Unsupervised video abstract generation method based on comparative learning and attention-driven adversarial network

Abstract

The invention discloses an unsupervised video abstract generation method based on comparative learning and an attention-driven adversarial network. The method comprises the following steps: reading a video frame sequence, and extracting frame-level visual features by using a feature extractor; inputting the frame-level visual features into a video abstract generation model, and generating an importance score vector of each frame through the model; constructing an objective function, optimizing the selection accuracy of a frame selector, and iteratively optimizing model parameters in combination with a discriminator of the generative adversarial network, so that the video abstract generation model can effectively learn the importance and time sequence characteristics of frames from the unlabeled video; and finally, by using the video abstract generation model obtained by training, according to an importance prediction score, selecting a most representative frame, and generating a final video abstract. The method can effectively capture diversity and representativeness of video key frames, and long-distance time dependence in a modeling video sequence is achieved.

Description

Technical Field

[0001] The invention relates to an unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network, and belongs to the technical field of computer vision. Background Art

[0002] With the rapid growth of video data, video summarization, as an effective information compression and content extraction technology, has been widely used in many fields. Video surveillance, film production, sports events and other scenes generate a large amount of video data. How to quickly extract key information from it and generate concise and representative summaries has become a core challenge of current technology. Although existing video summarization methods have made some progress, there are still problems such as insufficient diversity and representativeness of capturing video key frames, and limited ability to model long-distance temporal dependencies in video sequences.

[0003] Traditional video summarization methods usually rely on manually designed features or simple temporal selection strategies, such as timestamp-based frame selection or image feature extraction. Although these methods can extract certain representative frames, they have limitations in capturing complex video content, especially when dealing with long-time videos and diverse scenes, and lack a deep understanding of video content. For example, early video summarization methods based on keyframe selection selected keyframes through simple temporal methods, but failed to effectively handle long-time dependencies and diverse content.

[0004] With the rise of deep learning, video summarization methods based on neural networks have gradually become a research hotspot. Xu et al. proposed a video summarization method based on convolutional neural networks, which can generate summaries by extracting features from video frames. However, this method relies on manually designed convolutional features and fails to deeply model the spatiotemporal context information of the video. In addition, the video summarization method based on long short-term memory networks proposed by Liu et al. solves the problem of temporal modeling, but still has the problem of insufficient multi-scale feature modeling, especially in complex video scenes, and cannot effectively capture long-distance temporal dependencies.

[0005] In recent years, generative adversarial networks, or GANs, have made significant progress in the field of image and video generation and have become a powerful tool for unsupervised learning. Yi et al. proposed a GAN-based video summarization method that automatically learns video summaries by training generators and discriminators. However, existing GAN methods still face some challenges, such as the low quality of generated video summaries and unstable training of generators and discriminators. In particular, in the generated summaries, the diversity and representativeness of the frames cannot be fully guaranteed, which affects the overall quality of the video summaries.

[0006] Currently, unsupervised video summarization research has gradually attracted attention, and many studies have begun to explore methods that combine contrastive learning and multi-head self-attention mechanisms. The video representation learning method based on contrastive learning proposed by Wu et al. enhances the distinguishability between video frames by maximizing the feature similarity of similar video frames, but it still faces the shortcomings of frame selection and summary generation in unsupervised video summarization tasks, especially in dealing with long temporal dependencies and diversity, and there is still a lack of effective solutions. Summary of the invention

[0007] In view of the problems existing in the existing methods, the purpose of the present invention is to provide an unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network, aiming to solve the problems of the existing video summary methods in capturing the diversity and representativeness of video key frames, and the limited ability to model long-distance temporal dependencies in video sequences.

