Efficient target segment retrieval and localization method and system based on generative augmentation

Abstract

The application discloses a high-efficiency target segment retrieval and positioning method and system based on generative enhancement, adopts a bidirectional Mamba architecture based on a state space model to extract long-distance time sequence dependence of a video, reduces the calculation complexity to a linear level, introduces a generative query reconstruction task as a semantic regularization mechanism, promotes the model to realize deeper and more fine-grained cross-modal semantic alignment, and the lightweight decoder is only enabled in the training stage, which greatly improves the retrieval accuracy of the system without increasing any additional inference overhead; an explicit boundary perception and time sequence classification positioning strategy is adopted, the accurate transition of the action context is sharply captured by expanding the time sequence receptive field, irrelevant background noise interference is effectively filtered out, and high-precision frame-level segment positioning is realized. While realizing high-efficiency and high-precision retrieval, the application is highly automated in operation control, and can be seamlessly deployed in large-scale multimedia data processing scenes.

Description

Technical Field

[0001] This invention relates to a method and system for efficient target fragment retrieval and localization based on generative enhancement, belonging to the interdisciplinary field of multimedia information retrieval and artificial intelligence. Background Technology

[0002] With the explosive growth of multimedia data, video has become a major medium for information dissemination. Video clip retrieval requires models to efficiently retrieve relevant videos from massive unedited video libraries based on natural language queries and accurately locate specific time segments. This technology has enormous application potential in video search engines, intelligent surveillance, and multimedia content analysis. Currently, existing video clip retrieval and localization technologies mainly follow the "discriminative matching" paradigm. Its main structure and principle are typically as follows: adopting a "coarse-to-fine" pipeline, firstly, pre-trained visual and text encoders are used to extract single-modal features, then network structures such as Transformer are used for temporal dependency modeling and cross-modal feature interaction, and finally, in the localization stage, clip localization is usually treated as a regression task or a heuristic threshold setting method is used to output the final start and end times.

[0003] Meanwhile, when faced with massive, long-duration, and complex unedited video libraries, existing retrieval and localization methods still suffer from the following significant problems and drawbacks: 1) Long video processing is inefficient and computationally expensive. Most existing methods rely on Transformer encoders for temporal dependency modeling, and the computational complexity of their self-attention mechanism is quadratic with the length of the input video sequence, i.e. 1) When processing unedited long videos containing a large number of frames, this architecture consumes enormous computational and memory resources, leading to severe efficiency bottlenecks and making it difficult to meet the scalability requirements of massive video library retrieval. 2) It is prone to getting stuck in shallow feature matching and lacks deep, fine-grained semantic alignment capabilities. Existing discriminative matching mechanisms lack an intrinsic mechanism for verifying dense correspondences, causing models to often seek learning shortcuts, relying only on salient static objects in the video for surface feature matching, while ignoring the deep understanding of the temporal evolution of actions. This mechanism makes the model prone to attention drift when faced with difficult-to-bear samples with similar backgrounds but subtle differences in action semantics, and unable to accurately distinguish subtle differences. 3) The segment boundary localization is ambiguous and easily affected by noise. Existing regression-based localization methods struggle to handle the non-continuous time spans of complex actions, while threshold-based filtering strategies are easily affected by video background noise. Due to the lack of explicit boundary constraints on the temporal context transition, these methods ultimately result in significant deviations and ambiguities in the predicted segment start and end times.

[0004] Therefore, in response to the retrieval needs of massive unedited video libraries, there is an urgent need to establish a video library segment retrieval method with low computational overhead, deep semantic alignment capabilities, and accurate boundary positioning. Summary of the Invention

