A vehicle-mounted human-computer interaction software visual defect sample generalization method and system
By imposing similarity and difference constraints on the generative model and performing two rounds of screening, high-quality abnormal samples of in-vehicle display screens are generated, solving the problem of insufficient sample generalization ability in existing technologies and improving the accuracy and adaptability of the detection model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU CAERI AUTOMOTIVE ENG RES INST CO LTD
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-29
AI Technical Summary
In the existing technology, the visual recognition system for anomaly detection of vehicle display screens lacks an efficient sample generalization method, which limits the accuracy and generalization ability of the detection model, and the generated samples have insufficient similarity or difference with the original samples.
When generating samples using a pre-trained generative model, similarity and difference constraints are imposed, and a two-stage screening process is used to ensure structural consistency and content differences between the generated samples and the original samples, thus forming a high-quality generalized sample set.
It improves the generalization ability of the sample library, and the generated samples have sufficient diversity while retaining effective defect features, thereby improving the accuracy and adaptability of the detection model.
Smart Images
Figure CN122116031A_ABST
Abstract
Description
Technical Field
[0001] This specification relates to the field of computer technology, and in particular to a method and system for generalizing visual defect samples in vehicle-mounted human-computer interaction software. Background Technology
[0002] With the rapid development of automotive intelligence and connectivity, the functions of in-vehicle display screens have become increasingly complex. Abnormalities (such as black screens, red screens, white screens, green screens, blue screens, and distorted screens) can seriously impact driving safety and efficiency. Currently, the detection of anomalies in in-vehicle automated software mainly relies on visual recognition systems, but the training and testing of these systems require a large number of defect samples. However, because in-vehicle display defects are difficult to collect in real-world environments, the number of available defect samples is limited, severely restricting the accuracy and generalization ability of detection models.
[0003] In existing technologies, some methods expand samples using traditional image enhancement techniques (such as rotation, cropping, and brightness adjustment). However, these methods can only change the external morphology of the samples and cannot generate new defect features, making it difficult to meet the diverse sample requirements of detection systems. Other methods use ordinary generative models to generate defect samples, but lack adaptability to in-vehicle scenarios and fail to balance the similarity and differences between samples, resulting in generated samples that are either overly similar to the original samples or have invalid defect features. In other words, existing methods lack an efficient screening system—either relying solely on automatic screening leads to the omission of residual redundancy or invalid samples, affecting the quality of sample expansion. Summary of the Invention
[0004] This specification provides a method and system for generalizing visual defect samples in vehicle-mounted human-computer interaction software, in order to solve the problem of low quality of generalized samples in the prior art.
[0005] The following technical solution is adopted in this specification: A method for generalizing visual defect samples in vehicle-mounted human-computer interaction software includes the following steps: S1: Obtain visual defect sample data of in-vehicle human-machine interaction software, and classify and label the sample data to obtain multiple original sample datasets with different feature labels; S2: For each original sample in the original sample dataset, a pre-trained generative model is used to generate corresponding generated samples according to a preset ratio, and each generated sample must simultaneously satisfy a preset similarity constraint and a preset difference constraint. S3: Based on the pre-set screening conditions, each of the generated samples is subjected to a first screening and a second screening to obtain the final generated sample; S4: Mix the final generated sample with each sample in the original sample dataset to obtain the generalized sample set.
[0006] Based on the above technical means, when generating samples using the generative model, similarity constraints and difference constraints are applied. On the one hand, this ensures the structural consistency between the generated samples and the original samples, and on the other hand, it allows the generated samples to have content differences from the original samples. This ensures that the generated samples have sufficient diversity while retaining effective defect features, thereby improving the generalization ability of the sample library. Furthermore, it allows for two rounds of screening of the generated samples based on certain conditions, so that the final generated samples can better meet expectations, thus increasing the quality of the generated samples.
[0007] Furthermore, the similarity constraint in S2 specifically means that the similarity value between the original sample and the corresponding generated sample is not lower than a preset similarity value range.
[0008] Furthermore, the similarity value between the original sample and the corresponding generated sample in S2 is calculated as follows: = , in, For similarity values, For a given original sample, For the corresponding generated samples, It is a structural similarity index. For high-frequency information matching degree, For low-frequency information matching degree, The weighting coefficients are the structural similarity index, the high-frequency information matching degree, and the low-frequency information matching degree.