[0008] The present invention provides an unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network, the method comprising the following steps: S1: read the video frame sequence and use the feature extractor to extract frame-level visual features; S2: Input the frame-level visual features into the video summary generation model, and generate an importance score vector for each frame through the model. The video summary generation model includes: Contrast frame selector: The contrast frame selector receives visual features of video frames, calculates similarities between frames through contrast learning, automatically selects the most representative frames, aggregates similar frames and pushes dissimilar frames away through contrast loss, optimizes frame selection, and ensures that the video summary contains informative and representative frames; Multi-head attention autoencoder: The multi-head attention autoencoder receives the weighted feature vectors from the comparison frame selector before and after processing, uses the multi-head attention mechanism to model frame features in parallel, captures the complex temporal relationship and multi-dimensional information between frames; reconstructs features through the decoder to extract key visual information in the video; Generative adversarial network module: The generative adversarial network module includes a generator and a discriminator. The generator generates samples based on input features to imitate the real data distribution; the discriminator determines the difference between the generated samples and the real samples. The two are mutually optimized through adversarial training. The generator continuously improves the generation quality, and the discriminator improves the ability to distinguish between real and fake samples. Importance prediction module: The importance prediction module receives frame features from the generative adversarial network, analyzes the visual information of each frame, predicts the importance score of the frame, and thus determines the importance of each frame in the video summary; S3: Construct an objective function to optimize the selection accuracy of the frame selector, and combine it with the discriminator of the generative adversarial network to iteratively optimize the model parameters so that the video summary generation model can effectively learn the importance and temporal characteristics of frames from unlabeled videos; S4: Using the video summary generation model trained in S3, perform the above S1 and S2 steps on the input video, select the most representative frames according to the importance prediction scores, and generate the final video summary.

[0009] Preferably, when calculating frame similarity, the comparison frame selector enhances the ability to cluster similar frames and separate dissimilar frames by introducing inter-frame context information, i.e., temporal relationships and global features, and utilizes the comparison loss term to maximize the similarity between similar frames and minimize the similarity between dissimilar frames, thereby improving the accuracy of frame selection and the representativeness of video summaries.

[0010] Preferably, constructing the multi-head attention autoencoder comprises: The multi-head attention mechanism is used to model the feature vectors of video frames in parallel, capturing multi-dimensional frame-to-frame interaction information and ensuring full learning of features in each dimension. After the attention weighted results of the frame features are mapped to the latent space through the encoder, the features are reconstructed through the decoder to extract and retain the key timing information and visual features in the video.

[0011] Preferably, the multi-head attention autoencoder maintains feature consistency in the feature encoding and reconstruction stages through a residual connection mechanism to avoid network degradation problems.

[0012] Preferably, constructing the generative adversarial network module includes: The generator receives the weighted features output by the comparison frame selector to generate samples that are as close to the real data distribution as possible, and optimizes the generator's loss function to make the generated summary as realistic and diverse as possible. The discriminator reversely optimizes the generation quality of the generator by identifying the distribution difference between the generated samples and the real samples; The generator and discriminator optimize each other through adversarial learning to generate high-quality and diverse video summaries.

[0013] Preferably, constructing the importance prediction module includes: Performing weighted summation on the frame features to obtain global context features; The feature scores of each frame are normalized by a normalization method to generate an importance score vector of the frame; The importance score vector is used to select the most representative key frames in the video and generate a video summary.

[0014] Preferably, the objective function includes a contrast loss term, an adversarial loss term and a reconstruction loss term, wherein: The contrastive loss term optimizes the frame selector by clustering similar frames and separating dissimilar frames; The adversarial loss term improves the realism, diversity, and quality of generated summaries through adversarial learning between the generator and the discriminator. The reconstruction loss term ensures efficient reconstruction of features, information preservation, and consistency of visual features through a multi-head attention autoencoder.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: 1. The present invention provides an unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network. It optimizes the frame selection process through contrastive learning, automatically selects the most representative video frames, and combines the multi-head attention mechanism to capture the complex temporal relationship and multi-dimensional information between frames, which significantly improves the information content and representativeness of the video summary.

[0016] 2. The present invention proposes a contrast frame selector, which enhances the clustering ability of similar frames and the separation ability of dissimilar frames by introducing inter-frame context information, thereby improving the accuracy and robustness of frame selection and ensuring that the generated summary contains more informative key frames.

[0017] 3. The present invention designs a multi-head attention autoencoder, which models frame features in parallel through a multi-head attention mechanism, captures multi-dimensional inter-frame interaction information, and reconstructs video frame features in combination with an encoder-decoder structure, effectively retaining the key visual information and temporal features in the video, and improving the temporal consistency and content richness of the summary.

