisition module is configured to: acquire the new class dataset at time T, which includes images and labels of new class target objects; and acquire the complete set of old class memory networks that have been pre-learned before time T.

[0017] The incremental learning module is configured to input both the new class dataset and the old class memory network set at time T into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used for the replay process of the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can simultaneously recognize new and old class objects.

[0018] The prototype selection module is configured to use a prototype selection strategy based on density peaks to select a prototype for each new class in the new class dataset at time T, thereby obtaining a prototype set corresponding to each new class.

[0019] The memory network construction module is configured to: construct a self-organizing incremental learning neural network for each new class based on the new class dataset and the prototype set corresponding to each new class, thereby obtaining a new class memory network set; merge the new class memory network set with the old class memory network set to obtain the memory network set updated at time T, which will be used for the replay process at time T+1;

[0020] The output module is configured to: acquire the image to be recognized in the dynamic service scenario, input the image to be recognized into the updated classification network, and output the recognition result of the target object.

[0021] Furthermore, an electronic device is also provided, including:

[0022] Memory, used for non-transitory storage of computer-readable instructions; and

[0023] Processor, for executing the computer-readable instructions,

[0024] When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0025] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the instructions of the method described in the first aspect are executed.

[0026] In another aspect, a computer program product is also provided, including a computer program that, when run on one or more processors, is used to implement the method described in the first aspect above.

[0027] The above technical solution has the following advantages or beneficial effects:

[0028] This invention proposes a robot continuous learning network based on an improved incremental classifier and representation learning (iCaRL) method. The network uses replay and regularization strategies to train a ResNet to learn progressively increasing classes. This invention improves the replay method in two main ways. First, a prototype selection strategy based on density peaks is proposed, enabling the network to obtain accurate prototypes under data distributions of arbitrary shapes. Second, a self-organizing incremental neural network is introduced as a sample memory network for replay during ResNet training for each class. This not only learns a general representation for each class to obtain a sparse memory network, avoiding excessive storage consumption, but also, due to its dynamic architecture, it can incrementally learn real-time data streams and avoid catastrophic forgetting. Experimental results show that the proposed method achieves better learning performance and efficiency.

[0029] This invention proposes an improved incremental classifier and representation learning method to help elderly care robots perform incremental learning in complex and ever-changing human life scenarios. It can recognize labels of multiple target objects in human life scenarios and also handle newly added object labels in these scenarios. A memory network is used to enable the network to gradually learn new class representations without forgetting old classes. On one hand, a prototype selection strategy based on density peaks is designed to replace the mean-based method. This strategy allows for the selection of more diverse samples as prototypes, which is helpful for representing data distributions of arbitrary shapes. On the other hand, a self-organizing incremental neural network (SOINN) is introduced for each class as a memory network to replace the fixed sample set in iCaRL. This invention utilizes the generalization and dynamic architecture of SOINN to memorize an unlimited number of classes and learn generalized example representations for replay. Attached Figure Description

[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0031] Figure 1 This is a flowchart of the method in Example 1;

[0032] Figure 2(a) shows the CIFAR-100 accuracy results for each learning process in Example 1 with 10 classes as the class increment;

[0033] Figure 2(b) shows the CIFAR-100 accuracy results for each learning process in Example 1 with 20 classes as the class increment;

[0034] Figure 2(c) shows the CIFAR-100 accuracy results for each learning process in Example 1 with 50 classes as the class increment. Detailed Implementation

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] The main challenge of continuous learning is catastrophic forgetting. To address this, many compelling studies have been proposed. The most popular continuous learning strategies can be categorized into three types: replay methods, regularization methods, and parameter isolation methods. Replay methods typically employ a memory to store a small number of past samples or use a generative model to learn the data distribution of learned experiences, reusing this data when learning new categories to mitigate forgetting. This approach can achieve good learning results, but limited memory can limit the number of categories that can be learned. Regularization methods focus on preserving learned knowledge by modifying weight updates through knowledge distillation, such as LwF, EWC, and SI. However, changes in the distribution of learned knowledge can lead to significant errors. Parameter isolation methods fix the parameters of previous knowledge or use a dynamic architecture to extend the network for new categories. This approach allows new knowledge to be quickly added to the network, which is more in line with the nature of continuous learning. However, as knowledge accumulates, this method faces significant storage costs, and the learning rate gradually decreases.

