Redundant guidance dynamic sampling and prior knowledge guided surgical video time instant retrieval method

Abstract

The application discloses a redundant guidance dynamic sampling and prior knowledge guidance operation video time retrieval method, which comprises video data collection and cutting, prior knowledge base construction, lightweight feature extraction and interframe similarity calculation, redundant guidance dynamic sampling, prior knowledge prompt word construction and zero sample reasoning, model fine tuning and time retrieval, and finally multi-modal large language model output target video segment starting timestamp and ending timestamp in long operation video, completing video time retrieval. The application solves the problem of high visual redundancy of long operation video in the prior art, insufficient key information density under the fixed frame number sampling strategy, effectively integrates the prior knowledge in the operation field into the retrieval process, and realizes high-precision time sequence boundary positioning under the limited model input frame number budget.

Description

Technical Field

[0001] This invention relates to a method for retrieving surgical video moments guided by redundant dynamic sampling and prior knowledge, and relates to the fields of computer vision and medical artificial intelligence. Background Technology

[0002] In the surgical training system, surgical video teaching is a core component connecting theoretical knowledge with clinical practice. Residents and specialist trainees typically need to repeatedly watch numerous complete surgical videos to learn the standard procedures for different surgical techniques, master key operational skills, and become familiar with the timing and methods of using various surgical instruments. However, a complete surgery can easily last several hours, and the truly valuable teaching segments are often concentrated in only a few key periods. If doctors or medical students want to review a specific procedure, they often have to repeatedly drag and scroll through the lengthy videos, resulting in extremely low retrieval efficiency.

[0003] Video moment retrieval, a technology capable of accurately locating specific segments within videos based on natural language queries, has demonstrated immense application potential and practical value in the field of medical surgical education. This technology allows trainees to efficiently review specific surgical phases, operational details of key instruments, or important clinical events. However, existing video moment retrieval research largely focuses on general open domains, primarily developing for common scenarios such as daily activities and film clips, and has not yet been deeply optimized for the highly specialized field of medical surgery. Surgical videos differ significantly from general domain videos in visual features, procedural structure, and clinical semantics, making it difficult for general retrieval models to capture subtle visual changes during surgery.

[0004] In practical applications, surgical videos face severe challenges related to visual ambiguity and nuanced semantics. Under the limited endoscopic view, blood, fat, and surrounding tissues appear visually highly similar, and the identification of clinical significance often relies on subtle anatomical exposures, interactions between tools and tissues, or extremely subtle changes in movement, rather than obvious scene transitions. Furthermore, query requirements in surgical teaching scenarios are distinctly medically specific, such as "finding fragments of bile cystic artery transection using an ultrasonic scalpel." These queries require models to possess a professional understanding of surgical anatomy, instrument names, and operational actions, which general-purpose models struggle to meet. While domain-specific prior knowledge can provide strong cues for retrieval, how to effectively integrate this medical knowledge into model prompts to compensate for the compression loss of visual information remains a subject of in-depth exploration. Even more challenging is the high information redundancy present in long surgical videos, while critical target moments are typically extremely sparse. Current mainstream multimodal large language models, limited by memory and context windows, can usually only receive a fixed number of sampled frames when processing long videos. Traditional sampling strategies struggle to effectively capture those sparsely distributed critical clinical moments when faced with highly redundant videos, leading to significant visual information loss. However, existing technologies cannot meet this requirement. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a surgical video moment retrieval method guided by redundancy-guided dynamic sampling (RDS) and prior knowledge, so as to solve the problems of high visual redundancy and insufficient key information density in long-term surgical videos under the fixed frame number sampling strategy in the prior art. At the same time, it effectively integrates prior knowledge in the surgical field into the retrieval process, so as to achieve high-precision temporal boundary positioning under the limited model input frame number budget.