[0018] 4. The present invention is experimentally verified on two benchmark datasets SumMe and TVSum. The experimental results show that compared with the existing most advanced methods, the method proposed in the present invention has significantly improved the performance of video summary generation, proving its superiority in terms of information content, representativeness and temporal consistency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 A flowchart of the implementation of the unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network provided by the present invention; Figure 2 The overall framework diagram of the unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network provided by an embodiment of the present invention; Figure 3A schematic diagram of a multi-head attention autoencoder provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0021] like Figure 1 As shown, it is a flowchart of the implementation of the unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network provided by the present invention. This embodiment provides an unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network, including the following steps: S1: read the video frame sequence and use the feature extractor to extract frame-level visual features; S2: Input the frame-level visual features into the video summary generation model, and use the model to generate the importance score vector of each frame; S3: Construct an objective function to optimize the selection accuracy of the frame selector, and combine it with the discriminator of the generative adversarial network to iteratively optimize the model parameters so that the video summary generation model can effectively learn the importance and temporal characteristics of frames from unlabeled videos; S4: Using the video summary generation model trained in S3, perform the above S1 and S2 steps on the input video, select the most representative frames according to the importance prediction scores, and generate the final video summary.

[0022] Embodiment 1: The present invention provides a preferred embodiment to perform S1, read the video frame sequence, and use the feature extractor to extract frame-level visual features. The present invention can process videos of any length and any type of video, such as surveillance, movies, etc. The specific steps are as follows: First, the original video is pre-sampled at a sampling rate of 2FPS to obtain a video frame sequence V = {v1,v2,..., v n}, where v i represents the i-th video frame, and n represents the total number of video frames. Then, the GoogLeNet network pre-trained on the ImageNet dataset is used as a feature extractor to extract visual features for each frame. Specifically, the 1024-dimensional feature vector output by the pool5 layer of the GoogLeNet network is used as the feature representation of each frame, denoted as F = {f1,f2,..., f n}, where f iRepresents the feature vector of the i-th frame, which is used to characterize the visual content in the frame.

[0023] Embodiment 2: The present invention provides a preferred embodiment to perform S2, inputting the frame-level visual features into the video summary generation model, and generating an importance score vector for each frame through the model.

[0024] like Figure 2 As shown in the figure, it is the overall framework diagram of the network model of this embodiment. The network consists of 4 parts: contrast frame selector, multi-head attention autoencoder, generative adversarial network module and importance prediction module. The specific construction steps of each module will be described below: S21, build a comparison frame selector. Specifically, first input the frame features {f1,f2,..., f n}, these features come from the pre-trained GoogLeNet network. Then, by comparing the learning framework, the feature f of each frame is calculated i With other frame features f j The similarity between the two frames is calculated by using the contrast loss function to optimize the frame selection. The goal of contrastive learning is to maximize the similarity between similar frames and minimize the similarity between dissimilar frames, so as to select the most representative frames. The specific loss function is as follows: , Among them, d i is the Euclidean distance between frames i, y i is the label, y i = 1 means the frame pair is similar, y i = 0 means the frame pair is not similar, and is a hyperparameter that controls the distance threshold between similar and dissimilar frames. By minimizing this loss function, the frame selector can optimize the selection of video frames so that the summary can contain informative and representative frames.

[0025] S22, build a multi-head attention autoencoder, as shown in the following diagram Figure 3 As shown in Figure 2. In this step, the multi-head self-attention mechanism is first used to model the features of the video frames in parallel to capture multi-dimensional temporal information and spatial interactions. The multi-head attention mechanism is to weight the input features through multiple attention heads to enhance the model's learning ability for complex spatiotemporal dependencies. Each attention head processes the input features in a different subspace, allowing the model to focus on the relationship between frames in different dimensions, thereby improving the modeling ability of temporal information.