[0037] Since each method has its advantages and disadvantages, hybrid learning methods are becoming a trend, combining these strategies to find the optimal solution. Sprechmann et al. proposed a memory-based parameter adaptation method that employs both replay and parameter isolation. This method stores learned samples in memory and uses their replays to modify the weights of the neural network. He et al. proposed a concept-assisted backprojection method to prevent gradients from influencing previous learning tasks; this method incorporates regularization and dynamic architecture. Incremental Classifier and Representation Learning (iCaRL) is a continuous learning method based on replay and regularization strategies, suitable for incremental class scenarios. It uses a fixed prototype set to store learned classes for replay, and then trains a representation network simultaneously using classification loss and distillation loss. However, the selection of samples is based on the average of all samples in each class, which may not be suitable for non-spherical data distributions in dynamic data streams. Furthermore, as the number of classes learned increases, the number of samples per class decreases due to the fixed sample set. This may limit the number of classes it can learn.

[0038] Example 1

[0039] This embodiment provides an incremental classification learning method for target objects in dynamic service scenarios;

[0040] like Figure 1 As shown, the incremental classification learning method for target objects in dynamic service scenarios includes:

[0041] S101: Obtain the new class dataset at time T, which includes images and labels of new class target objects; obtain the set of all old class memory networks that have been pre-learned before time T;

[0042] S102: Input both the new class dataset and the old class memory network set at time T into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used to replay the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can recognize both new and old class objects.

[0043] S102-1: During training, a prototype selection strategy based on density peaks is used to select a prototype for each new class in the new class dataset at time T, resulting in a prototype set corresponding to each new class.

[0044] S102-2: During training, based on the new class dataset and the prototype set corresponding to each new class, a self-organizing incremental learning neural network is constructed for each new class, thereby obtaining a new class memory network set; the new class memory network set is merged with the old class memory network set to obtain the memory network set updated at time T, which will be used for the replay process at time (T+1);

[0045] S103: Obtain the image to be recognized in the dynamic service scenario, input the image to be recognized into the updated classification network, and output the recognition result of the target object.

[0046] It should be understood that iCaRL is a continuous learning method based on replay and regularization. Because the data stream appears as class increments during continuous learning, iCaRL can learn multiple classes simultaneously. This method defines a prototype set P of size K for replay and a feature extractor. Used to represent learning.

[0047] Assume the data stream in continuous learning is X 1 ,…,X s-1 Class y∈N contains n y Sample iCaRL allocates a capacity of size m = K / 's to each class in the prototype set P. For class y, the average feature vector μ is as shown in formula (1). The k-th sample p in this class is closest to μ. k As shown in formula (2), m samples are selected from each class as the prototype set p of that class. y ={p1,p2,…,p m}

[0048]

[0049]

[0050] Then, new class X is added gradually using regularization strategies. s ,…,Xt All samples D = {X} in the new class and the learned prototype set. s ,…,X t}∪P together use a single loss function to train the feature extractor The loss function L(θ) includes the classification loss L for the new class of data. c (θ) and distillation loss L for learned categories of data d (θ), as shown in formulas (3)-(5).

[0051] L(θ)=L c (θ)+L d (θ), (3)

[0052]

[0053]

[0054] Among them, g y (·) represents the sigmoid output of the feature extractor. It is the output of the current feature extractor on the sample set.

[0055] In the next learning process, iCaRL has learned t classes, and the number of existing prototypes for each class is updated to m = K / t. Excess prototypes are removed from the prototype set. Finally, iCaRL uses a classification strategy based on the closest sample mean for classification.

[0056] This invention proposes an improved iCaRL method for elderly care robots, enabling the robot to learn new class samples without forgetting knowledge of already learned classes. In this method, the entire continuous learning process is as follows: Figure 1 As shown, it can be described in three steps: First, a prototype for each new class is selected using a density peak-based method; second, a self-organizing neural network is used for each class to learn a sparse memory network, which can also be regarded as a memory for replay; third, the parameters of ResNet are updated using samples of the new class and including the prototypes of the old classes, so that the network can gradually learn the new class representation without forgetting the old classes. Algorithm 1 illustrates the class incremental learning process.

