A method for detecting unknown defect samples based on energy distribution

Abstract

The present invention relates to the technical field of defect detection of printed circuit board assemblies, and discloses a method for detecting unknown defect samples based on energy distribution. The test samples of printed circuit board assemblies are input into a trained detection model to obtain the detection results of the test samples. The construction and training process of the detection model includes: constructing a detection model; an optimal transmission mechanism based on energy; expanding the energy scores of known defect categories and unknown defect categories of printed circuit board assemblies through unsupervised training of inter-cluster expansion strategies, thereby enhancing the energy transmission cost, and then guiding the samples in the unlabeled data set to enter different clusters. The present invention assigns correct defect category labels to some unlabeled PCBA samples that meet the known defect categories, and then allows them to participate in joint training, thereby solving the problem of low defect detection accuracy outside the training data in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of PCBA defect detection, and in particular to a method for detecting unknown defect samples based on energy distribution. Background Art

[0002] During the production process of the SMT production line, defects in the printed circuit board assembly (PCBA) occur due to the complex and changeable process flow, which endangers the product function. At present, the defect detection methods of the PCBA production line can be specifically divided into manual visual inspection and automated optical inspection (AOI). In some production lines with high detection accuracy requirements, a joint re-inspection of the two is adopted, that is, after AOI, manual re-judgment is carried out. At present, structural defects are the main defects of assembled circuit boards, including insufficient solder, voids and short circuits. However, due to improvements in PCBA production line technology, rule-based systems have become unable to adapt to PCBA defect detection due to the complexity and cumbersomeness of managing a large number of rules, the increasing demand for expert knowledge, and the inability to handle fuzzy and uncertain boundaries. They are gradually being replaced by data-driven model-based systems, such as defect detection based on deep learning methods. In the prior art, there are already a large number of cases where deep learning has been applied to PCBA defect detection. Although these cases have improved the accuracy of defect category detection within the training distribution, deep learning still does not have generalizability for defects outside the training data generated in real time by the PCBA production line. That is, the application of deep learning to PCBA defect detection still has the following problems: 1) Since the real-time data stream may have a distribution different from that of the training samples, the offline training model does not have adaptive capabilities; 2) The offline training model cannot detect new defect categories.

[0003] Optimal Transport (OT) is a mathematical theory that describes the optimal way to transfer data between two different distributions. This theory can be used in a variety of application scenarios, such as image processing, machine learning, statistics, and computational fluid dynamics.

[0004] OT theory regards two distributions as two "quality" distributions. It is necessary to find a transmission method that can most effectively transmit data between the two distributions while keeping the "quality" of the two distributions unchanged. This transmission method can be achieved by solving a mathematical problem, called the "optimal transmission problem."

[0005] OT theory can be used to solve the problem of transfer learning, that is, how to transfer learning between different data sets. It can also be used for training in generative adversarial networks (GANs), as well as adversarial sample mining, etc. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a method for detecting unknown defect samples based on energy distribution.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A method for detecting unknown defect samples based on energy distribution, inputting a test sample of a printed circuit board assembly into a trained detection model to obtain a defect type detection result of the test sample; the construction and training process of the detection model includes the following steps:

[0009] Step 1: Build a detection model:

[0010] The detection model includes the backbone network, the classifier H cls , and the classifier H cls Parallel K-dimensional optimal transport clusterer H ot ;

[0011] Extract training samples x through the backbone network i The feature code z i , then the i-th training sample x i Probability for: c is a K-dimensional vector, used to represent K clusters, softmax(·) represents the softmax function; training sample x i It is an image acquired of a printed circuit board assembly;

[0012] Step 2: The optimal transmission mechanism based on energy includes the following steps:

[0013] Step S21, in the optimal transmission problem, an energy score is introduced, and samples with energy scores greater than or equal to a threshold are assigned to the same cluster, and samples with energy scores less than the threshold are assigned to any other cluster; the energy transmission cost of the energy-based optimal transmission mechanism is obtained through the energy score and the transmission cost of the optimal transmission problem; the allocation matrix is ​​optimized by calculating the optimal transmission distance of the optimal transmission problem through the energy transmission cost and the allocation matrix in the optimal transmission problem; N training samples are mapped to K clusters through the optimized allocation matrix;