[0006] Technical Solution: A method for retrieving key moments in surgical videos guided by redundant dynamic sampling and prior knowledge, applicable to professional videos with extremely high visual information redundancy (such as medical surgical videos). This invention decouples sampling optimization at the data end from text prompt optimization at the inference end, achieving accurate localization of sparse key moments under limited computational budget. Specifically, it includes the following steps: Step 1, Video Data Collection and Cropping: Long surgical video data is collected, and domain experts annotate the time boundaries of specific clinical events to obtain the start and end timestamps of the set clinical events in the complete surgical video. Unlike existing methods that divide videos into fixed-length segments, this solution adopts a dynamic cropping strategy based on "virtual slicing." This eliminates the need to pre-divide the complete surgical video into fixed-length segments. Instead, the start and end timestamps of the set clinical events are used as a benchmark to determine the event time boundaries. Random expansion ratios and time offsets are applied to these event time boundaries, and the start and end times of the cropped segments are determined by combining this with a preset minimum segment length. This allows for the extraction of video segments containing the target event and its preceding and following context from the complete surgical video. By applying random expansion ratios and time offsets to the annotated event boundaries to extract video segments, this method significantly increases the diversity of sample contexts while preserving the continuity of local surgical procedures. It also effectively transforms the model's supervision signal from absolute timestamps to relative timestamps within segments, improving the model's adaptability to different temporal scales.

[0007] Step 2, Construction of the Prior Knowledge Base: Based on the experience of medical experts, a structured domain-specific prior knowledge base is constructed for different types of surgical stages. The innovation of this approach lies in the construction of a three-in-one knowledge structure consisting of "surgical stage objectives," "key anatomical structures," and "key operational steps of the stage." This structured description can compensate for the deficiencies of existing general-purpose multimodal large language models in professional medical semantic understanding. By injecting deterministic information verified by clinical experts, it provides precise medical logical guidance for subsequent reasoning paths. Surgical stage objectives describe the clinical purpose of the stage; key anatomical structures describe the identified major tissues or organs involved in the stage; and key operational steps describe the identified major surgical actions with sequential or functional relationships within the stage.

[0008] Step 3, Lightweight Feature Extraction and Inter-Frame Similarity Calculation: For the original RGB video sequence with a total of T frames obtained in Step 1... To reduce the computational load caused by the high-resolution original frames, this scheme first performs spatial downsampling and global spatial average pooling to extract the global spatial average feature vector of each frame. The calculation formula is as follows: in, This represents a video frame after spatial downsampling, where x and y represent spatial location indices in the downsampled image. and These represent the height and width of the downsampled image, respectively. Then, the cosine similarity between adjacent frames is calculated. The cosine similarity sequence of adjacent frames is obtained. As the basis for subsequent dynamic similarity threshold calculation, local stability judgment and redundancy mask generation, it quantifies the dynamic change amplitude of video with extremely low computational overhead, providing a foundation for subsequent redundancy detection.

[0009] Step 4, Redundancy-Guided Dynamic Sampling (RDS): This step is the core sampling innovation of this invention, used for the supervised fine-tuning stage of the video time-of-view retrieval model. Unlike the uniform or random sampling used in existing technologies, the Redundancy-Guided Dynamic Sampling step, based on the cosine similarity sequence of adjacent frames, performs visual redundancy detection through dual criteria. First, the extracted cosine similarity sequence of adjacent frames... Sort by numerical value in ascending order, and select the similarity value corresponding to the preset quantile as the dynamic similarity threshold. The dynamic similarity threshold Redundancy determination for adjacent frame pairs is used to adapt to different surgical scene motion patterns; secondly, a sliding window is introduced to calculate the local similarity variance. To preserve slow but meaningful camera movements while exploring the scene. Define a binarized redundancy indicator. for: in The set stability threshold is used. A binarized redundancy indicator is used to mark whether adjacent frame pairs are in a visually redundant state. It filters out minor pauses by applying duration constraints, i.e., only continuous redundant segments with a duration exceeding a preset length threshold are identified as redundant regions. A binarized redundancy mask is generated, and sampling weights are assigned to each frame. (Redundant frames have low weights, non-redundant frames have high weights), and the micro-pause refers to a continuous redundant marked segment whose length is less than a frame number threshold. Based on this, a cumulative distribution function is constructed: A fixed number of high-information-density video frame sequences are generated through sampling using the inverse cumulative distribution function. These high-information-density video frame sequences refer to sequences that possess both global temporal coverage and a non-redundant frame ratio exceeding a set threshold. This mechanism ensures that the video moment retrieval model input maintains global temporal coverage while forcing visual focus on non-redundant key information regions, significantly improving the information density of the input.