[0057] Table 1. Category Incremental Learning Process

[0058]

[0059]

[0060] During continuous learning, knowledge representation is learned by a ResNet-based feature extraction network, whose output node number varies with the number of learned classes. When a new class appears, the network parameters are updated by minimizing a loss function (as shown in equations (3)-(5)), which includes the classification loss for the new class and the distillation loss for the medical class. In particular, the constraint of the distillation loss is crucial to prevent catastrophic forgetting in the network. This invention retains this part of the network in iCARL and focuses on improving the algorithm by constructing a more efficient replay dataset, specifically including the following two aspects.

[0061] First, this invention employs a prototype selection strategy based on density peaks instead of a mean-based approach. Distance-based average vectors best approximate class features with spherical data distributions but are unsuitable for non-spherical categories. Inspired by the CFS algorithm, this invention selects samples with density peaks as prototype centers for each class, thus handling data distributions of arbitrary shapes. Unlike traditional clustering tasks, this invention only extracts cluster centers, eliminating the need to cluster the remaining samples. Furthermore, multiple density peaks are selected as prototypes for each class, rather than a single average. Therefore, prototypes based on density peaks can more accurately represent the data distribution of dynamic data streams during continuous learning.

[0062] Second, this invention proposes a replay memory network based on an incremental self-organizing neural network to improve upon the fixed prototype set in iCaRL. Because a fixed sample set not only limits the number of classes to be learned but also gradually reduces the number of samples per class, this means that fewer and fewer prototypes are available for replaying each class. The training results become more biased towards new classes, leading to a degree of catastrophic forgetting. The incremental self-organizing neural network can overcome these shortcomings. On one hand, this network has a dynamic structure, allowing for unlimited learning of new classes; on the other hand, each node stores a generalized representation of several similar samples. Therefore, it can store an appropriate number of prototypes for each class for replaying.

[0063] Further, S101: Obtaining a new class dataset, the new class dataset including images and labels of target objects of the new class, specifically including:

[0064] In the scenario of elderly care robots accompanying the elderly, the new class dataset refers to the images and labels of all unknown categories of target objects in the care scenario, such as: medicine images, cabinet images, bed images, water cup images, window images, shoe images, clothes hanger images, and television images.

[0065] Further, S101: Obtaining a pre-learned set of old category memory networks, wherein the learning process includes:

[0066] Obtain the old dataset, which consists of images and labels of target objects of known categories;

[0067] For each category in the old dataset, a prototype selection strategy based on density peaks is used to select several samples as the prototype set for the current category.

[0068] Construct a memory network for the current category based on the prototype set for each category;

[0069] Based on the memory networks of all categories, construct memory networks for all categories to obtain the set of memory networks for the old categories.

[0070] Further, in step S102: the new class dataset at time T and the old class memory network set are both input into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used for the replay process of the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can simultaneously recognize new and old class objects. Specifically, this includes:

[0071] The new class dataset and the old class memory network set at time T are both input into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used to replay the classification network at time T to prevent catastrophic forgetting. When the total loss function value no longer decreases or the number of iterations exceeds the set number, training is stopped, and the trained classification network is obtained. The total loss function includes the classification loss function and the distillation loss function.

[0072] Further, S102-1: The prototype selection strategy based on density peaks selects a prototype for each new class in the new class dataset at time T, obtaining a prototype set corresponding to each new class, specifically including:

[0073] S102-11: Calculate the distance between sample i and sample j within a category; determine the local density ρ of sample i based on the distance between sample i and sample j and the cutoff distance. i The values ​​of i and j are both in the range of 1 to n; n is a positive integer.

[0074] S102-12: Calculate the maximum distance δ between sample i and other samples of the same class when the local density of sample i is maximized. i Based on the local density ρ of sample i i And the minimum distance δ between sample i and all other samples of the same category with higher density than sample i. i ;

[0075] S102-13: Determine the gamma value γ of sample i iThis leads to the set of gamma values ​​γ for all samples.