[0014] Step S22: The training process of the detection model has multiple training stages. In the tth training stage, the number of known defect categories y∈y in cluster k is calculated. (t) The sample proportion The unlabeled dataset The samples exceeding the set threshold τ are transferred to the labeled data set, and the samples in the unlabeled data set are evenly distributed among the M known defect categories; y (t)is a collection of known defect categories for printed circuit board assemblies;

[0015] Step S23: Sample x i The feature code z i Input to classifier H cls , the output of the classifier H cls (z i ) as the score of the unknown defect category of the printed circuit board assembly;

[0016] Step three, the energy scores of known defect categories and unknown defect categories of printed circuit board assemblies are expanded through unsupervised training of the inter-cluster expansion strategy, thereby enhancing the discrimination between known defect samples and unknown defect samples, and then guiding the samples in the unlabeled dataset into different clusters.

[0017] Furthermore, step S21 specifically includes:

[0018] The cost matrix in the optimal transport problem is defined as Cluster Probability Represents x i Belongs to the jth cluster c j probability; represents the allocation matrix, Q ji =Q(c j |x i ) represents x i Assign to c j The posterior probability; when the N-dimensional vector β is used to represent the N training samples and the K-dimensional vector α is used to represent the distribution of the K clusters, all feasible solutions U(α,β) of the assignment matrix Q in the transport polyhedron are expressed as:

[0019]

[0020] Where I is a vector of all 1s of the corresponding dimension, and vector α and vector β represent the marginal projections of the allocation matrix Q on rows and columns, respectively;

[0021] Introducing energy scores for distinguishing known defect classes from unknown defect classes

[0022]

[0023] Where l(c j |x i )=H ot (z i ), represents the sample x i Belongs to cluster c j Logical fraction of e i represents the energy score of the i-th sample;

[0024] The training samples with energy scores greater than or equal to the energy threshold are assigned to the same cluster, and the training samples with energy scores less than the energy threshold are assigned to other K clusters; the allocation matrix Q based on energy e is expressed as:

[0025]

[0026] Energy transmission cost P en It is expressed as:

[0027]

[0028] The energy score e is first broadcast into a K×N matrix and then multiplied element-wise with the cost matrix P; therefore, the optimal transmission distance OT(α,β) between α and β is:

[0029] OT(α,β)=min Q∈∏(α,β) - <Q,P en >

[0030] in <Q,P en > represents the Frobenius inner product, ∏(α,β) represents the continuous multiplication operation;

[0031] The entropy regularization term H(Q) is introduced into the Wasserstein distance, and the optimization objective of the optimal transmission problem is expressed as:

[0032] OT(α,β)=min Q∈∏(α,β) - <Q,P en >+εH(Q);

[0033] Where ε>0, H(Q)=∑ ji Q ji log D ji , optimize the allocation matrix Q:

[0034]

[0035] And through the Sinkhorn algorithm, u and v are solved;

[0036] The N training samples are mapped to K clusters through the assignment matrix Q.

[0037] Furthermore, step S22 specifically includes:

[0038] The training samples belonging to the kth cluster in the tth training stage form a set D k :

[0039]

[0040] represents the cluster index of the i-th sample in the t-th training stage;

[0041] In the tth training stage, define the set y of known defect categories (t) for:

[0042]

[0043] Where M represents the number of known defect categories of PCB assemblies, y l The labeled training sample set D l The ground truth labels in represents the pseudo-label assigned to the training samples in the unlabeled dataset; at t = 0, And it will be updated during the training process, and y l Remain unchanged; calculate the known defect category y∈y in cluster k (t) The proportion of training samples

[0044]

[0045] When all unlabeled samples in cluster k When it exceeds the threshold τ, it will be added to the labeled dataset D L :

[0046]

[0047] represents the labeled dataset for the tth training stage; the training samples added to the labeled dataset are added to the unlabeled dataset Removed; the final progressive dataset optimization goal is:

[0048]

[0049]

[0050] represents the classification loss, represents the equilibrium loss, p(y|x i ) represents the training sample x i The predicted probability of belonging to all known defect categories, and y i One-hot encoding, indicating x i The corresponding label of represents a uniform posterior distribution over all M PCB assemblies’ known defect classes, and Will force The training samples in are evenly distributed among the M known defect categories of PCB assemblies.