[0009] Based on the aforementioned technical means, the similarity constraints include structural similarity, high-frequency information matching degree, and low-frequency information matching degree, ensuring the structural similarity between the generated sample and the original sample through similarity constraints.
[0010] Furthermore, the difference constraint in S2 specifically means that the difference value between the original sample and the corresponding generated sample is not lower than a preset difference value range.
[0011] Furthermore, the calculation method for the difference constraint value between the original sample and the corresponding generated sample in S2 is as follows: , in For the difference value, The pixel distribution difference rate This refers to the degree of difference in detail within the defective area. For the uniqueness of local features, , , These are the weighting coefficients for pixel distribution difference rate, defect area detail difference, and local feature uniqueness, respectively.
[0012] Based on the aforementioned technical means, the difference constraints include pixel distribution difference rate, defect area detail difference degree, and local feature uniqueness, ensuring the content difference between the generated sample and the original sample through the difference constraints.
[0013] Furthermore, in S2, the similarity constraint is that the similarity value between the original sample and the corresponding generated sample is not lower than a preset similarity value range; in S2, the difference constraint is that the difference value between the original sample and the corresponding generated sample is not lower than a preset difference value range.
[0014] Furthermore, the first screening in S3 specifically includes: calculating the similarity between each pair of generated samples to obtain a first similarity; when the value of the first similarity is not less than a preset first threshold, randomly removing one of the generated samples.
[0015] Based on the above technical means, for the generated samples, the first similarity is calculated between each pair. When the first similarity exceeds a certain threshold, it means that the two generated samples are too similar and cannot achieve the purpose of generalization. In addition, one of the corresponding generated samples is removed, that is, redundant samples that are too similar within the generated samples are removed, which increases the quality of the generalized samples.
[0016] Furthermore, the second screening in S3 specifically includes: calculating the similarity between the generated sample that has undergone the first screening and each sample in the original sample dataset to obtain the corresponding second similarity; when the value of the second similarity is not less than a preset second threshold, the corresponding generated sample is removed.
[0017] Based on the above technical means, by comparing the similarity of each generated sample with all samples in the original library, it is possible to effectively identify and remove those generated samples that are too similar to existing defective features, thus preventing invalid expansion. Specifically, for each generated sample and other original samples, a second similarity is calculated. When the second similarity exceeds a certain threshold, it indicates that the generated sample is too similar to the original sample, and the corresponding generated sample is removed. This ensures that the generated samples that pass the second screening need to be different from the original samples, thereby continuously expanding the coverage of the sample library in the feature space.
[0018] Furthermore, the calculation method for the first similarity and / or the second similarity includes the following steps: S31: Extract the feature vectors of the samples to be compared, and calculate the cosine similarity between the feature vectors as the first similarity component; S32: Obtain the structural similarity index between the samples to be compared as the second similarity component; S33: The first similarity component and the second similarity component are weighted and added together to obtain the first similarity and / or the second similarity.
[0019] Based on the aforementioned technical means, when calculating the first similarity and / or the second similarity, the calculation of the first similarity and / or the second similarity integrates the cosine similarity algorithm and the structural similarity fusion algorithm. The first similarity component mainly reflects the similarity of images in high-level semantics and abstract content by extracting and comparing the feature vectors of the images. The second similarity component directly calculates the similarity between the two images in low-level visual features such as brightness, contrast, and structure, while capturing similarity information from semantic concepts to visual details. This overcomes the limitations of a single measurement perspective and makes the final similarity evaluation result more comprehensive.
[0020] Furthermore, the pre-trained generative model in S2 is specifically a stable diffusion model trained based on defect samples.
[0021] Furthermore, a generalization system for visual defect samples in vehicle-mounted human-machine interaction software is provided, applying the aforementioned method, comprising: a data acquisition and annotation module, a constraint generation module, a filtering module, and a fusion module connected in sequence; the data acquisition and annotation module is used to acquire and process original sample data to form an original sample dataset with multi-dimensional feature labels; the constraint generation module is used to generate generated samples that simultaneously satisfy preset similarity constraints and difference constraints based on the original sample dataset and using a pre-trained generative model; the filtering module is used to sequentially perform a first filtering of the similarity between generated samples and a second filtering of the similarity between generated samples and the original samples to obtain the final generated samples; the fusion module is used to mix the final generated samples with the original sample dataset to obtain and output the generalized sample set.