[0076] S102-14: Sort the γ set in descending order. Starting from the first sample in the sorted γ set, check each sample i, δ... i ≥2σ(δ i And ρ i ≥μ(ρ i The following condition is considered: ) whether the following condition holds true, where σ(·) represents the variance and μ(·) represents the mean; if true, then sample i is considered the cluster center of the current category, sample i is selected as the prototype of label y, and the label y of sample i is recorded in the set C = {C...} s ,…,C t In}, among which n represents category y c The process involves using a center to examine subsequent samples in the sorted γ set, thereby obtaining the prototype set corresponding to the new category; if the condition is not met, the process ends.

[0077] Further, in S102-11: calculate the distance between sample i and sample j within a category; based on the distance between sample i and sample j and the cutoff distance, determine the local density ρ of sample i. i Specifically, it includes:

[0078] For each sample in the current input sample set, first calculate the local density ρ. i Local density ρ i Defined as:

[0079]

[0080] Where, d ij It is the distance between sample i and sample j, d c It is the cutoff distance, which is defined as the value at the 2nd percentile position after sorting the distances between any two samples in ascending order.

[0081] Further, in S102-12: when the local density of sample i is maximized, the maximum distance δ between sample i and other samples of the same category is calculated. i Based on the local density ρ of sample i i And the minimum distance δ between sample i and all other samples of the same category with higher density than sample i. i Specifically, this includes:

[0082]

[0083] Where, δ i Let δ represent the distance between sample i and any other sample with a higher density. When sample i's own density is at its maximum, δ... iThis indicates the maximum distance between this sample and other samples.

[0084] Further, in S102-13: determine the median value γ of sample i. i Specifically, it includes:

[0085] Using γ i =ρ i δ i Calculate γ for each sample i value.

[0086] The prototype selection strategy based on density peaks primarily selects samples with high density and large distances within a category as prototypes for that category. Algorithm 2 details the prototype selection strategy based on density peaks.

[0087] Table 2 Prototype Selection Strategy Based on Density Peak

[0088]

[0089]

[0090] Further, S102-2: Based on the new class dataset and the prototype set corresponding to each new class, a self-organizing incremental learning neural network is constructed for each new class, thereby obtaining a new class memory network set, specifically including:

[0091] S102-21: Using the prototype set corresponding to the new category y, initialize the self-organizing incremental neural network of the current category y, and use each prototype as a node to initialize the network;

[0092] S102-22: Learn each sample of category y one by one, for each sample Determine if the number of nodes in the initial network is less than 2. If the number of nodes in the initial network is less than 2, then directly add the sample. Create a new node and establish connections between it and existing nodes.

[0093] S102-23: If the number of nodes in the initial network is greater than or equal to 2, then find the Euclidean distance sample from the initial network. The two nearest nodes are b1 and b2;

[0094] S102-24: If the sample Simultaneously meet the node activation conditions and Among them, w b1 w b2 TH represents the weights of nodes b1 and b2 respectively. b1 TH b2Let b1 and b2 represent the similarity thresholds, respectively. Treating it as a new sample in the current network, create a new node to store the new sample;

[0095] Otherwise, it is believed Similar to b1, and b2 can be considered a neighbor of b1; establish a topological connection between b1 and b2, and simultaneously use... To update the weights of node b1 and its neighboring node b2;

[0096]

[0097]

[0098] Where, ξ b1 w represents the update rate of node b1. Nb1 This represents the weights of the neighboring nodes of node b1.

[0099] S102-25: Nodes that are activated less than a set threshold are deleted as noise nodes to obtain a simplified memory network;

[0100] S102-26: After traversing all the new categories, we obtain the new category memory network set.

[0101] This invention assigns a self-organizing incremental learning neural network as a memory network to each class, enabling it to learn sparse and generalized sample representations. Unlike traditional self-organizing incremental neural networks (SOINN), this invention utilizes prototype samples. To initialize the network for category y, first create n c Nodes

[0102] Then, SOINN learns all samples of category y one by one according to the Hebbian learning rule. Each sample appears once in a sequential learning manner.