[0005] Objective: To address the problems and shortcomings of existing technologies, this invention provides a highly efficient target segment retrieval and localization method and system based on generative enhancement. Specifically, addressing the issue that existing methods using the Transformer architecture for temporal modeling result in quadratic computational complexity and severe memory consumption, leading to low efficiency in retrieving massive amounts of unedited long videos, this invention aims to provide a long temporal feature extraction scheme with linear complexity to overcome the efficiency bottleneck. Simultaneously, addressing the problems of existing discriminative retrieval models being prone to shallow visual feature matching, attention drift, and lacking deep, fine-grained semantic alignment capabilities, this invention introduces a generative deep semantic reconstruction mechanism as a regularization method, enabling the model to accurately understand the underlying temporal action semantics of the video and achieve fine alignment of cross-modal features. Furthermore, addressing the problem that existing regression- or threshold-based localization methods lack explicit constraints on temporal context transitions, resulting in ambiguous start and end time boundaries and susceptibility to irrelevant background noise, this invention provides a clear boundary awareness and classification localization mechanism to ensure accurate frame-level temporal boundary localization in complex video streams.

[0006] Technical solution: The core design concept of this invention is to transform the video library segment retrieval task from the traditional "shallow feature matching" to "generative semantic understanding". By introducing an autoregressive text reconstruction task during the training phase, the model is forced to focus on deep visual cues that are truly relevant to the query. At the same time, the decoder is discarded during the inference phase to ensure retrieval efficiency.

[0007] An efficient target fragment retrieval and localization method based on generative augmentation is proposed. A retrieval and localization model is trained to achieve target fragment retrieval and localization. The training process of the retrieval and localization model specifically includes the following steps: Step 1, Multimodal Feature Extraction: Obtain training set data, which includes pairs of natural language queries, corresponding video sequences, and the actual start and end time boundary labels of the queries in the videos; use a pre-trained text encoder to extract word vector feature sequences of the queries, use a pre-trained visual encoder to extract frame-level visual feature sequences of the videos at a fixed frame rate, and map the word vector feature sequences and frame-level visual feature sequences to a unified shared implicit dimension space by setting a modality linear projection layer.

[0008] Step 2, Bidirectional Mamba Temporal Encoding: The mapped video frame sequence is input into a bidirectional Mamba network based on a state-space model; the bidirectional Mamba network scans the video frame features along both forward and backward time dimensions, and the bidirectional outputs are concatenated and fused through a linear layer to generate a video feature sequence with global temporal context awareness; at the same time, the Transformer layer is used to refine the word vector feature sequence of the mapped query to generate a query feature sequence.

[0009] Step 3, Coarse-grained retrieval and distribution matching based on soft alignment: Select the Top-K videos and perform coarse-grained retrieval; calculate the soft alignment score between the query feature sequence and the context-aware video feature sequence within the batch, i.e., the global semantic similarity score; based on this, abandon the traditional binary contrast loss and adopt a distribution matching scheme based on Kullback-Leibler (KL) divergence to calculate the video retrieval loss. Specifically, the predicted matching distribution is calculated using a Softmax function scaled by a temperature coefficient. At the same time, a label smoothing technique is introduced to construct a smooth real target distribution. The robustness of the embedding space of the bidirectional Mamba and Transformer models is improved by minimizing the KL divergence between the predicted distribution and the target distribution.

[0010] Step 4, Cross-modal attention calculation: Input the context-aware video feature sequence and the corresponding query feature sequence into the multimodal Transformer fusion encoder, use the cross attention layer to calculate the relevance affinity matrix of each video frame to the text query, and aggregate along the text dimension to generate frame-level relevance attention weight sequence and cross-modal fusion features.

[0011] Step 5, Generative Feature Reconstruction and Semantic Regularization: A lightweight autoregressive decoder is connected after the fusion encoder; this decoder uses a one-way causal masking mechanism, taking the cross-modal fusion features output in Step 4 as input, and reconstructs the input natural language query text word by word; the negative log-likelihood loss between the reconstructed text and the real text is calculated to obtain the generation loss. This step, semantic regularization, forces the fusion encoder to precisely focus on the relevant visual semantic regions, making the reconstructed text as similar as possible to the real text, thereby implicitly calibrating the cross-attention weights in step 4.