[0022] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: When generating samples, the generative model applies similarity and difference constraints. On the one hand, it ensures the structural consistency between the generated samples and the original samples, and on the other hand, it allows the generated samples to have content differences from the original samples. This ensures that the generated samples have sufficient diversity while retaining effective defect features, thereby improving the generalization ability of the sample library. It can also perform two screenings on the generated samples based on certain conditions, so that the final generated samples can better meet expectations, thereby increasing the quality of the generated samples. Attached Figure Description
[0023] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall structure of Example 1; Figure 2 This is a flowchart of the first similarity / second similarity acquisition process in Example 1; Figure 3 This is a structural diagram of Example 2.
[0024] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings. The same or similar reference numerals correspond to the same or similar components. The terms describing positional relationships in the drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.
[0026] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.
[0027] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.
[0028] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.
[0029] In the embodiments of this application, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium.
[0030] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0031] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.
[0032] Example 1 like Figure 1 As shown, this embodiment provides a method for generalizing visual defect samples of in-vehicle human-machine interaction software, including the following steps: acquiring visual defect sample data of in-vehicle human-machine interaction software, and classifying and labeling the sample data to obtain an original sample dataset with multiple labels with different features.
[0033] In this preferred embodiment, the classification types of each original sample data specifically include: divided into six categories according to defect type: black screen, red screen, white screen, green screen, blue screen, and distorted screen.
[0034] In this preferred embodiment, the annotation of each original sample data specifically involves: secondary classification based on fine-grained feature dimensions such as blur ratio (0%-100%, graded in 10% intervals), defect area ratio (0%-100%, graded in 10% intervals), brightness uniformity (0-10 points, with higher scores indicating better uniformity), color deviation (0-5 levels, with higher levels indicating more significant deviation), and texture complexity (simple / medium / complex levels), assigning a unique multi-dimensional feature label to each sample. For example, the feature labels for a sample with a distorted screen are: distorted screen—blur ratio 30%—defect area ratio 60%—brightness uniformity 4 points—color deviation level 2—texture complexity medium.
[0035] like Figure 1As shown, in this embodiment, S2: For each original sample in the original sample dataset, a pre-trained generative model is used to generate corresponding generated samples according to a preset ratio, and each generated sample must simultaneously satisfy preset similarity constraints and preset difference constraints.
[0036] In this embodiment, the pre-trained generative model in S2 is specifically a stable diffusion model trained based on defect samples. Specifically, the pre-trained generative model optimizes the parameters of the stable diffusion model through pre-training with vehicle-mounted visual defect samples, ensuring the authenticity of the defect features and scene adaptability of the generated samples.
[0037] In this embodiment, the preset ratio is 1:5.
[0038] In other embodiments, the preset ratio is 1:n, where n∈[3,8].
[0039] In this embodiment, the similarity constraint in S2 is specifically: the similarity value between the original sample and the corresponding generated sample is not lower than a preset similarity value range.
[0040] The similarity value between the original sample and the corresponding generated sample in S2 is calculated as follows: = , in, For similarity values, For a given original sample, For the corresponding generated samples, It is a structural similarity index. For high-frequency information matching degree, For low-frequency information matching degree, These are the weighting coefficients for the structural similarity index, the high-frequency information matching degree, and the low-frequency information matching degree, respectively. The sum of is 1.
[0041] In this embodiment, the formula for the high-frequency information matching degree is: , in, This represents the pixel value of the original sample at coordinates (x, y), where W and H are the width and height of the sample image; This represents the pixel value at coordinates (x, y) of the generated generalized sample; HF(I(x, y)): the high-frequency component of the image pixel, with the high-frequency information matching degree ranging from [0, 1]. The closer the value is to 1, the higher the matching degree of the generated sample with the high-frequency defect details (such as screen noise, red screen boundaries, etc.) of the original sample. HF(I(x,y)) The result is: .
[0042] In this embodiment, the formula for the low-frequency information matching degree is: , Where: LF(I(x,y)) represents the low-frequency component of the image pixel, which is the pixel value obtained after convolution with a Gaussian blur kernel; the range of low-frequency information matching degree is [0,1]. The closer the value is to 1, the more consistent the overall structure of the generated sample with the original sample (such as the position of the defect on the screen, the full-screen / local defect morphology).
[0043] In this embodiment, the difference constraint in S2 is specifically: the difference value between the original sample and the corresponding generated sample is not lower than a preset difference value range.