[0010] Step 5, Prior Knowledge Prompt Construction and Zero-Shot Inference: During the zero-shot evaluation test, the video time-lapse model receives natural language retrieval instructions and retrieves corresponding medical prior knowledge from the domain prior knowledge base constructed in Step 2. The retrieved surgical stage objectives, anatomical structures, and operational steps are injected as complementary contextual information into the prompts, constructing enhanced retrieval prompts. The advantage of this approach lies in that it does not change the fine-tuned model parameters, but rather establishes a strong alignment between professional language and visual features directly during the inference stage by optimizing the Prompt semantic space, thereby significantly improving localization accuracy without increasing the model training burden.

[0011] Step 6, Model Fine-tuning and Time-of-Sight Retrieval: The high-information-density video frame sequence obtained from the redundantly guided dynamic sampling in Step 4, along with standardized query instructions, is input into the multimodal large language model for supervised fine-tuning. Finally, the multimodal large language model outputs the start and end timestamps of the target video segment within the long surgical video, completing the video time-of-sight retrieval.

[0012] 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 redundant guided dynamic sampling and prior knowledge-guided surgical video moment retrieval method described above.

[0013] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the redundant-guided dynamic sampling and prior knowledge-guided surgical video moment retrieval method described above.

[0014] Beneficial effects: This invention provides a method for retrieving surgical video moments guided by redundant dynamic sampling and prior knowledge. The technology is mainly based on a multimodal large language model, which has achieved significant beneficial effects in improving the efficiency of long video feature sampling, enhancing fine-grained semantic localization capabilities, and establishing industry evaluation benchmarks.

[0015] To verify the practical application effectiveness of this method, this invention first constructed a video time-based retrieval dataset (SurgScope) covering multiple types of surgeries. This dataset originates from real clinical surgical scenarios and covers five typical surgical types, including pituitary surgery, mitral valve repair, cholecystectomy, hysterectomy, and cataract surgery. The dataset contains over 16,000 expert-annotated query-time pairs, covering 50 surgical stages and 26 surgical instruments. This large-scale and highly specialized benchmark dataset not only provides a rigorous testing environment for the technical solution proposed in this invention but also fills the gap in the medical field regarding the lack of dedicated video retrieval evaluation standards.

[0016] Validated on the aforementioned benchmark datasets, the redundant guided dynamic sampling strategy proposed in this invention demonstrates superior sampling efficiency and information density enhancement capabilities. As a lightweight, plug-and-play module, RDS can robustly identify and filter highly redundant segments in long videos. Experimental data shows that RDS significantly outperforms traditional uniform or random sampling strategies when processing surgical videos: taking the VTimeLLM model as an example, after adopting RDS, the average intersection-over-union ratio (IoU) for surgical stage localization increased from 43.30 to 45.73, and the high-precision localization index R1@0.7 increased from 20.07 to 22.76; for the TimeChat model, mIoU also increased from 48.47 to 50.38. This proves that this invention can maximize the retention of core clinical information frames within a limited computational budget, effectively alleviating the information loss problem in long video processing.

[0017] Furthermore, this invention significantly enhances the model's ability to understand complex medical semantics by introducing a domain-specific prior knowledge prompting mechanism. By injecting expert knowledge such as stage goals, anatomical structures, and operational procedures into the prompt words, the model can better bridge the semantic gap between visual features and professional language. In the evaluation, after adding prior knowledge prompts, the localization mIoU of the VideoChat-Flash model increased significantly from 29.01 to 37.02, and the mIoU of the TimeChat model increased even more dramatically from 5.49 to 39.48. These quantitative results fully demonstrate the significant advantages of this invention in handling long-term program contexts, improving retrieval accuracy, and robustness, providing strong support for the practical application of medical artificial intelligence technology. Attached Figure Description

[0018] Figure 1 This is a flowchart of the dataset generation process according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] A method for retrieving surgical video moments guided by redundant dynamic sampling and prior knowledge includes the following steps: Step 1, Training Data Preparation and Expert-Level Annotation: The implementation was first performed on the SurgScope benchmark dataset, which contains 243 surgical videos. This dataset covers five surgical procedures: endoscopic pituitary surgery, thoracoscopic mitral valve repair, cholecystectomy, laparoscopic hysterectomy, and cataract surgery. During annotation, clinical medical experts precisely defined the appearance times of 50 surgical stages and 26 surgical instruments in the videos, ultimately generating 16,169 "instruction-time" pairs. Invalid samples due to frame corruption or blurry annotations were removed, providing standardized supervision signals for subsequent fine-tuning and zero-shot validation of the algorithm. After obtaining the start and end timestamps of the target events, video segment trimming was further performed based on the time boundary annotations. Specifically, this embodiment does not employ a fixed-length slicing method, but instead uses a dynamic cropping strategy based on "virtual slicing": Based on the start and end timestamps of the target event, a random expansion ratio and time offset are applied to the target event boundary, and the start and end times of the cropped segment are determined in conjunction with a preset minimum segment length. Subsequently, the corresponding video segment is extracted from the complete surgical video, ensuring that the cropped video segment contains the target event and its preceding and following local context. For the cropped video segment, the absolute start and end timestamps of the target event in the complete surgical video are converted into relative start and relative end timestamps within the cropped segment, serving as temporal boundary supervision information for supervised fine-tuning, zero-shot evaluation, and time-bound retrieval of the multimodal large language model.