[0026] Specifically, for each frame, the feature representation f i, we calculate the attention distribution through the query, key, and value matrices. Through the multi-head self-attention mechanism, the relationship between frames can be captured from multiple angles, and the model can simultaneously focus on the important features of different frames in multiple subspaces in a weighted manner. The calculation formula of the attention mechanism is as follows: , Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively. k is the dimension of the key vector. Each attention head calculates these weighted sums independently, and then concatenates the outputs of multiple heads to get the final multi-head attention result.

[0027] After that, the attention-weighted frame feature vector is input into the encoder for mapping into the latent space. The encoder maps the frame features into the latent space through linear transformation and retains the key timing information and visual features of the video. In this process, in order to avoid information loss and feature degradation, a residual connection mechanism is usually used so that the input information can be effectively transmitted through the network layer, maintaining feature consistency and ensuring that the network can be effectively trained.

[0028] S23, constructing a generative adversarial network module. The generative adversarial network module is the core part of the present invention, and its purpose is to generate high-quality and diverse video summaries through adversarial training between the generator and the discriminator. The module consists of two parts, the generator and the discriminator, which optimize each other through adversarial learning.

[0029] Generator: The main task of the generator is to generate virtual summary samples that are as close to the real video summary as possible based on the input frame features. Specifically, the generator receives the weighted frame features output by the contrast frame selector as input, and generates a video summary after processing by a multi-layer fully connected network or a convolutional neural network. The goal of the generator is to imitate the real video summary distribution as much as possible, so that the generated summary has a high degree of authenticity and information content. The optimization goal of the generator is to maximize the judgment value of the discriminator on its generated samples, that is: , Among them, G(z) represents the sample generated by the generator, and D(·) is the output of the discriminator, which represents the probability similarity between the generated sample and the real data.

[0030] Discriminator: The task of the discriminator is to distinguish the difference between the virtual summary generated by the generator and the real summary. The discriminator receives the real summary and the generated summary, calculates their probability values, and outputs a score on whether the sample is real. The optimization goal of the discriminator is to minimize the logarithmic loss of its output, that is: , in, is the real video summary, G(z) is the summary generated by the generator. The discriminator is continuously optimized to enhance its ability to distinguish generated samples from real samples.

[0031] Adversarial learning: The generator and the discriminator optimize each other through adversarial learning. The generator generates more realistic summaries by minimizing the discriminator's loss, while the discriminator improves its ability to distinguish between generated summaries and real summaries by minimizing its own loss. The two compete with each other, and ultimately the generator can generate video summaries that are close to reality, while the discriminator can more accurately distinguish between real and fake samples.

[0032] S24, construct an importance prediction module. The importance prediction module is used to evaluate the importance of each frame in the video summary based on the representation of frame features, and then select the most representative key frame to generate a summary. This module includes three key steps: Weighted summation to obtain global context features: In order to ensure that the context information of the frame can be fully considered in the process of generating the summary, the importance prediction module first performs a weighted summation of the features of all frames to extract the global context information of the video. In this way, the temporal information of the entire video can be integrated to ensure that the selection of each frame is not only based on its own features, but also takes into account the relationship with other frames.

[0033] Normalization: Normalize the feature scores of each frame to ensure that the scores of different frames are compared under the same standard. This process can eliminate the differences in scales of features of different frames, making the score calculation more fair and stable. Specifically, standardization or maximum and minimum normalization methods are used to make the scores of each frame fall within a uniform range, which is convenient for subsequent key frame selection.

[0034] Key frame selection: Based on the normalized score of each frame, select the frame with the highest score as the key frame of the video summary. In this way, it can ensure that the video summary contains the most informative and representative frames, thereby improving the quality and diversity of the video summary.

[0035] Embodiment three: Based on the above embodiments, the present invention provides a preferred embodiment to execute S3, construct an objective function, optimize the selection accuracy of the frame selector, and combine the discriminator of the generative adversarial network to iteratively optimize the model parameters, so that the video summary generation model can effectively learn the importance and timing characteristics of frames from unlabeled videos.

