A small sample class incremental recognition method based on reinforcement learning

Abstract

The application provides a small sample class incremental identification method based on reinforcement learning, and the identification method comprises the following steps: downloading a small sample class incremental learning data set; designing a small sample incremental classification identification network based on reinforcement learning; training the network by using training data to obtain a trained network; generating a small sample class incremental learning model according to the trained network; and performing small sample incremental classification identification by using the small sample class incremental learning model. The problem that the existing technology cannot alleviate the small sample incremental learning disaster forgetting caused by the non-adaptive exploration of the new class classifier learning strategy is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of reinforcement learning, small sample classification processing and computer vision technology, and in particular to a small sample class incremental recognition method based on reinforcement learning. BACKGROUND

[0002] In some scenarios, such as industrial visual inspection, new concepts usually have only a few samples, such as new defect categories, which creates a very challenging scenario - a small sample learning scenario, at which time we hope that the artificial intelligence system can quickly integrate these new concepts while retaining old knowledge, which is called small sample incremental learning. The standard process of small sample class incremental learning is to input training data of different categories in stages, and at each stage, a group of specific class samples are input into the model, and these samples are usually very small in number, and data of other categories is also unavailable. Incremental learning gives the model the ability to iteratively and dynamically learn, which has natural advantages and places in intelligent tutoring systems and risk control fields. However, when the artificial intelligence model is updated with limited samples of new categories, the "catastrophic forgetting" phenomenon often occurs, that is, the knowledge obtained from previous data is rapidly degraded. In addition, due to insufficient training data, the model may also be severely overfit on new classes. Therefore, how to alleviate catastrophic forgetting and overfitting becomes a problem to be solved. Current small sample incremental recognition methods can be divided into "based on replay", "based on regularization" and "based on network architecture". Specifically, there are:

[0003] 1) Incremental learning method based on replay. Main technical means: alleviate the forgetting of old knowledge in each stage of incremental learning by implicitly retaining relevant old knowledge, and then use the corresponding classifier for incremental recognition. Problems and shortcomings: These methods require additional computing resources and storage space to recall old knowledge, and when the number of tasks increases, either the training cost will increase, or the representativeness of the representative samples will decrease, and in the actual production environment, this method may also have the problem of data privacy leakage.

[0004] 2) Incremental learning method based on regularization. The model output result adopts cross-entropy / feature-based and the intermediate result adopts L2 norm or cosine similarity, so that the new model is similar to the old model. Problems and shortcomings: This method is highly dependent on the correlation between new and old tasks, and when the task difference is too large, the task confusion phenomenon will occur, and the training time of a task will increase linearly with the number of learning tasks, and the introduced regularization term often cannot effectively constrain the optimization process of the model on the new task.

[0005] The incremental learning method based on network architecture. The main technical means: through expanding the model structure or using the topological structure for incremental learning, the network structure is constantly modified. In this way, the old task is maintained through the old model weight, and the newly added model structure adapts to the incremental task, so as to achieve the purpose of adapting to the old task and the new task.

[0006] The prior art has the technical problems and defects that the increase of the model structure inevitably aggravates the task storage burden and the operation burden, and how to expand the model in a more optimal way. SUMMARY

[0007] In view of the above problems, the present application is proposed in order to provide a small sample class incremental identification method based on reinforcement learning, which overcomes the above problems or at least partially solves the above problems.

[0008] According to one aspect of the present application, a small sample class incremental identification method based on reinforcement learning is provided, the identification method comprising:

[0009] Downloading a small sample class incremental learning data set;

[0010] Designing a small sample incremental classification identification network based on reinforcement learning;

[0011] Training the network using training data to obtain a trained network;

[0012] Generating a small sample class incremental learning model according to the trained network;

[0013] Using the small sample class incremental learning model for small sample incremental classification identification.