[0021] Step 2, Domain Prior Knowledge Base Construction and Prompt Enhancement: For the surgical stages mentioned above, a structured prior knowledge base was constructed, comprising three dimensions: "surgical stage objectives," "key anatomical structures," and "key operational steps of the stage." This knowledge base was rigorously reviewed and revised line by line by surgeons using the initial draft generated by the large language model to ensure the medical professionalism of the descriptions, with each description averaging approximately 143.7 words. During zero-shot prediction, the prior knowledge from this knowledge base was dynamically injected into the natural language query. Enhanced prompts were constructed to guide the model in establishing complex contextual relationships between procedural semantics and visual features during the inference stage.

[0022] Step 3, Lightweight Feature Representation and Inter-Frame Similarity Calculation: For the input original RGB video frames First, a spatial downsampling operation with a stride of 4 is used to reduce the computational load. Then, global spatial average pooling is performed to extract the global average feature vector for each frame. The calculation formula is as follows: in, For downsampled frames, and These represent the downsampled height and width, respectively. After obtaining the feature vectors, L2 normalization is applied, and the cosine similarity between consecutive adjacent frame feature pairs is calculated. This allows for the quantification of the dynamic amplitude of video stream changes over time, providing a mathematical basis for subsequent redundancy determination.

[0023] Step 4, Redundancy-Guided Dynamic Sampling (RDS) Algorithm Logic: To optimize the input efficiency in the fine-tuning stage, a dual criterion is used for visual redundancy detection, namely, selecting the 40th percentile of the similarity sequence arranged in ascending order as the dynamic similarity threshold. Simultaneously, the similarity variance is calculated within a 5-frame sliding window. And set a stability threshold Define a binary redundancy indicator. Furthermore, a duration constraint is imposed, meaning that only segments consecutively marked as redundant are ultimately considered valid redundancy if their length exceeds 12 frames, thus preserving clinically significant minor movements. Subsequently, redundant frames are assigned lower sampling weights. Assign higher weights to non-redundant frames This effectively increases the probability of non-redundant areas being sampled.

[0024] Step 5, Visual sequence compression based on the inverse cumulative distribution function: Construct the cumulative distribution function (CDF) based on the assigned weight sequence, the formula of which is: For the 100-frame fixed input budget of the multimodal large language model, a set of target values ​​is generated at equal intervals in the range [0,1]. The target values ​​are then sampled and mapped back to the original frame index through the inverse cumulative distribution function. This generates a set of compressed visual sequences with a fixed number of frames, global distribution and high information density without changing the underlying network structure of the model.

[0025] Step 6, Model Fine-tuning Training and Multi-path Boundary Prediction: Two independent application paths are set up. The first is to input the high-density visual sequence generated in Step 5 into the multimodal large language model for supervised fine-tuning to optimize the model's sensitivity to localization in long surgical videos. The second is to combine the enhanced cue words generated in Step 2 with video features to perform cross-modal reasoning under zero-shot evaluation testing, and finally output the relative start and end timestamps of the target event in the video.