[0036] The first step is to construct a contrast loss function. In order to optimize the selection accuracy of the frame selector, this embodiment introduces a contrast learning framework. By constructing positive and negative sample pairs, the model can better distinguish key frames from non-key frames. Specifically, the feature representation of the input video frame sequence is F = {f1, f2,..., f n}, where f i Represents the feature vector of the i-th frame. Define the key frame set F + As positive samples, the non-key frame set F − As negative samples, the similarity between frames is calculated. A loss function based on contrastive learning is used to reduce the feature distance between similar frames and increase the feature distance between dissimilar frames. The loss function is as follows: , where d(·,·) represents the similarity measure between frame features, and y i is the identifier of the frame pair, y i = 1 indicates a positive sample pair, y i = 0 indicates a negative sample pair, is the set threshold. By optimizing this loss function, the model can learn the deep relationship between frames, thereby improving the accuracy of the frame selector in the summary generation process.

[0037] The second step is to optimize the generative adversarial network. In this embodiment, the generator G of the GAN is responsible for generating a video summary based on the output of the frame selector, while the discriminator D is used to distinguish the difference between the generated summary and the real video frame set. The goal of the generator is to maximize the judgment error of the discriminator so that the generated video summary is closer to the real data distribution.

[0038] The third step is joint training and optimization. By jointly optimizing the contrast loss and adversarial loss, the optimization objective function of this embodiment is as follows: , Among them, λ1,λ2,λ3 are weight hyperparameters used to balance the contribution of different loss terms. During the training process, contrastive learning is first used to optimize the frame selector so that it can accurately distinguish key frames and non-key frames; then, adversarial training is performed under the GAN framework to make the generated video summary more representative in terms of both vision and content.

[0039] Through the above optimization process, the method of this embodiment can effectively learn the importance and temporal characteristics of video frames in an unsupervised environment, ensure that the generated video summary has high diversity, representativeness and temporal consistency, and demonstrate excellent summary generation performance on multiple video datasets.

[0040] In order to verify the effectiveness of the above embodiments, the present invention is applied in practice, and the F score (%) is calculated to compare and analyze with the existing advanced methods. Experiments are conducted on two benchmark datasets, SumMe and TVSum, to evaluate the performance of the present invention in terms of representativeness, diversity and temporal consistency of key frame selection, and verify its advantages in the unsupervised video summarization task. SumMe is a popular video summary dataset containing 25 videos covering different types of scenes, and each video is provided with frame-level importance annotations by at least 15 users. This dataset is used to evaluate the similarity between the summaries generated by the model and the manual summaries. TVSum consists of 50 videos from YouTube, covering 10 different categories (such as news, documentaries, travel, etc.), and each video is provided with importance annotations by 20 users. This dataset is suitable for testing the generalization ability of the model under different video types.

[0041] The present invention uses the standard 5-fold cross validation (5FCV) to conduct experiments. All videos are evenly divided into 5 subsets. Each time, 4 groups are selected as training sets and 1 group is selected as test sets. Finally, the average value of 5 experiments is taken as the final result. This method can effectively reduce the result deviation caused by random data division and ensure the fairness and stability of the experiment.

[0042] This experiment selects a variety of mainstream unsupervised video summarization methods for comparison, including: SUM-GAN, DR-DSN, SUM-FCN, VASNet, AC-SUM-GAN. Under the standard evaluation setting, the experimental results are shown in Table 1. The proposed method achieves the best summary performance on all datasets.

[0043]

[0044] Experimental results show that the proposed method is better than the existing best methods on the SumMe and TVSum datasets, and has achieved significant improvements in summary quality and temporal consistency. This is because the proposed method uses contrastive learning to optimize the frame selector, which enables the model to more accurately distinguish key frames from non-key frames, and improves the model's ability to model long-term dependencies through a multi-head self-attention mechanism. In addition, compared with methods such as VASNet and DR-DSN, the proposed method combines generative adversarial networks to ensure the diversity and coherence of summaries.