[0051] Further, step S23 specifically includes:

[0052] Using the classifier H cls The output of l(y|x i )=H cls (z i ) as the score of the unknown defect class of the printed circuit board assembly:

[0053]

[0054] Where l(y m |x i ) represents x i Belong to y m The logical score of the defect, the higher the score, the more it belongs to the known defect category, y m represents the mth known category label, e(x i ) represents the energy fraction;

[0055] In order to smooth the energy distribution and avoid the unknown defect category samples from concentrating at the minimum energy, the temperature T is set, and the T energy fraction is expressed as follows:

[0056]

[0057] T e (x i ) indicates a PCBA sample without label.

[0058] Furthermore, step three specifically includes: in the unlabeled data, using the labeled data set D L The training sample set and from the unlabeled dataset D U The training sample set Construct a joint training sample set {x i} i＝1...B ={x (l,i)}∪{x (u,i)},B=B 1 +B 2 ;x (l,i) Represents a labeled dataset D L The i-th training sample of (u,i) Represents a training sample from an unlabeled dataset; then, two different dataset augmentation strategies are used to obtain the augmented image of the same training sample: Indicates that the first data set enhancement strategy A 0 The first enhanced image obtained; Indicates that the second data set enhancement strategy A 1 The obtained second enhanced image;

[0059] Extracted through the same backbone network Features:

[0060]

[0061] where e θ represents the backbone network parameters, H m Represents a multi-layer perceptron; considering the lack of data labels, contrastive learning loss L is used rep :

[0062]

[0063] The feature code z j ∈Z 1,(t′) , set Z 1,(t′) express t′ represents the dynamic memory pool, n represents the latest iteration number, and cos represents the pre-similarity; through L rep The obtained enhanced representations of training samples will be mapped to a low-dimensional logical space in order to assign more discriminative cluster distributions to the training samples.

[0064] Compared with the prior art, the beneficial technical effects of the present invention are:

[0065] The present invention designs an optimal transmission (OT) scheme based on uncertainty perception, which includes an energy transmission mechanism (ET) and an inter-cluster expansion strategy (Lrep). It assigns correct defect category labels to some unlabeled PCBA samples that meet known defect categories, and then allows them to participate in joint training, thereby solving the problem of low defect detection accuracy outside training data in the prior art.

[0066] The energy transfer mechanism introduces the energy score as an uncertainty measure and estimates the uncertainty-based transfer cost to guide the cluster distribution of all samples.

[0067] To further promote the discrimination in the logical space, where uncertainty can significantly reflect the differences between samples of known defect categories and unknown defect categories, the inter-cluster expansion strategy enhances the global feature representation of mixed known defect categories and unknown defect category samples, and then the enhanced global feature representation will be mapped into a more discriminative logical space. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0069] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0070] Figure 1The overall process of the present invention is demonstrated. The unknown defect sample detection method based on energy distribution in the present invention specifically includes the following contents.

[0071] 1. Energy-based transmission mechanism

[0072] 1.1 Clustering of unknown samples

[0073] Different from the known PCBA defect category detection, the task of our present invention is to assign labels under the interference of unknown defect samples. We propose an energy-based transfer (ET) mechanism based on logical space to more fully explore the semantic differences and hidden knowledge in unlabeled datasets.