[0014] Optionally, the downloaded small sample class incremental learning data set specifically comprises:

[0015] Collecting a mainstream small sample incremental learning data set CIFAR-100, including 100 classes, each class containing 600 images;

[0016] Each class has 500 training images and 100 test images, and the 100 classes in the small sample incremental learning data set CIFAR-100 are divided into 20 superclasses;

[0017] Each image has a fine label and a coarse label;

[0018] In accordance with the small sample incremental N-way K-shot setting, the small sample incremental learning data set CIFAR-100 data set is divided into 60 base classes and 40 new classes;

[0019] After the base phase, the remaining 40 classes are further divided into 8 incremental phases, each of which is a 5-way 5-shot classification task.

[0020] Optionally, the design based on reinforcement learning small sample incremental classification recognition network specifically includes two parts of classifier update and exploratory optimization.

[0021] Optionally, the classifier update specifically includes:

[0022] The model input is the CIFAR-100 dataset, the encoder φ uses ResNet20 as the backbone, and W t is the weight of the new class classifier at stage t, and if M classes are included in each incremental phase, then W t =[w1,…,w k ,…,w M ] T where w k represents the classifier weight of the kth class, and d represents the feature dimension;

[0023] For an input sample x i from the batch data X, the cosine similarity is used to measure the extracted feature embedding and the corresponding classification weight W t to describe the classification performance, and is defined as the classifier state After obtaining the current batch data prediction state S t , the cross-entropy loss of the current classifier is calculated:

[0024]

[0025] where, is the kth element of the one-hot encoded label vector y i , and represents the kth element of the classifier state , which corresponds to the cosine distance of the kth class in the new class on the sample x i ; The greater, the more similar to the kth class;

[0026] The gradient t of the current classifier weight parameter W is obtained, and the classifier learning strategy a t obtained by the exploratory optimization module is used to update the weight parameter together:

[0027]

[0028] The next classifier state Through this exploration, the updated weight W was obtained. t+1 Using sample x i Calculations show that learning strategy a is obtained simultaneously. t Rewards Where R is the reward function.

[0029] According to claim 1, the few-shot class incremental recognition method based on reinforcement learning is characterized in that the exploratory optimization specifically includes:

[0030] Using the critic ψ on the classifier state S t Conduct an evaluation to obtain long-term returns. t The reward r updated one step with the classifier t different;

[0031] v t From state S t Begin evaluating the classifier's performance, and use a t Continuously update the classifier;

[0032] Next classifier state S t+1 The return value v is obtained through the exploratory critic ψ. t It describes the next state S. t+1 Strategy a is used for a long time t Then, the impact on classifier performance;

[0033] Using the idea of ​​single-step difference in reinforcement learning to guide the learning of the critic ψ:

[0034]

[0035] Here, γ is a scaling factor used to scale future returns;

[0036] Current differences and current classifier strategy a t The log value simultaneously adaptively supervises the Gaussian actors. Explore optimization:

[0037]

[0038] Loss L as a supervisory signaling actor a Guiding Gaussian Actors Updated to

[0039] Next classifier state S t+1 Through the updated actors The classifier policy is learned by taking the mean μ and variance σ, and a new classifier learning policy a is obtained by Gaussian sampling N(μ, σ). t+1 ;

[0040] New strategy a t+1 for pushing the classifier parameter W t+1 , alternately pushing the classifier update and the exploratory optimization of the classifier strategy, and realizing the integration of the update of the new class classifier and the exploratory optimization of the new class classifier into an end-to-end closed loop learning process.

[0041] The small sample class incremental identification method based on reinforcement learning according to claim 4, wherein the reward function R specifically comprises:

[0042] The reward r intra for evaluating the intra-class compactness in the new class:

[0043]

[0044] wherein represents the operation corresponding to the label y i ;

[0045] The reward r inter for evaluating the inter-class dispersion of the new class data:

[0046]

[0047] When the new classifier parameter W t+1 makes the input sample x i closer to other classes, the reward score r inter becomes lower, and the largest one is taken as the penalty;

[0048] The reward r cross for evaluating the dispersion between the new and old classes:

[0049]

[0050] wherein Y old represents the old class label at the current stage. The closer the feature embedding of the sample x i to the classifier of the old class, the smaller the reward r cross value, indicating the more misclassification cases;

[0051] r intra represents the gain of the correct classification probability of the updated classifier compared with the previous classifier, wherein the correct classification probability is proportional to the cosine measure;

[0052] r inter represents the change of the maximum misclassification probability before and after the update;

[0053] r cross represents the sum of the probability values of the new class samples being misclassified into the old classes, rcross The larger, the more knowledge of the old class can be retained;

[0054] Total reward r of reinforcement learning t (x i ) formalization: r t (x i ) = r intra (x i ) + r inter (x i ) + r cross (x i ).