[0103] When samples in category y When this occurs, SOINN will find the two nearest nodes b1 and b2 in the network.

[0104] If the sample Able to simultaneously meet node activation conditions and but This can be considered a new sample in the current network, and a new node needs to be created to store this new sample.

[0105] Otherwise, it is believed Similar to b1, and b2 can be considered a neighbor of b1. Therefore, the network will establish a topological connection between b1 and b2, while using... This is used to update the weights of node b1 and its neighboring nodes, which helps to obtain a generalized prototype representation.

[0106] Furthermore, when the initial network has fewer than 2 nodes, the sample... After inputting the data, a new node should be created immediately to prevent the case where there is only one prototype for the category.

[0107] The network's online incremental learning capability can help select more representative samples from each class for replay, rather than just selecting samples that are close to the average and decrease over time.

[0108] Once all samples of a class have been learned, nodes that are rarely activated after being created in this SOINN will be considered noise and removed, thus simplifying the memory network.

[0109] After traversing all new classes, the entire memory network set M = {M} can be obtained. 1 ,…,M s M s+1 ,…,M t This is used to replay the process in the next learning cycle. Algorithm 3 illustrates the details of the memory network construction process.

[0110] Table 3 Details of the memory network construction process

[0111]

[0112]

[0113] Further, in S102-2: the new category memory network set is merged with the old category memory network set to obtain the memory network set updated at time T, which will be used for the replay process at time (T+1); specifically, the union of the new and old category memory network sets is taken to obtain the complete memory network set.

[0114] This invention evaluates the proposed method using experiments on the CIFAR-100 dataset. In scenarios with increasing number of classes, each class appears sequentially and is processed only once. Therefore, this invention divides the dataset into 10, 20, or 50 classes sequentially. To verify the effectiveness of the two improvements, this invention compares iCaRL. Proposed Method 1 represents iCaRL improved solely through a prototype selection strategy based on density peaks, while Proposed Method 2 represents the entire algorithm with two improvements. Experiments are conducted using the avalanche library in Python, and the evaluation metric is multi-class accuracy on a classification test set.

[0115] Table 4. Average accuracy on CIFAR-100

[0116]

[0117] Figures 2(a) to 2(c) The results in Table 1 demonstrate that the proposed method 2 achieves a higher average accuracy than iCaRL across all three batch types. Furthermore, Figures 2(a) to 2(c) This demonstrates that this excellent performance is consistent throughout the entire learning process. Therefore, these two improvements do indeed enhance the learning effectiveness of iCaRL.

[0118] The accuracy of the proposed method 1 is only slightly better than iCaRL. It seems that using a prototype selection strategy based solely on density peaks does not improve learning efficiency. However, as shown in Figure 2(a), the proposed method 1 outperforms iCaRL in the early stages of learning, but gradually falls behind as the number of classes increases. This indicates that the prototype selection strategy based on density peaks has a positive impact on learning. The fixed sample set in iCaRL limits ResNet's learning of a comprehensive distribution of class data. This is because the number of prototypes m = K / t per class decreases as the number of classes learned increases. In later stages of learning, only a small number of prototypes are selected for distillation loss to update the parameters of ResNet. Therefore, the extracted features are only associated with a few density peaks, but the final closest-to-mean classifier tends to correctly classify samples within the spherical neighborhood of the mean.

[0119] When used in conjunction with the improved SOINN-based memory network proposed in this invention, the prototype selection strategy based on density peaks significantly improves performance. This is because SOINN can learn sparse and generalized representations for each class, which is equivalent to updating ResNet with the entire data distribution for each class. Simultaneously, the prototype selection strategy based on density peaks helps SOINN store density peak samples, which facilitates learning data distributions of arbitrary shapes, not just the spherical distribution of recent nearest averaging methods.

[0120] For elderly care robots, this invention proposes an improved incremental classifier and representation learning method. This method utilizes a prototype selection strategy based on density peaks to select suitable cluster centers as prototypes for each class. Furthermore, each class is equipped with a memory network based on an improved self-organizing incremental neural network, initialized by the prototype. A series of comparative experiments on the CIFAR-100 dataset with iCaRL demonstrate that these improvements can enhance learning efficiency. Therefore, the proposed method is an effective continuous learning approach. Future work will consider applying this method to physical robots such as Nao to test its practical effectiveness.