[0074] Specifically, the present invention is based on the classifier H cls The K-dimensional OT clusterer is introduced in parallel and is denoted as H ot , used to cluster samples into K cluster centers. For a given N samples And the feature encoding z∈{z 1 ,z 2 ,...,z N}, define the probability of the i-th training class: c is a K-dimensional vector used to represent K clusters. Then, the present invention can define the cost matrix in the optimal transmission problem as Represents x i Belongs to c j The probability of c j is the jth cluster. Similarly, the present invention will Expressed as a distribution matrix, Q ji =Q(c j |x i ) represents x i Assign to c j It should be noted that the assignment matrix Q only represents the assignment of a cluster to each sample, rather than directly assigning labels to the samples. When the present invention uses an N-dimensional vector β to represent the distribution of N samples and a K-dimensional vector α to represent the distribution of K clusters, all feasible solutions of the assignment matrix Q in the transmission polyhedron can be expressed as:

[0075]

[0076] where I is a vector of all ones of the corresponding dimension, and α and β are the marginal projections of the assignment matrix Q onto its rows and columns, respectively.

[0077] The present invention will then be used to distinguish the energy scores of known defect classes and unknown defect classes. The transmission optimization module is introduced and is defined as follows:

[0078]

[0079] Where l(c j |x i )=H ot (z i ), represents the sample x i Belongs to cluster c j The logical fraction of .

[0080] The energy-based transfer mechanism (ET) will encourage samples with higher energy scores to be assigned to the same cluster, while those samples with lower energy scores, meaning they have greater uncertainty in cluster distribution, will tend to be evenly distributed among the K clusters. The energy-based allocation matrix Q can be expressed as:

[0081]

[0082] Energy transmission cost P en It can be expressed as:

[0083]

[0084] e is first broadcasted into a K×N matrix and then multiplied element-wise with P. Therefore, the Wasserstein distance OT(α,β) between α and β is defined as:

[0085] OT(α,β)=min Q∈∏(α,β) - <Q,P en >

[0086] Where <*, *> represents the Frobenius inner product.

[0087] In order to avoid linear programming problems that require a lot of computational costs, the present invention introduces the entropy regularization term H(Q) into the Wasserstein distance and expresses the optimization problem as:

[0088] OT(α,β)=min Q∈∏(α,β) - <Q,P en >+εH(Q);

[0089] Where ε>0, H(Q)=∑ ji Q ji log Q ji , and optimize Q at this time:

[0090]

[0091] Here, the exponential operation is performed element by element, and through the Sinkhorn algorithm, u and v can be solved faster. The assignment matrix Q maps N samples to K clusters. That is, samples less than the energy threshold are samples of known defect categories, and the types of known defect categories are K categories.

[0092] So far, the present invention can allocate all samples to K clusters through the allocation matrix Q.

[0093] 1.2 Progressive training sample construction

[0094] Specifically, the present invention allows the samples belonging to the kth cluster in the tth training stage to form a set D k , recorded as:

[0095]

[0096] In the tth training stage, the present invention defines the set of known PCBA defect category labels as:

[0097]

[0098] Where M represents the number of known categories of PCBA defects, y l By D l The ground truth labels in represents the pseudo label assigned to the unlabeled sample. Note that at t = 0, And it will be updated during the training process, and y l Then, the present invention calculates the category y∈y belonging to cluster k (t) The sample proportions are as follows:

[0099]

[0100] When all unlabeled samples in cluster k When it exceeds the threshold τ, it will be added to the labeled dataset D L :

[0101]

[0102] At the same time, remove this part of data from the unlabeled dataset. The final progressive dataset optimization objective can be written as:

[0103]

[0104]

[0105] p(y|x i ) represents x i The predicted probability of belonging to all known defect categories, and y iOne-hot encoding, indicating x i In addition, represents the uniform posterior distribution over all M PCBA known defect categories, and Will force The samples in are evenly distributed among the M known PCBA categories.

[0106] 1.3 Energy score calculation

[0107] Considering H cls and H OT They all map the same features to the logical space through different fully connected layers, and there is a certain correlation between them. Therefore, the present invention uses the H cls Output l(y|x i )=H cls (z i ) is used as the score of the unknown defect category of PCBA and is used to detect unknown samples after training. Specifically:

[0108]

[0109] Where l(y m |x ii ) represents x i Belong to y m The higher the score, the more it belongs to the known defect category.