[0055] Optionally, the training of the network using the training data specifically includes:

[0056] In the basic training stage, a large amount of base class data is used to train the basic model;

[0057] In the incremental stage, the backbone parameters of the model are frozen, and only the classifier parameters are optimized;

[0058] The model is trained for 2000 rounds using the Adam optimizer, and the learning rate is 0.0003;

[0059] During the entire training process, data augmentation is performed.

[0060] Optionally, the data augmentation method specifically includes: using random cropping, random scaling and random horizontal flipping method for data augmentation.

[0061] Optionally, the identification method further includes:

[0062] After each incremental stage, the trained model is evaluated on the corresponding test set, and the classification accuracy is reported.

[0063] The present application provides a kind of based on reinforcement learning small sample class incremental identification method, the identification method includes: download small sample class incremental learning dataset;Design small sample incremental classification identification network based on reinforcement learning;Training data is used to train the network, and training network is obtained;According to the training network, small sample class incremental learning model is generated;Small sample incremental classification identification is carried out using the small sample class incremental learning model.The problem that the above prior art cannot alleviate small sample incremental learning disaster forget problem caused by not self-adaptive exploration new class classifier learning strategy is solved.

[0064] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, which can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 A flowchart illustrating a few-shot class incremental recognition method based on reinforcement learning, provided in an embodiment of the present invention;

[0067] Figure 2 A schematic diagram of a small-sample incremental learning model based on reinforcement methods provided in an embodiment of the present invention. Detailed Implementation

[0068] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0069] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0070] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0071] like Figure 1 As shown, the data preparation stage

[0072] 1.1 We collected the mainstream few-shot incremental learning dataset CIFAR-100, which has 100 classes, each containing 600 images. Each class has 500 training images and 100 test images. The 100 classes in CIFAR-100 are divided into 20 superclasses. Each image has a "fine" label (its class) and a "coarse" label (its superclass). Following the few-shot incremental N-way K-shot setup, the CIFAR-100 dataset is divided into 60 base classes and 40 new classes. After the base class training in the base stage is completed, the remaining 40 classes are further divided into 8 incremental stages, each stage being a 5-way 5-shot classification task.

[0073] During the model design phase, the end-to-end model design is as follows:

[0074] 2.1 Overall model as follows Figure 2 As shown, it includes two parts: classifier update and exploratory optimization. In the classifier update part, the model input is the CIFAR-100 dataset, and the encoder φ uses ResNet20 as the backbone. Let W... t It represents the weights of the new class classifier at time t in a certain stage. If the new class in each incremental stage contains M categories, then... W t =[w1,…,w k ,…,w M ] T , where w k Let represent the classifier weights for the k-th class, and d represent the feature dimension. For an input sample x from the batch data X... i We use cosine similarity to measure the extracted feature embeddings. and the corresponding classification weight W t To describe classification performance, and define it as the classifier state. Obtain the current batch data prediction status S t Then, calculate the cross-entropy loss of the classifier at the current time step:

[0075]

[0076] in, The label vector y is a one-hot encoded vector. i The kth element, Indicates the classifier state The k-th element corresponds to the sample x in the k-th class of the new class. i The cosine distance. The larger the value, the more it resembles the k-th class. Then, the classifier weight parameters W at the current time step are obtained. t gradient And combined with the classifier learning strategy a obtained from the exploratory optimization module t Update the weight parameters together:

[0077]