[0044] The difference between the original sample and the corresponding generated sample in S2 is calculated as follows: , in For the difference value, The pixel distribution difference rate This refers to the degree of difference in detail within the defective area. For the uniqueness of local features, , , These are the weighting coefficients for pixel distribution difference rate, defect region detail difference, and local feature uniqueness, respectively. , , The sum of is 1.
[0045] In this embodiment, the formula for calculating the degree of detail difference in the defect area is: , in: R This represents the defect area of the original sample; the range of the defect area detail difference is [0,1]. The closer the value is to 0, the more overlapping the defect area details are, and the higher the value is, the more significant the difference in defect details is.
[0046] In this embodiment, the formula for calculating the uniqueness of local features is as follows: , in, This represents the number of SIFT feature points in the original sample. This indicates the number of SIFT feature points in the generated sample; This indicates the number of matching feature points between the original sample and the generated sample (matched by Euclidean distance, with a distance threshold of 1.5). The range of local feature uniqueness is [0,1]. The closer the value is to 1, the more unique the local features of the generated sample are. In this embodiment, the similarity constraint in S2 is that the similarity value between the original sample and the corresponding generated sample is not lower than a preset similarity value range; the difference constraint in S2 is that the difference value between the original sample and the corresponding generated sample is not lower than a preset difference value range.
[0047] In this embodiment, the preferred similarity value range is [0.7, 1.0], and the preferred difference value range is [0.2, 0.5].
[0048] like Figure 1 As shown, in this embodiment, S3: according to the preset screening conditions, each of the generated samples is subjected to a first screening and a second screening to obtain the final generated sample.
[0049] The first screening in S3 specifically includes: calculating the similarity between each pair of generated samples to obtain a first similarity; when the value of the first similarity is not less than a preset first threshold, randomly removing one of the generated samples.
[0050] In this preferred embodiment, during the first screening, when the value of the first similarity is not less than a preset first threshold, the generated sample with richer defect features is retained or the sample with lower defect features is removed by comparing the differences between the two generated samples. For example, the generated sample with a more uniform distribution of fuzziness and more defect details is retained. If the defect features of the two generated samples are not significantly different, one generated sample is randomly removed or retained.
[0051] The second screening in S3 specifically includes: calculating the similarity between the generated sample that has undergone the first screening and each sample in the original sample dataset to obtain the corresponding second similarity; when the value of the second similarity is not less than a preset second threshold, the corresponding generated sample is removed.
[0052] In this embodiment, the first threshold is 0.9; the second threshold is 0.85.
[0053] like Figure 2 As shown, in this embodiment, the calculation method for the first similarity and / or the second similarity includes the following steps: S31: Extract the feature vectors of the samples to be compared, and calculate the cosine similarity between the feature vectors as the first similarity component; S32: Obtain the structural similarity index between the samples to be compared as the second similarity component; S33: The first similarity component and the second similarity component are weighted and added together to obtain the first similarity and / or the second similarity.
[0054] In this preferred embodiment, the weights of the first similarity component and the second similarity component are respectively set to 0.5.
[0055] In other embodiments, a third manual screening is also included, which involves a manual review of the generated samples that have passed the second screening. Specifically, testers verify the samples according to three core criteria: authenticity of defect features: whether the sample truly simulates defects in vehicle displays; sample quality: whether the sample is clear, without severe blurring or distortion, and without invalid or cluttered pixels; and residual repeatability: whether it still shows significant similarity to samples in the existing sample library. Samples that do not meet any of the above criteria are then removed, and all samples that meet the requirements are retained to form the final generalized sample set.
[0056] like Figure 1 As shown, in this embodiment, S4: the final generated sample is mixed with each sample in the original sample dataset to obtain a generalized sample set.
[0057] In summary, the beneficial effects of this embodiment are as follows: When generating samples, the generative model applies similarity and difference constraints. On the one hand, it ensures the structural consistency between the generated samples and the original samples, and on the other hand, it allows the generated samples to have content differences from the original samples. This ensures that the generated samples have sufficient diversity while retaining effective defect features, thereby improving the generalization ability of the sample library. It can also perform two screenings on the generated samples based on certain conditions, so that the final generated samples can better meet expectations, thereby increasing the quality of the generated samples.
[0058] During the screening process, not only can the redundancy problem within the generated samples be solved, but also the redundancy problem between the generated samples and the original samples can be solved, thereby increasing the quality and generalization ability of the generated samples.
[0059] Example 2 This embodiment is similar to Embodiment 1, and the same parts are described in Embodiment 1. The following description only focuses on the improved parts.