[0110] In order to smooth the energy distribution and avoid the concentration of the unknown defect class sample at the minimum energy, the present invention sets a larger temperature T. The specific T energy fraction is expressed as follows:

[0111]

[0112] 2. Inter-cluster expansion strategy

[0113] In order to expand the energy scores of known defect categories and unknown defect categories of PCBA, thereby enhancing the energy transfer cost and the ability to guide unlabeled samples into different clusters, the present invention further introduces an unsupervised training module called inter-cluster expansion strategy to obtain improved feature representation.

[0114] In the case of unlabeled data, the present invention uses two L and D U Data batches and Construct joint training data {x i} i＝1...B ={x (l,i)}∪{x (u,i)},B=B1 +B 2 Afterwards, two different dataset enhancement strategies are used to obtain the same enhanced image, which is recorded as: Extract features through the same backbone network:

[0115]

[0116] where e θ represents the backbone network parameters, H m Represents a multi-layer perceptron. Considering the lack of data labels, the present invention uses InfoNCE loss L rep :

[0117]

[0118] where z j ∈Z 1,(t') , cos represents the pre-similarity. rep The obtained enhanced representation will be mapped to a low-dimensional logical space, in which the present invention can better obtain the energy scores of specific categories to better reflect the semantic differences between samples of different categories and assign them more discriminative cluster distributions.

[0119] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0120] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for detecting unknown defect samples based on energy distribution, wherein a test sample of a printed circuit board assembly is input into a trained detection model to obtain a defect type detection result of the test sample; the construction and training process of the detection model includes the following steps: Step 1: Build a detection model: The detection model includes the backbone network, the classifier H cls , and the classifier H cls Parallel K-dimensional optimal transport clusterer H ot ; Extract training samples x through the backbone network i The feature code z i , then the i-th training sample x i Probability for: c is a K-dimensional vector, used to represent K clusters, softmax(·) represents the softmax function; training sample x i It is an image acquired of a printed circuit board assembly; Step 2: The optimal transmission mechanism based on energy includes the following steps: Step S21, in the optimal transmission problem, an energy score is introduced, and samples with energy scores greater than or equal to a threshold are assigned to the same cluster, and samples with energy scores less than the threshold are assigned to any other cluster; the energy transmission cost of the energy-based optimal transmission mechanism is obtained through the energy score and the transmission cost of the optimal transmission problem; the allocation matrix is ​​optimized by calculating the optimal transmission distance of the optimal transmission problem through the energy transmission cost and the allocation matrix in the optimal transmission problem; N training samples are mapped to K clusters through the optimized allocation matrix; Step S22: The training process of the detection model has multiple training stages. In the tth training stage, the number of known defect categories y∈y in cluster k is calculated. (t) The sample proportion The unlabeled dataset The samples exceeding the set threshold τ are transferred to the labeled data set, and the samples in the unlabeled data set are evenly distributed among the M known defect categories; y (t) is a collection of known defect categories for printed circuit board assemblies; Step S23: Sample x i The feature code z i Input to classifier H cls , the output of the classifier H cls (z i ) as the score of the unknown defect category of the printed circuit board assembly; Step three, the energy scores of known defect categories and unknown defect categories of printed circuit board assemblies are expanded through unsupervised training of the inter-cluster expansion strategy, thereby enhancing the discrimination between known defect samples and unknown defect samples, and then guiding the samples in the unlabeled dataset into different clusters.