[0121] Example 2

[0122] This embodiment provides an incremental classification learning system for target objects in dynamic service scenarios;

[0123] An incremental classification learning system for target objects in dynamic service scenarios includes:

[0124] The acquisition module is configured to: acquire the new class dataset at time T, which includes images and labels of new class target objects; and acquire the complete set of old class memory networks that have been pre-learned before time T.

[0125] The incremental learning module is configured to input both the new class dataset and the old class memory network set at time T into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used for the replay process of the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can simultaneously recognize new and old class objects.

[0126] The prototype selection module is configured to use a prototype selection strategy based on density peaks to select a prototype for each new class in the new class dataset at time T, thereby obtaining a prototype set corresponding to each new class.

[0127] The memory network construction module is configured to: construct a self-organizing incremental learning neural network for each new class based on the new class dataset and the prototype set corresponding to each new class, thereby obtaining a new class memory network set; merge the new class memory network set with the old class memory network set to obtain the memory network set updated at time T, which will be used for the replay process at time (T+1);

[0128] The output module is configured to: acquire the image to be recognized in the dynamic service scenario, input the image to be recognized into the updated classification network, and output the recognition result of the target object.

[0129] It should be noted that the acquisition module, incremental learning module, prototype selection module, memory network construction module, and output module described above correspond to steps S101 to S103 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.

[0130] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0131] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0132] Example 3

[0133] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0134] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0135] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0136] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0137] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0138] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0139] Example 4

[0140] This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An incremental classification learning method for target objects in dynamic service scenarios, characterized by: include: Obtain the new class dataset at time T, which includes: images and labels of the new class target objects; obtain the complete set of old class memory networks that have been pre-learned before time T; The new class dataset and the old class memory network set at time T are both input into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used to replay the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can recognize both new and old class objects. During training, a prototype selection strategy based on density peaks selects a prototype for each new class in the new class dataset at time T, resulting in a prototype set for each new class. Specifically, this includes: Calculate the sample under a category With sample Distance between; based on samples With sample The distance between and the cutoff distance determine the sample. Local density ; and The values ​​of all values ​​are in the range of 1 to n; n is a positive integer; sample When the local density is at its maximum, calculate the sample Maximum distance between the sample and other samples in the same category According to the sample Local density Calculate samples Compared with other density ratio samples in the same category Higher minimum distance between all samples ; Determine the sample gamma value ; and then the gamma values ​​of all samples are obtained. gather; sorted in descending order Sort the set, and then... Starting with the first sample in the set, examine each sample one by one. , Whether it is valid, among which Represents variance. This represents the mean; if true, then the sample... Select samples as cluster centers for the current category. As a label The prototype, recording samples tags Recorded in the set Among them Representative category of Each center, after sorting The subsequent samples in the set are checked to obtain the prototype set corresponding to the new category; if the condition is not met, the process ends. During training, a self-organizing incremental learning neural network is constructed for each new category based on the new category dataset and the prototype set corresponding to each new category, thereby obtaining a new category memory network set. The new category memory network set is merged with the old category memory network set to obtain the memory network set updated at time T. The memory network set updated at time T is used for the replay process at time T+1. The system acquires the image to be identified in the dynamic service scenario, inputs the image into the updated classification network, and outputs the recognition result of the target object.

2. The incremental classification learning method for target objects in a dynamic service scenario as described in claim 1, characterized in that, Calculate the sample under a category With sample Distance between; based on samples With sample The distance between and the cutoff distance determine the sample. Local density Specifically, it includes: For each sample in the current input sample set, first calculate the local density. Local density Defined as: ，（6） in, It is a sample With sample The distance between them It is the cutoff distance, which is defined as the value at the 2nd percentile position after sorting the distances between any two samples in ascending order.

3. The incremental classification learning method for target objects in a dynamic service scenario as described in claim 1, characterized in that, sample When the local density is at its maximum, calculate the sample Maximum distance between the sample and other samples in the same category According to the sample Local density Calculate samples Compared with other density ratio samples in the same category Higher minimum distance between all samples Specifically, this includes: （7） in, Indicates sample The distance between the sample and any other sample with a higher density, when the sample When its density is at its maximum, This indicates the maximum distance between this sample and other samples.