[0012] Step 6, Boundary Probability Decoupling and Classification Supervision: The relevance attention weight sequence output from Step 4 is input into the boundary-aware localization module. This module contains stacked one-dimensional convolutional layers. By expanding the temporal receptive field, it captures the transitional changes in the local context, decouples the attention weights, and outputs the initial boundary probability distribution sequence and the final boundary probability distribution sequence through the Softmax function. The start and end times of the ground truth annotations are converted into one-hot hard labels. The standard cross-entropy loss is used to maximize the log-likelihood of the ground truth boundary frames to obtain the localization loss. .

[0013] Step 7, Joint Optimization and Parameter Update: Calculate the video retrieval loss obtained in the above steps. , generation loss and positioning loss By performing weighted summation, we obtain the final overall training loss function for the retrieval and localization model. Its definition is: (1) in, and These are preset hyperparameters used to balance the weights of different tasks. The gradient is calculated using the backpropagation algorithm, and the network weights of the entire system are updated until the retrieval and localization models converge, completing the training.

[0014] Furthermore, when performing actual inference retrieval after completing model training, the method of the present invention specifically includes the following steps: The autoregressive decoder used in the training phase is removed; the user-input query and the video library to be retrieved are acquired in real time, and the above feature extraction and Mamba encoding processes are repeated; the global semantic similarity score is calculated through the coarse-grained retrieval module, and the Top-K candidate videos are selected; the candidate videos and query features are input into the fusion encoder and boundary-aware localization module, and high-quality start / end boundary probabilities are directly output (without generative decoding computational overhead); the valid time interval candidate segments in the video are traversed. The initial boundary probability is multiplied by the termination boundary probability to obtain the local confidence score; finally, the global semantic similarity score and the local confidence score are geometrically weighted and fused according to formula (2): (2) in, A weighting factor to balance relevance and confidence level; The similarity scores between the Top-K candidate videos and the query. The first in the video Confidence scores for each candidate interval; output the total score. The highest video ID and its corresponding precise start and end time boundaries.

[0015] A high-efficiency target fragment retrieval and localization system based on generative enhancement is disclosed. The system mainly comprises four logical modules: a multimodal feature extraction module, a coarse-grained video retrieval module based on bidirectional Mamba, a generative enhancement fusion module, and a boundary-aware localization module. The multimodal feature extraction module acquires initial independent features from video sequences and text queries. The coarse-grained video retrieval module based on bidirectional Mamba scans video frame features along both forward and reverse time dimensions, concatenates the bidirectional outputs, and fuses them through a linear layer to generate a video feature sequence with global temporal context awareness, used to capture long-distance temporal dependencies in videos and recall candidate videos. The generative enhancement fusion module adopts an encoder-decoder architecture, with the decoder only enabled during the training phase. It uses the output cross-modal fusion features as input to reconstruct the input natural language query text word-by-word, forcing the fusion encoder to precisely focus on relevant visual semantic regions, implicitly calibrating cross-modal attention as a semantic regularizer. The boundary-aware localization module decouples the calibrated attention into the start and end boundary probabilities of the fragments. This module contains stacked one-dimensional convolutional layers that capture transitional changes in local context by expanding the temporal receptive field, decouple attention weights, and output the start and end boundary probability distribution sequences through the Softmax function.

[0016] The implementation process and methods of the system are the same and will not be described again.

[0017] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the efficient target fragment retrieval and localization method based on generative enhancement as described above.

[0018] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the efficient target fragment retrieval and localization method based on generative enhancement as described above.

[0019] Compared with existing technologies, this invention provides a highly efficient target segment retrieval and localization method and system based on generative enhancement. Its significant advantages are as follows: First, this invention employs a bidirectional Mamba architecture based on a state-space model to extract long-distance temporal dependencies in videos, reducing computational complexity to a linear level and significantly saving computational and memory resources, thereby significantly breaking through the retrieval efficiency bottleneck of massive amounts of unedited long videos. Second, this invention innovatively introduces a generative query reconstruction task as a semantic regularization mechanism, enabling the model to achieve deeper and finer-grained cross-modal semantic alignment. Furthermore, this lightweight decoder is only enabled during the training phase, significantly improving the system's retrieval accuracy without adding any additional inference overhead. Finally, this invention adopts a clear boundary awareness and temporal classification localization strategy, which expands the temporal receptive field to keenly capture the precise transitions of action context, effectively filtering out irrelevant background noise interference, solving the problem of blurred localization boundaries in complex video streams, and achieving high-precision frame-level segment localization. In summary, this invention achieves high-efficiency and high-precision retrieval while maintaining highly automated operation and control. It can be seamlessly deployed in large-scale multimedia data processing scenarios and has extremely high industrial value and broad application prospects in fields such as intelligent video search engines, smart security monitoring, and multimedia large-scale model content analysis. Attached Figure Description