[0026] Step 7, Performance Testing and Result Verification: To verify the effectiveness of the above method, this embodiment conducted tests and evaluations on the SurgScope surgical benchmark. In the localization task during the model fine-tuning stage, when the number of model input frames is fixed at 100 frames, compared with the default uniform or random sampling, using redundant guided dynamic sampling frames as input significantly improves model performance. As shown in Table 1, after fine-tuning, the average intersection-over-union ratio (mIoU) of the VTimeLLM model increased from 43.30 to 45.73, and the high-precision localization index R1@0.7 increased from 20.07 to 22.76; under the same fine-tuning conditions, the average intersection-over-union ratio of the TimeChat model also increased from 48.47 to 50.38, and R1@0.7 jumped from 26.70 to 31.72. Furthermore, in finer-grained instrument localization fine-tuning tasks, redundant guided dynamic sampling also improved the average intersection-over-union (IoU) of the TimeChat model from 5.37 to 6.52, and the R1@0.7 more than doubled from 0.63 to 1.48. These data fully demonstrate that our method can effectively improve the model performance of long video multi-granularity moment retrieval tasks by dynamically filtering redundant features. As shown in Table 2, in the zero-shot localization task, adding surgical prior knowledge prompts showed a significant enhancement in professional semantic alignment capabilities. The average IoU of the TimeChat model jumped significantly from 5.49 to 39.48, the average IoU of the VideoChat-Flash model also increased significantly from 29.01 to 37.02, and the average IoU of the Qwen2.5-VL-7B-Instruct model also achieved a significant increase from 11.20 to 17.14. The experimental results of the above multi-model studies consistently demonstrate that this prior knowledge prompting mechanism significantly improves the model's ability to align professional medical semantics and locate temporal boundaries when targeted training is lacking.

[0027] Table 1. Performance of multimodal large language models fine-tuned with different sampling methods in surgical localization tasks.

[0028] Table 2. Impact of Adding Surgical Prior Knowledge on Zero-Shot Surgical Stage Localization Performance of Multimodal Large Language Model

[0029] Step 8, Method Output and Practical Application: The method described in this invention ultimately implements an end-to-end algorithm for retrieving timeframes from long surgical videos. In practical applications, the user inputs a complete original long surgical video and text prompts describing the target retrieval event (e.g., a natural language query: "When does the nasal passage creation stage appear in this endoscopic pituitary surgery video?"). Based on the input, the algorithm performs video data feature dimensionality reduction, redundant filtering, dynamic sampling, and prior knowledge fusion prompts. Then, a multimodal large language model performs cross-modal alignment parsing to obtain and output the start and end timestamps of the target query segment within the video. Using this timestamp output, surgeons or medical students can overcome the viewing burden of lengthy surgical videos, directly and quickly locating and jumping to the corresponding specific surgical segment, greatly saving time in retrieving specific surgical events. This can directly serve postoperative safety assessment, surgeon skill assessment, and clinical surgical teaching.

[0030] In summary, the algorithm disclosed in this invention optimizes the information density of visual input through a redundancy-guided dynamic sampling strategy and enhances professional semantic understanding by combining expert prior knowledge, effectively solving the problem of inaccurate localization caused by visual redundancy and semantic complexity in long surgical video retrieval.

[0031] This invention discloses a method for retrieving surgical video moments guided by redundancy-guided dynamic sampling and prior knowledge. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention. For example, they can adjust the similarity percentile ratio for redundancy determination, modify the similarity calculation formula (e.g., using Euclidean distance), or apply it to multimodal base models with different architectures. 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 technologies.