Claims

1. An unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network, characterized in that: The following steps are involved: S1: read the video frame sequence and use the feature extractor to extract frame-level visual features; S2: Input the frame-level visual features into the video summary generation model, and generate an importance score vector for each frame through the model. The video summary generation model includes: Contrast frame selector: The contrast frame selector receives visual features of video frames, calculates similarities between frames through contrast learning, automatically selects the most representative frames, aggregates similar frames and pushes dissimilar frames away through contrast loss, optimizes frame selection, and ensures that the video summary contains informative and representative frames; Multi-head attention autoencoder: The multi-head attention autoencoder receives the weighted feature vectors from the comparison frame selector before and after processing, uses the multi-head attention mechanism to model frame features in parallel, captures the complex temporal relationship and multi-dimensional information between frames; reconstructs features through the decoder to extract key visual information in the video; Generative adversarial network module: The generative adversarial network module includes a generator and a discriminator. The generator generates samples based on input features to imitate the real data distribution; the discriminator determines the difference between the generated samples and the real samples. The two are mutually optimized through adversarial training. The generator continuously improves the generation quality, and the discriminator improves the ability to distinguish between real and fake samples. Importance prediction module: The importance prediction module receives frame features from the generative adversarial network, analyzes the visual information of each frame, predicts the importance score of the frame, and thus determines the importance of each frame in the video summary; S3: Construct an objective function to optimize the selection accuracy of the frame selector, and combine it with the discriminator of the generative adversarial network to iteratively optimize the model parameters so that the video summary generation model can effectively learn the importance and temporal characteristics of frames from unlabeled videos; S4: Using the video summary generation model trained in S3, perform the above S1 and S2 steps on the input video, select the most representative frames according to the importance prediction scores, and generate the final video summary.

2. The unsupervised video summarization method based on contrastive learning and attention-driven adversarial network as claimed in claim 1, characterized in that: When calculating frame similarity, the comparison frame selector introduces inter-frame context information, namely, temporal relations and global features, to enhance the ability to cluster similar frames and separate dissimilar frames, and utilizes the comparison loss term to maximize the similarity between similar frames and minimize the similarity between dissimilar frames.

3. The unsupervised video summarization method based on contrastive learning and attention-driven adversarial network as claimed in claim 1, characterized in that: Construct the multi-head attention autoencoder, including: The multi-head attention mechanism is used to model the feature vectors of video frames in parallel, capturing multi-dimensional frame-to-frame interaction information and ensuring full learning of features in each dimension. After the attention weighted results of the frame features are mapped to the latent space through the encoder, the features are reconstructed through the decoder to extract and retain the key timing information and visual features in the video.

4. The unsupervised video summarization method based on contrastive learning and attention-driven adversarial network as claimed in claim 1, characterized in that: The multi-head attention autoencoder maintains feature consistency in the feature encoding and reconstruction stages through a residual connection mechanism, thereby avoiding the problem of network degradation.

5. The unsupervised video summarization method based on contrastive learning and attention-driven adversarial network as claimed in claim 1, characterized in that: Constructing the generative adversarial network module includes: The generator receives the weighted features output by the comparison frame selector to generate samples that are as close to the real data distribution as possible, and optimizes the generator's loss function to make the generated summary as realistic and diverse as possible. The discriminator reversely optimizes the generation quality of the generator by identifying the distribution difference between the generated samples and the real samples; The generator and discriminator optimize each other through adversarial learning to generate high-quality and diverse video summaries.

6. The unsupervised video summary generation method based on contrastive learning and attention-driven adversarial network as claimed in claim 1, characterized in that: Constructing the importance prediction module includes: Performing weighted summation on the frame features to obtain global context features; The feature scores of each frame are normalized by a normalization method to generate an importance score vector of the frame; The importance score vector is used to select the most representative key frames in the video and generate a video summary.

7. The unsupervised video summarization method based on contrastive learning and attention-driven adversarial network as claimed in claim 1, characterized in that: The objective function includes a contrast loss term, an adversarial loss term, and a reconstruction loss term, where: The contrastive loss term optimizes the frame selector by clustering similar frames and separating dissimilar frames; The adversarial loss term improves the realism, diversity, and quality of generated summaries through adversarial learning between the generator and the discriminator. The reconstruction loss term ensures efficient reconstruction of features, information preservation, and consistency of visual features through a multi-head attention autoencoder.

Images