[0078] Subsequently, the next classifier state Through this exploration, the updated weight W was obtained. t+1 Using sample x i Calculated. Simultaneously, learning strategy a is obtained. t Rewards On the other hand, in exploring the optimization module, the first step is to use the critic ψ to evaluate the classifier state S. t Conduct an evaluation to obtain long-term returns. t (Reward r updated one step with the classifier) t (Different). Then vt From state S t Begin evaluating the classifier's performance, and use a t The classifier is continuously updated. Similarly, the next classifier state S... t+1 The return value v is obtained through the exploratory critic ψ. t It describes the next state S. t+1 Strategy a is used for a long time t Then, the impact on classifier performance. Next, the idea of ​​single-step difference in reinforcement learning is used to guide the learning of the critic ψ:

[0079]

[0080] Here, γ is a scaling factor used to scale future returns. On the other hand, the current difference and the current classifier policy a... t The log value simultaneously adaptively supervises the Gaussian actors. Explore optimization:

[0081]

[0082] Next, as a monitoring signaling actor, the loss L a Guiding Gaussian Actors Updated to Accordingly, the next classifier state S t+1 Through the updated actors The classifier policy is learned by taking the mean μ and variance σ, and a new classifier learning policy a is obtained by Gaussian sampling N(μ, σ). t+1 New strategy a t+1 Used to drive classifier parameter W t+1 The next update will alternately advance the classifier update and the exploratory optimization of the classifier strategy, and integrate the update of the new classifier and the exploratory optimization of the new classifier into an end-to-end closed-loop learning process.

[0083] 2.2 Design of reward function R.

[0084] (1) Evaluate the reward r for intraclass tightness in the new class. intra :

[0085]

[0086] in, This indicates taking the label y. i The operation.

[0087] (2) Evaluate the reward r for the inter-class dispersion of new class data. inter :

[0088]

[0089] When the new classifier parameters W t+1 Make the input sample x i The closer to other classes, the higher the reward score r. inter The lower the value, the greater the change will be, and the penalty will be applied accordingly.

[0090] (3) Evaluate the distributed reward r between the old and new classes. cross :

[0091]

[0092] Among them, Y old This indicates the old category label for the current stage. Sample x i Feature embedding The closer the classifier is to the old class, the higher the reward r. cross The smaller the value, the more likely it is to be misclassified.

[0093] In summary, r intra This represents the gain in the correct classification probability of the updated classifier compared to the previous classifier, where the correct classification probability is proportional to the cosine metric. inter This represents the change in the maximum misclassification probability before and after the update. cross r represents the sum of probabilities that a new class sample is misclassified into an old class. cross The larger the value, the more knowledge from older classes can be retained. The final total reward r for reinforcement learning... t (x i The formalization is as follows:

[0094] r t (x i ) = r intra (x i )+r inter (x i )+r cross (x i ).

[0095] 3. Model Training Phase

[0096] In the basic training phase, a large amount of base class data was used to train the base model. In the incremental phase, the backbone parameters of the model were frozen, and only the classifier parameters were optimized. The model was trained for 2000 epochs using the Adam optimizer with a learning rate of 0.0003. Throughout the training process, data augmentation methods such as random pruning, random scaling, and random horizontal flipping were used.

[0097] 4. Model Testing Phase

[0098] After each incremental stage, we evaluate the trained model on the corresponding test set and report the classification accuracy.

[0099] Small sample exploratory incremental learning method: As shown in the figure, the method includes two parts: new classifier update and exploratory optimization. Figure 2

[0100] Image data is first extracted by the encoder, and then the metric value of each class is obtained by the classifier, and then the relevant metric value of the new class classifier weight W t is obtained as the classifier state S t , and then the classifier learning strategy a t is calculated by the gradient, and the new class classifier weight is updated from W t to W t+1 . When the classifier weight is updated, the new classifier metric is obtained by W t+1 , so that the next state S t+1 is obtained. Among them, the classifier update strategy a t is completed by the strategy exploration optimization part, and the classifier state S t is first sent to the Gaussian actor to obtain the mean value μ and variance σ of the strategy. On this basis, a classifier learning strategy can be selected by Gaussian sampling. On the other hand, the state S t is returned by the exploratory critic to obtain the current classifier state value v t , v t evaluates the classifier state from a long-term perspective. Then use the optimization reward value r t and the classifier state return value v t to calculate the critic loss L c (r t ;v t ), to supervise the update of the critic. At the same time, the classifier behavior loss L a (r t ;v t ; a t ) is used to supervise the update of the Gaussian actor . In the training process, the method ensures that the model can explore the optimal new class classifier learning strategy in the incremental stage through multiple jointly acting losses, thereby improving the final small sample class incremental recognition performance.