[0020] Figure 1 This is an overall framework diagram of efficient target fragment retrieval and localization based on generative enhancement, according to an embodiment of the present invention. Figure 2 This is a flowchart of the training and inference method for generative augmented video segment retrieval according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0022] This invention provides a highly efficient target segment retrieval and localization method and system based on generative enhancement. The method can efficiently retrieve target videos from a massive, unedited long video library based on user-input natural language queries, and accurately locate the precise start and end timestamps of actions. By introducing a bidirectional Mamba architecture with linear complexity and a generative semantic reconstruction mechanism, this invention effectively overcomes the shortcomings of traditional methods, such as high computational overhead and susceptibility to shallow visual feature matching. Figure 1The diagram illustrates the overall framework for efficient target fragment retrieval and localization based on generative enhancement as described in this invention. The system in this invention mainly comprises four core logical modules: a multimodal feature extraction module, a coarse-grained video retrieval module based on bidirectional Mamba, a generative enhancement fusion module, and a boundary-aware localization module. Their specific composition and structural functions are described below: Multimodal Feature Extraction Module: This module uses a pre-trained visual encoder, such as ResNet-152, to extract frame-level feature sequences from the video. Simultaneously, it uses a pre-trained text encoder, such as RoBERTa, to extract word vector feature sequences from the query, and maps them to a unified dimensional space through linear projection. Coarse-grained Video Retrieval Module Based on Bidirectional Mamba: This module leverages the linear complexity advantage of the state-space model, scanning video frame features using both forward and backward time dimensions, and fusing them to generate a video feature sequence with global temporal context awareness. This module is used to calculate coarse-grained similarity and quickly recall videos. Generative Enhancement Fusion Module: This module uses a multimodal Transformer fusion encoder to calculate cross-modal cross-attention. More importantly, this module is followed by a lightweight autoregressive decoder that is only enabled during training to reconstruct the original text query, implicitly calibrate the attention weights, and filter out background noise. The boundary-aware localization module uses a stacked one-dimensional convolutional network to decouple the calibrated attention weights into independent start and end boundary probability distributions.

[0023] like Figure 2 As shown, the efficient target fragment retrieval and localization method based on generative augmentation includes training and inference phases. To make the concept of this invention clearer, the "model training phase" and "actual inference phase" of the system are described in detail here. The model training phase process is described as follows: Step 201, Input training set: The system obtains training data containing pairs of "natural language queries", "corresponding unedited video sequences", and "real time boundary annotations".

[0024] Step 202, Multimodal Feature Extraction and Encoding: The textual and visual features of the shared space are obtained through the multimodal feature extraction module described above; then the video frame sequence is input into the bidirectional Mamba network for temporal context encoding.

[0025] Step 203, coarse-grained retrieval and distribution matching training: Calculate the soft alignment score between the query sequence and the video sequence, i.e., the global semantic similarity; the system abandons the binary contrast loss and instead constructs the true target distribution by introducing label smoothing technology, calculates the Kullback-Leibler (KL) divergence between the predicted distribution and the target distribution, and derives the video retrieval loss. .

[0026] Step 204, cross-modal feature fusion: Input the context-aware video features and query features into the Transformer fusion encoder, and output the frame-level relevance attention weight sequence and fused features.

[0027] Step 205, Generate Enhanced Regularization Training: Input the fused features into the autoregressive decoder, which employs a causal masking mechanism to reconstruct the original natural language query word-by-word. Calculate the negative log-likelihood loss between the reconstructed text and the original text. This loss forces the fusion encoder to focus on semantically relevant visual cues in case the decoder lacks effective information.