[0032] Obviously, those skilled in the art should understand that the steps of the redundant guided dynamic sampling and prior knowledge-guided surgical video moment retrieval method described 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 optionally, they can be implemented using device-executable program code, thereby storing them in a storage device for execution by the computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than 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 method for retrieving surgical video moments guided by redundant dynamic sampling and prior knowledge, characterized in that, Includes the following steps: Step 1, Video Data Collection and Editing: Collect surgical video data and mark the time boundaries of the defined clinical events; Video segment extraction based on boundary annotation; Step 2, Construction of Prior Knowledge Base: Construct a structured domain prior knowledge base for different types of surgical stages; Step 3, Lightweight Feature Extraction and Inter-Frame Similarity Calculation: For a surgical video data sequence with a total of T frames... First, spatial downsampling and global spatial average pooling are performed to extract each frame. Global space average eigenvector Subsequently, the cosine similarity between adjacent frames is calculated to obtain the cosine similarity sequence between adjacent frames. As the basis for subsequent dynamic similarity threshold calculation, local stability judgment and redundancy mask generation, it realizes the quantification of the dynamic change amplitude of video and provides a basis for subsequent redundancy detection. Step 4, Redundancy-Guided Dynamic Sampling: This step is used for the supervised fine-tuning stage of the video time-lapse model. First, the extracted cosine similarity sequences of adjacent frames are... Sort by numerical value in ascending order, and select the similarity value corresponding to the preset quantile as the dynamic similarity threshold. The dynamic similarity threshold Redundancy determination for adjacent frame pairs is used to adapt to different surgical scene motion patterns; secondly, a sliding window is introduced to calculate the local similarity variance of adjacent frame sequences. A binary redundancy indicator is defined to mark whether adjacent frame pairs are in a visually redundant state. A duration constraint is applied to filter out minor pauses, i.e., only continuous redundant segments with a duration exceeding a preset length threshold are identified as redundant regions. A binary redundancy mask is generated, and sampling weights are assigned to each frame. A cumulative distribution function is constructed, and a fixed number of high-information-density video frame sequences are generated through sampling using the inverse cumulative distribution function. The high-information-density video frame sequence refers to a video frame sequence that has both global temporal coverage and a non-redundant frame ratio greater than a set threshold. The minor pause refers to a continuous redundant marked segment whose length is less than the frame number threshold. Step 5, Construction of prior knowledge prompts and zero-shot inference: When the model is tested for zero-shot evaluation, it receives natural language retrieval instructions and retrieves the corresponding medical prior knowledge from the domain prior knowledge base constructed in Step 2; the retrieved task objectives, anatomical structures and operation steps are injected into the prompts as complementary contextual information to construct enhanced retrieval prompts; Step 6, Model Fine-tuning and Time-of-Sight Retrieval: The high-information-density video frame sequence obtained from the redundantly guided dynamic sampling in Step 4, along with standardized query instructions, is input into the multimodal large language model for supervised fine-tuning. Finally, the multimodal large language model outputs the start and end timestamps of the target video segment within the long surgical video, completing the video time-of-sight retrieval.

2. The surgical video moment retrieval method guided by redundant dynamic sampling and prior knowledge as described in the claim, characterized in that, In step 1, long surgical video data is collected, and domain experts mark the time boundaries of the set clinical events to obtain the start and end timestamps of the set clinical events in the complete surgical video. A dynamic cropping strategy based on "virtual slicing" is adopted, that is, without pre-dividing the complete surgical video into fixed-length segments, the event time boundary is determined based on the start and end timestamps of the set clinical events, a random expansion ratio and time offset are applied to the event time boundary, and the start and end times of the cropped segments are determined in combination with the preset minimum segment length, thereby extracting video segments containing the target event and its preceding and following context from the complete surgical video.

3. The surgical video moment retrieval method guided by redundant dynamic sampling and prior knowledge as described in the claim is characterized in that, In step 2, a three-in-one knowledge structure is constructed, consisting of "surgical stage objectives", "key anatomical structures" and "key operational steps of the stage". The surgical stage objectives are used to describe the clinical purpose of the stage; the key anatomical structures are used to describe the identified major tissues or organs involved in the stage; and the key operational steps of the stage are used to describe the identified major surgical actions in the stage that have sequential or functional relationships.

4. The surgical video moment retrieval method guided by redundant dynamic sampling and prior knowledge as described in the claim, characterized in that, In step 3, the video sequence obtained in step 1 First, spatial downsampling and global spatial average pooling are performed to extract the global spatial average feature vector for each frame. The calculation formula is as follows: in, This represents a video frame after spatial downsampling, where x and y represent spatial location indices in the downsampled image. and Let the height and width of the downsampled image be represented respectively; then the cosine similarity between adjacent frames is calculated. This enables the quantification of the dynamic changes in video, providing a foundation for subsequent redundancy detection.

5. The surgical video moment retrieval method guided by redundant dynamic sampling and prior knowledge as described in claim 4, wherein the redundant guided dynamic sampling step based on the cosine similarity sequence of adjacent frames performs visual redundancy detection through dual standards; firstly, a dynamic similarity threshold is calculated. To adapt to different surgical movement patterns; secondly, a sliding window is introduced to calculate the local similarity variance. Define a binary redundancy indicator. for: in A stability threshold is set; after applying duration constraints to filter out minor pauses, a binary redundancy mask is generated, and sampling weights are assigned to each frame. Construct the cumulative distribution function: A fixed number of video frame sequences are generated by sampling using the inverse cumulative distribution function. The micro-pause refers to a continuous redundant marker segment whose length is less than the frame number threshold.

6. 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 surgical video moment retrieval method guided by redundant guided dynamic sampling and prior knowledge as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instruction is executed by the processor, it implements the steps of the surgical video moment retrieval method guided by redundant guided dynamic sampling and prior knowledge as described in any one of claims 1-5.

Images