[0101] Design reward function using distance between new and old classes: In the optimization of the classifier, the performance of the updated classifier needs to be evaluated each time. On the other hand, the exploratory critic needs to evaluate the classifier in the long term, so the design of the reward function R is the key to the optimization direction. Comprehensive design from three aspects: (1) Intra-class compactness in new classes. Given the sample x i and the obtained classifier learning strategy a​t After the classifier weights are updated from W t to W t+1 , the performance of the sample x i on W t+1 should be better than on W t , so the cosine measure of the sample x i on W t+1 with the corresponding label y i should be greater than the cosine measure on the classifier W t , so a distance difference can be defined to represent the difference between the two cosine measures. 2) The inter-class dispersion of new class data. By analyzing the inter-class distance of the new class, by constraining the distance of different classes to be large, the sample is not easy to misclassify. (3) The dispersion between new and old classes. In the incremental process, all classes should be as dispersed as possible to reduce misclassification in the feature space. To this end, the distance between the new class sample and the old class weight is considered. Finally, by considering the above comprehensive consideration of intra-class and inter-class distance, the reward function is designed, which reduces the misclassification probability in the incremental stage. The disaster forgetting in small sample incremental learning is effectively alleviated.

[0102] Beneficial effects: the present application provides a small sample exploratory incremental learning method based on reinforcement learning, integrates the update of new class classifier and the exploratory optimization of new class classifier into an end-to-end closed loop learning process, and significantly improves the accuracy of small sample class incremental learning. Using the characteristics of trial and error learning in reinforcement learning, a small sample exploration strategy based on reinforcement learning is proposed to optimize the new classifier, and a Gaussian inference strategy sampling method is used to adaptively adjust the incremental model parameters to adapt to new knowledge data; a classifier reward function is designed to balance the influence of new and old classes, and the classifier is evaluated in the long run, and the optimization direction is adjusted in real time, which reduces the performance decline of the old class while improving the distinguishability of the new class, further alleviates the disaster forgetting problem, and improves the recognition performance of small sample incremental learning.