[0060] like Figure 3As shown, this embodiment provides a generalization system for visual defect samples in vehicle-mounted human-computer interaction software, applying the method of Embodiment 1, including: a data acquisition and annotation module, a constraint generation module, a filtering module, and a fusion module connected in sequence; the data acquisition and annotation module is used to acquire and process original sample data to form an original sample dataset with multi-dimensional feature labels; the constraint generation module is used to generate generated samples that simultaneously satisfy preset similarity constraints and difference constraints based on the original sample dataset and using a pre-trained generative model; the filtering module is used to perform a first filtering of the similarity between generated samples and a second filtering of the similarity between generated samples and the original samples in sequence to obtain the final generated samples; the fusion module is used to mix the final generated samples with the original sample dataset to obtain and output the generalized sample set.
[0061] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.
Claims
1. A method for generalizing visual defect samples in vehicle-mounted human-computer interaction software, characterized in that, Includes the following steps: S1: Obtain visual defect sample data of in-vehicle human-machine interaction software, and classify and label the sample data to obtain a raw sample dataset with multiple labels with different features. S2: For each original sample in the original sample dataset, a pre-trained generative model is used to generate corresponding generated samples according to a preset ratio, and each generated sample must simultaneously satisfy a preset similarity constraint and a preset difference constraint. S3: Based on the pre-set screening conditions, each of the generated samples is subjected to a first screening and a second screening to obtain the final generated sample; S4: Mix the final generated sample with each sample in the original sample dataset to obtain the generalized sample set.
2. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 1, characterized in that, The similarity constraint in S2 is specifically defined as follows: the similarity value between the original sample and the corresponding generated sample is not lower than a preset similarity value range.
3. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 2, characterized in that, The similarity value between the original sample and the corresponding generated sample in S2 is calculated as follows: = , in, For similarity values, For a given original sample, For the corresponding generated samples, It is a structural similarity index. For high-frequency information matching degree, For low-frequency information matching degree, The weighting coefficients are the structural similarity index, the high-frequency information matching degree, and the low-frequency information matching degree.
4. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 1, characterized in that, The specific difference constraint in S2 is that the difference value between the original sample and the corresponding generated sample is not lower than a preset difference value range.
5. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 4, characterized in that, The difference between the original sample and the corresponding generated sample in S2 is calculated as follows: ， in For the difference value, The pixel distribution difference rate This refers to the degree of difference in detail within the defective area. For the uniqueness of local features, , , These are the weighting coefficients for pixel distribution difference rate, defect area detail difference, and local feature uniqueness, respectively.
6. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 1, characterized in that, The first screening in S3 specifically includes: calculating the similarity between each pair of generated samples to obtain a first similarity; when the value of the first similarity is not less than a preset first threshold, randomly removing one of the generated samples.
7. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 6, characterized in that, The second screening in S3 specifically includes: calculating the similarity between the generated sample that has undergone the first screening and each sample in the original sample dataset to obtain the corresponding second similarity; when the value of the second similarity is not less than a preset second threshold, the corresponding generated sample is removed.
8. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 7, characterized in that, The calculation of the first similarity and / or the second similarity includes the following steps: S31: Extract the feature vectors of the samples to be compared, and calculate the cosine similarity between the feature vectors as the first similarity component; S32: Obtain the structural similarity index between the samples to be compared as the second similarity component; S33: The first similarity component and the second similarity component are weighted and added together to obtain the first similarity and / or the second similarity.
9. The method for generalizing visual defect samples in vehicle-mounted human-computer interaction software according to claim 1, characterized in that, The pre-trained generative model in S2 is specifically a stable diffusion model trained based on defect samples.
10. A generalization system for visual defect samples in vehicle-mounted human-computer interaction software, characterized in that, The method described by any one of claims 1-9 comprises: The data acquisition and annotation module, constraint generation module, filtering module, and fusion module are connected sequentially via communication. The data acquisition and labeling module is used to acquire and process raw sample data to form a raw sample dataset with multidimensional feature labels. The constraint generation module is used to generate samples that simultaneously satisfy preset similarity constraints and difference constraints based on the original sample dataset and a pre-trained generative model. The filtering module is used to sequentially perform a first filtering of the similarity between generated samples and a second filtering of the similarity between the generated samples and the original samples to obtain the final generated samples. The fusion module is used to mix the final generated samples with the original sample dataset to obtain and output the generalized sample set.