4. The incremental classification learning method for target objects in a dynamic service scenario as described in claim 1, characterized in that, Determine the sample median value Specifically, this includes: utilizing Calculate the value of each sample value.

5. The incremental classification learning method for target objects in a dynamic service scenario as described in claim 1, characterized in that, Based on the new class dataset and the prototype set corresponding to each new class, a self-organizing incremental learning neural network is constructed for each new class, thereby obtaining a set of new class memory networks, specifically including: Utilizing new categories The corresponding prototype set for the current category The self-organizing incremental neural network is initialized by using each prototype as a node in the initialized network. Learning categories one by one For each sample, for each sample Determine if the number of nodes in the initial network is less than 2. If the number of nodes in the initial network is less than 2, then directly add the sample. Create a new node and establish connections between it and existing nodes. If the number of nodes in the initial network is greater than or equal to 2, then find the Euclidean distance sample from the initial network. The two most recent nodes and ; If the sample Simultaneously meet the node activation conditions and ,in, , Representing nodes respectively and The weights, , Representing nodes respectively and The similarity threshold, then Treating it as a new sample in the current network, create a new node to store the new sample; Otherwise, it is believed and Similar, and It can be regarded as The neighboring nodes; in and Establish a topology connection between them, and use Update the node and its neighboring nodes The weights; Nodes that are activated less than a set threshold are removed as noise nodes, resulting in a simplified memory network. After traversing all the new categories, we obtain a set of new category memory networks.

6. The incremental classification learning method for target objects in a dynamic service scenario as described in claim 5, characterized in that, use Update the node and its neighboring nodes The weights; ，（8） ，（9） in, Represents a node update rate, Represents a node The weights of the neighboring nodes.

7. An incremental classification learning system for target objects in dynamic service scenarios, characterized by: include: The acquisition module is configured to: acquire the new class dataset at time T, which includes images and labels of new class target objects; and acquire the complete set of old class memory networks that have been pre-learned before time T. The incremental learning module is configured to input both the new class dataset and the old class memory network set at time T into the classification network. The new class dataset is used to train the classification network to achieve incremental learning, and the old class memory network is used for the replay process of the classification network at time T to prevent catastrophic forgetting. Through training, the parameters of the classification network are updated to obtain a classification network that can simultaneously recognize new and old class objects. The prototype selection module is configured to use a density peak-based prototype selection strategy to select a prototype for each new class in the new class dataset at time T, resulting in a prototype set for each new class. Specifically, this includes: Calculate the sample under a category With sample Distance between; based on samples With sample The distance between and the cutoff distance determine the sample. Local density ; and The values ​​of all values ​​are in the range of 1 to n; n is a positive integer; sample When the local density is at its maximum, calculate the sample Maximum distance between the sample and other samples in the same category According to the sample Local density Calculate samples Compared with other density ratio samples in the same category Higher minimum distance between all samples ; Determine the sample gamma value ; and then the gamma values ​​of all samples are obtained. gather; sorted in descending order Sort the set, and then... Starting with the first sample in the set, examine each sample one by one. , Whether it is valid, among which Represents variance. This represents the mean; if true, then the sample... Select samples as cluster centers for the current category. As a label The prototype, recording samples tags Recorded in the set Among them Representative category of Each center, after sorting The subsequent samples in the set are checked to obtain the prototype set corresponding to the new category; if the condition is not met, the process ends. The memory network construction module is configured to: construct a self-organizing incremental learning neural network for each new class based on the new class dataset and the prototype set corresponding to each new class, thereby obtaining a new class memory network set; merge the new class memory network set with the old class memory network set to obtain the memory network set updated at time T, which will be used for the replay process at time T+1; The output module is configured to: acquire the image to be recognized in the dynamic service scenario, input the image to be recognized into the updated classification network, and output the recognition result of the target object.

8. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-6.

9. A storage medium for non-transitory storage of computer-readable instructions, characterized in that, in, When non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-6 are executed.

Images