[0103] The above detailed description further describes the purpose, technical scheme and beneficial effects of the present application. It should be understood that the above is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A small sample class incremental identification method based on reinforcement learning, characterized in that, The recognition method comprises: Downloading a small sample class incremental learning data set; Designing a small sample incremental classification recognition network based on reinforcement learning; Training the network using training data to obtain a trained network; Generating a small sample class incremental learning model according to the trained network; Using the small sample class incremental learning model for small sample incremental classification recognition; The classifier updating specifically comprises: The model input is a data set CIFAR-100, the encoder φ uses ResNet20 as the backbone, and W t is the weight of the new class classifier at a certain stage t, if the new class contains M categories each time in the incremental stage, then W t =[w k ,…,w M ] T , wherein w k represents the classifier weight of the kth class, and d represents the feature dimension; For an input sample x from a batch data X i , the cosine similarity is used to measure the extracted feature embedding and the corresponding classification weight W t to describe the classification performance, and is defined as the classifier state After obtaining the current batch data prediction state S t , the cross-entropy loss of the current classifier is calculated: wherein, is a one-hot encoded label vector y i kth element of y, denotes the classifier state kth element of the classifier state, corresponding to the cosine distance of the sample x i to the kth class in the new class; the larger, the more similar to the kth class; obtaining the current time classifier weight parameter W t gradient of and the classifier learning strategy a obtained by the exploratory optimization module t together update the weight parameter: Next classifier state The updated weight W through this exploration t+1 , with sample x i Calculated; while learning strategy a t The reward Where R is the reward function; The reward function R specifically comprises: The reward r for evaluating intra-class tightness in the new class intra : wherein represents taking the operation corresponding to the label y i ; reward r for evaluating inter-class dispersion of new class data inter : When new classifier parameters W t+1 The input sample x i The reward score r inter becomes lower as it gets closer to other classes, take the maximum one as the penalty; Assessing the reward r dispersed between the new and old classes cross : where Y old represents the old class label of the current stage, the feature embedding of the sample x i the closer the classifier is to the old class, the smaller the reward r cross value, indicating the more misclassification cases.​ r intra represents the gain in the probability of correct classification of the updated classifier compared to the previous classifier, where the probability of correct classification is proportional to the cosine measure; r inter denotes the change in the maximum misclassification probability before and after the update; r cross r represents the sum of the probability values that a new class sample is misclassified into the old classes, cross The larger r is, the more knowledge of the old classes can be preserved. Total reward r of reinforcement learning t (x i ) formalization: r t (x i ) = r intra (x i ) + r inter (x i ) + r cross (x i ); The exploratory optimization specifically comprises: Utilizing critic ψ to evaluate classifier state S t to obtain long-term return v t , which is different from reward r t one step update of the classifier v t From state S t Start evaluating the performance of the classifier and use a t Continuously update the classifier; Next classifier state S t+1 The return value v is obtained by the exploratory critic ψ t , describes the next state S t+1 After long use of the policy a t The impact on the classifier performance; Using the idea of single-step difference in reinforcement learning to guide the learning of the critic ψ: Where γ is a proportionality coefficient used to scale future returns; Current difference and current classifier policy a t Adaptive supervision of the gaussian actor with log values Exploration optimization: as a supervisory signal actor loss L a guiding gaussian actor update to Next classifier state S t+1 By updated actor Learn classifier policy mean μ and variance σ, and get new classifier learning policy a by Gaussian sampling N(μ, σ) t+1 ; New strategy a t+1 For pushing the next update of the classifier parameters W t+1 of the new class classifier, alternating pushing the classifier update and the exploratory optimization of the classifier strategy, and implementing the integration of the update of the new class classifier and the exploratory optimization of the new class classifier into one end-to-end closed loop learning process.

2. The small sample class incremental identification method based on reinforcement learning according to claim 1, characterized in that, The downloading of the small sample class incremental learning data set specifically comprises: Collecting mainstream small sample incremental learning data set CIFAR-100, including 100 classes, each class containing 600 images; Each class has 500 training images and 100 test images, and the 100 classes in the small sample incremental learning data set CIFAR-100 are divided into 20 superclasses; Each image has a fine label and a coarse label; In accordance with the small sample incremental N-way K-shot setting, the small sample incremental learning data set CIFAR-100 data set is divided into 60 base classes and 40 new classes; After the base class training in the base stage is completed, the remaining 40 classes are further divided into 8 incremental stages, each of which is a 5-way 5-shot classification task.

3. The small sample class incremental identification method based on reinforcement learning according to claim 1, characterized in that, The design of the small sample incremental classification recognition network based on reinforcement learning specifically comprises two parts of classifier updating and exploratory optimization.

4. The small sample class incremental identification method based on reinforcement learning according to claim 1, characterized in that, The training of the network using training data to obtain a trained network specifically comprises: In the base training stage, the base model is trained using a large amount of base class data; In the incremental stage, the backbone parameters of the model are frozen, and only the classifier parameters are optimized; The model is trained for 2000 rounds using the Adam optimizer, with a learning rate of 0.0003; During the entire training process, data augmentation is performed.

5. The small sample class incremental identification method based on reinforcement learning according to claim 4, characterized in that, The method of data augmentation specifically comprises using random cropping, random scaling and random horizontal flipping for data augmentation.

6. The small sample class incremental identification method based on reinforcement learning according to claim 1, characterized in that, The recognition method further comprises: After the end of each incremental stage, the trained model is evaluated on the corresponding test set, and the classification accuracy is reported.

Images