[0028] Step 206, Boundary Localization and Classification Training: Input the attention weights into a one-dimensional convolutional network, decoupling them into start and end probabilities for each frame. Calculate the classification cross-entropy localization loss using real start and end one-hot hard labels. .

[0029] Step 207, Joint Optimization: The system combines the three loss functions according to the formula... Perform a weighted summation and update all network weights through backpropagation until the model converges.

[0030] The actual reasoning and retrieval process is described as follows: Step 301, Decoding and Real-Time Input: The system completely removes / discards the autoregressive decoder module from the training phase to eliminate computational overhead. It then receives real-time user input of query text and massive video library data.

[0031] Step 302, Feature Extraction and Mamba Encoding: Similar to the training phase, extract the features of the real-time input and perform bidirectional Mamba encoding to obtain context-aware video features.

[0032] Step 303, coarse-grained rapid recall: calculate the global semantic similarity score, and at this stage, select only the Top-K (e.g., K=100) most relevant candidate videos.

[0033] Step 304, Refined cross-modal attention extraction: Input the Top-K candidate videos and query features into the fusion encoder, and directly output a high-quality attention weight sequence that has been implicitly calibrated.

[0034] Step 305, Boundary Probability Acquisition: Decouple attention weights through a one-dimensional convolution module to obtain the precise start and end probabilities of candidate frames.

[0035] Step 306, Joint Likelihood Search: Globally traverse candidate segments of valid time intervals , will the The starting probability of a frame and the first The local confidence score is obtained by multiplying the frame termination probabilities.

[0036] Step 307, Score Fusion Decision: The global semantic similarity in step 303 and the local confidence in step 306 are geometrically weighted and fused to output the video ID with the highest final score and its precise start and end timestamps.

[0037] In summary, this invention provides a highly efficient target fragment retrieval and localization method and system based on generative enhancement. This technology can be applied to many fields such as multimedia content analysis, intelligent video search engines, and smart security monitoring playback. For example, in the field of smart security, security personnel only need to input natural language describing a specific abnormal event, and the system can leverage the linear processing advantages of the Mamba architecture to find the corresponding video from massive amounts of surveillance footage with extremely low latency. Furthermore, through a boundary awareness module calibrated by generative regularization, it can accurately extract a few seconds of surveillance footage showing the event. Because this technology maintains or even improves retrieval accuracy while incurring zero additional computational overhead during the inference phase, it significantly reduces hardware resource consumption and thus has extremely high industrialization value. This invention provides a specific implementation method and system idea for video fragment retrieval based on generative enhancement. There are many software and hardware approaches to implement this technical solution; the above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principles of this invention, such as adjusting the number of Mamba network layers, modifying the hyperparameters of the loss weights, or replacing the basic feature extraction network. These improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing computer vision and natural language processing technologies.