2. The unknown defect sample detection method based on energy distribution according to claim 1 is characterized in that: Step S21 specifically includes: The cost matrix in the optimal transport problem is defined as Cluster Probability Represents x i Belongs to the jth cluster c j probability; represents the allocation matrix, Q ji =Q(c j |x i ) represents x i Assign to c j The posterior probability of; when the N-dimensional vector β is used to represent the N training samples and the K-dimensional vector α is used to represent the distribution of the K clusters, all feasible solutions U(α, β) of the assignment matrix Q in the transport polyhedron are expressed as: Where I is a vector of all 1s of the corresponding dimension, and vector α and vector β represent the marginal projections of the allocation matrix Q on rows and columns, respectively; Introducing energy scores for distinguishing known defect classes from unknown defect classes Where l(c j |x i )=H ot (z i ), represents the sample x i Belongs to cluster c j Logical fraction of e i represents the energy score of the i-th sample; The training samples with energy scores greater than or equal to the energy threshold are assigned to the same cluster, and the training samples with energy scores less than the energy threshold are assigned to other K clusters; the allocation matrix Q based on energy e is expressed as: Energy transmission cost P en It is expressed as: The energy score e is first broadcast into a K×N matrix and then multiplied element-wise with the cost matrix P; therefore, the optimal transmission distance OT(α, β) between α and β is: OT(α,β)=min Q∈Π(α，β )- <Q,P en >; in <Q,P en > represents the Frobenius inner product, Π(α, β) represents the continuous multiplication operation; The entropy regularization term H(Q) is introduced into the Wasserstein distance, and the optimization objective of the optimal transmission problem is expressed as: OT(α,β)=min Q∈H(α，β) - <Q,P en >+εH(Q); Where ε>0, H(Q)=∑ ji Q ji log Q ji , optimize the allocation matrix Q: And through the Sinkhorn algorithm, u and v are solved; The N training samples are mapped to K clusters through the assignment matrix Q.

3. The unknown defect sample detection method based on energy distribution according to claim 1, characterized in that: Step S22 specifically includes: The training samples belonging to the kth cluster in the tth training stage form a set D k : represents the cluster index of the i-th sample in the t-th training stage; In the tth training stage, define the set y of known defect categories (t) for: Where M represents the number of known defect categories of PCB assemblies, y l The labeled training sample set D l The ground truth labels in represents the pseudo-label assigned to the training samples in the unlabeled dataset; at t = 0, And it will be updated during the training process, and y l Remain unchanged; calculate the known defect category y∈y in cluster k (t) The proportion of training samples When all unlabeled samples in cluster k When it exceeds the threshold τ, it will be added to the labeled dataset D L : represents the labeled dataset for the tth training stage; the training samples added to the labeled dataset are added to the unlabeled dataset Removed; the final progressive dataset optimization goal is: represents the classification loss, represents the equilibrium loss, p(y|x i ) represents the training sample x i The predicted probability of belonging to all known defect categories, and y i One-hot encoding, indicating x i The corresponding label of represents a uniform posterior distribution over all M PCB assemblies’ known defect classes, and Will force The training samples in are evenly distributed among the M known defect categories of PCB assemblies.

4. The unknown defect sample detection method based on energy distribution according to claim 1, characterized in that: Step S23 specifically includes: Using the classifier H cls The output of l(y|x i )=H cls (z i ) as the score of the unknown defect class of the printed circuit board assembly: Where l(y m |x i ) represents x i Belong to y m The logical score of the defect, the higher the score, the more it belongs to the known defect category, y m represents the mth known category label, e(x i ) represents the energy fraction; In order to smooth the energy distribution and avoid the unknown defect category samples from concentrating at the minimum energy, the temperature T is set, and the T energy fraction is expressed as follows: T e (x i ) represents unlabeled samples.

5. The unknown defect sample detection method based on energy distribution according to claim 1, characterized in that: Step 3 specifically includes: In the unlabeled data, use the labeled data set D L The training sample set and from the unlabeled dataset D U The training sample set Construct a joint training sample set {x i } i＝1...B ={x (l，i) }∪{x (u，i) }, B = B1 + B2; x (l，i) Represents a labeled dataset D L The i-th training sample of (u，i) Represents a training sample from an unlabeled dataset; then, two different dataset augmentation strategies are used to obtain the augmented image of the same training sample: Indicates that the first data set enhancement strategy A 0 The first enhanced image obtained; Indicates that the second data set enhancement strategy A 1 The obtained second enhanced image; Extracted through the same backbone network Features: where e θ represents the backbone network parameters, H m Represents a multi-layer perceptron; considering the lack of data labels, contrastive learning loss L is used rep : The feature code z j ∈Z 1，(t′) , set Z 1，(t′) express t′ represents the dynamic memory pool, n represents the latest iteration number, and cos represents the pre-similarity; through L rep The obtained enhanced representations of training samples will be mapped to a low-dimensional logical space in order to assign more discriminative cluster distributions to the training samples.

Images