[0038] Obviously, those skilled in the art should understand that the steps of the efficient target fragment retrieval and localization method based on generative enhancement or the modules of the efficient target fragment retrieval and localization system based on generative enhancement in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. A highly efficient target fragment retrieval and localization method based on generative augmentation, characterized in that, Training a retrieval and localization model to achieve target fragment retrieval and localization. The training process of the retrieval and localization model specifically includes the following steps: Step 1: Obtain training set data. The training set includes pairs of natural language queries, corresponding video sequences, and the actual start and end time boundary labels of the queries in the videos. Use a pre-trained text encoder to extract word vector feature sequences of the queries, and use a pre-trained visual encoder to extract frame-level visual feature sequences of the videos at a fixed frame rate. Map the word vector feature sequences and frame-level visual feature sequences to a unified shared implicit dimension space. Step 2: Input the mapped video frame sequence into a bidirectional Mamba network based on a state-space model; and concatenate the bidirectional outputs and fuse them through a linear layer to generate a video feature sequence; at the same time, use a Transformer layer to refine the mapped query word vector feature sequence to generate a query feature sequence. Step 3: Calculate the soft alignment score between the query feature sequence and the video feature sequence within the batch, i.e., the global semantic similarity score; calculate the video retrieval loss using a distribution matching scheme based on KL divergence. ; Step 4: Input the video feature sequence and the corresponding query feature sequence into the multimodal Transformer fusion encoder, use the cross attention layer to calculate the relevance affinity matrix of each video frame to the text query, and aggregate along the text dimension to generate frame-level relevance attention weight sequence and cross-modal fusion features. Step 5: An autoregressive decoder is connected after the fusion encoder. This decoder uses a one-way causal masking mechanism, taking cross-modal fusion features as input, and reconstructs the input natural language query text word by word. The negative log-likelihood loss between the reconstructed text and the real text is calculated to obtain the generation loss. ; Step 6: Input the relevance attention weight sequence into the boundary-aware localization module, decouple the attention weights, and output the start and end boundary probability distribution sequences through the Softmax function; convert the start and end times of the ground truth annotations into one-hot hard labels, and maximize the log-likelihood of the ground truth boundary frames using the standard cross-entropy loss to obtain the localization loss. ; Step 7, reduce video retrieval loss , generation loss and positioning loss By performing weighted summation, we obtain the final overall training loss function for the retrieval and localization model. Its definition is: (1) in, and The hyperparameters are preset; the gradient is calculated and the network weights of the retrieval and localization model are updated through the backpropagation algorithm until the retrieval and localization model converges, thus completing the training.

2. The efficient target fragment retrieval and localization method based on generative enhancement according to claim 1, characterized in that, Step 3 uses a distribution matching scheme based on KL divergence to calculate the video retrieval loss. Specifically, the predicted matching distribution is calculated using a Softmax function scaled by a temperature coefficient. At the same time, a label smoothing technique is introduced to construct a smooth real target distribution. The robustness of the model's embedding space is improved by minimizing the KL divergence between the predicted distribution and the target distribution.

3. The efficient target fragment retrieval and localization method based on generative enhancement according to claim 1, characterized in that, When performing actual inference retrieval after model training is completed, the method executes the following process: The autoregressive decoder used in the training phase is removed; the user-input query and the video library to be retrieved are acquired in real time, and steps 1-2 are executed; the global semantic similarity score is calculated through the coarse-grained retrieval module, and the Top-K candidate videos are selected; the candidate videos and query features are input into the fusion encoder and boundary-aware localization module to directly output high-quality start / end boundary probabilities; and the valid time interval candidate segments in the video are traversed. The initial boundary probability is multiplied by the termination boundary probability to obtain the local confidence score; finally, the global semantic similarity score and the local confidence score are geometrically weighted and fused according to formula (2): (2) in, A weighting factor to balance relevance and confidence level; The similarity scores between the Top-K candidate videos and the query. The first in the video Confidence scores for each candidate interval; output the total score. The highest video ID and its corresponding precise start and end time boundaries.

4. A high-efficiency target fragment retrieval and localization system based on generative enhancement, characterized in that, The system comprises: a multimodal feature extraction module, a coarse-grained video retrieval module based on bidirectional Mamba, a generative enhancement fusion module, and a boundary-aware localization module. The multimodal feature extraction module acquires initial independent features from video sequences and text queries. The coarse-grained video retrieval module based on bidirectional Mamba captures long-range temporal dependencies in videos and recalls candidate videos by inputting video frame sequences into a state-space model-based bidirectional Mamba network. The bidirectional outputs are then concatenated and fused through a linear layer to generate a new video feature sequence. The generative enhancement fusion module employs an encoder-decoder architecture, with the decoder only enabled during training. It reconstructs text queries and implicitly calibrates cross-modal attention as a semantic regularizer. This decoder uses a one-way causal masking mechanism, taking cross-modal fusion features as input and reconstructing the input natural language query text word-by-word. The boundary-aware localization module decouples the calibrated attention into start and end boundary probabilities of segments, outputting start and end boundary probability distribution sequences through a Softmax function.

5. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the efficient target fragment retrieval and localization method based on generative enhancement as described in any one of claims 1-3.

6. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the efficient target fragment retrieval and localization method based on generative enhancement as described in any one of claims 1-3.

Images