An audio-visual aesthetic automatic evaluation method, device, equipment and storage medium

Abstract

The application discloses an audio-visual aesthetics automatic evaluation method and device, equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: constructing an audio-visual aesthetics evaluation task context and a plurality of segment-level audio-visual units; analyzing to obtain scores of a plurality of evaluation dimensions; determining adaptive weights of each evaluation dimension to obtain an overall aesthetic score; performing similarity evaluation on each segment-level audio-visual unit, and corresponding grouping into similar audio-visual sample groups or comparable segment groups; performing pairwise comparison on the similar audio-visual sample groups or the comparable segment groups to obtain a relative ranking calibration result; and fusing the scores corresponding to the plurality of evaluation dimensions, the adaptive weights and the relative ranking calibration result to form an audio-visual aesthetics total score and a dimension-level diagnosis result. The application solves the technical problems of single evaluation dimension, insufficient audio-visual collaborative modeling, difficult identification of fine-grained differences, subjective key parameters and insufficient interpretability of evaluation results in the prior art.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an automatic evaluation method, apparatus, device, and storage medium for audiovisual aesthetics. Background Technology

[0002] With the rapid growth in the number of short videos, advertising videos, film clips, music videos, live stream clips, interview videos, educational training videos, documentary clips, and social media audio and video content, users are not only concerned with whether the picture is clear and the sound is audible, but also with the composition of the picture, color style, audio quality, rhythm changes, emotional expression, and the harmony between audio and video. How to automatically, stably, and finely evaluate the aesthetic quality of existing audiovisual content has become an important technical issue in content management, material selection, version quality inspection, video editing, platform recommendation, and automatic review.

[0003] Existing image aesthetic evaluation methods typically classify or score the quality of individual images, which can, to some extent, determine the image's composition, color, subject prominence, and overall visual appeal. While these methods are effective for evaluating static images, they struggle to express continuous images, audio signals, rhythmic structures, emotional changes, and cross-modal correspondences.

[0004] Existing video quality assessment methods focus more on objective quality indicators such as sharpness, compression distortion, noise, and stuttering, making it difficult to fully measure compositional aesthetics, musical atmosphere, rhythmic tension, emotional consistency, and audio-visual harmony. Current multimodal assessment methods typically use a single model to directly output an overall score, lacking a specialized assessment mechanism that addresses dimensions such as visual, audio, rhythm, synchronization, and emotion. This results in insufficient interpretability of the assessment process and difficulty in pinpointing specific problems in the audiovisual content.

[0005] Furthermore, existing evaluation methods typically rely on absolute scores, making it difficult to distinguish subtle aesthetic differences when multiple audiovisual segments are similar in theme, scene, or content. For example, the difference in matching between shot transitions and musical beats can be expressed by the beat matching score R_beat in the rhythm structure rating; the difference in the temporal correspondence between visual and audio events can be expressed by the audio-visual synchronization score Q_sync; the difference between audio atmosphere and visual emotion can be expressed by the emotional consistency score Q_emo; and the relative difference between different edited versions can be expressed by the sorting calibration score S_i'. However, a single overall score is insufficient to consistently reflect these objective differences. For music videos, interview videos, advertising videos, short films, documentaries, instructional videos, and social media audio-visual content, the importance of different aesthetic dimensions varies, and fixed weights can easily lead to inconsistencies between evaluation results and annotation / verification results or review results. In addition, existing methods rely heavily on human experience to judge parameters such as "segment importance," "shot type," and "semantic function," lacking calculable indicators, mathematical mappings, and reproducible processes, making it difficult to reproduce and review evaluation results.

[0006] Therefore, there is a need for an automated audiovisual aesthetics evaluation scheme that can simultaneously handle visual, audio, and cross-modal relationships, define the above parameters as feature-based computable indicators, and enable multiple agents to collaboratively complete fine-grained evaluation, weight estimation, relative ranking calibration, and comprehensive diagnosis. Summary of the Invention

[0007] In view of the deficiencies in the existing technology, the purpose of this application is to provide an automatic evaluation method, device, equipment and storage medium for audiovisual aesthetics, so as to solve the technical problems of the existing technology, such as the single dimension of audiovisual aesthetics evaluation, insufficient audio-visual collaborative modeling, difficulty in identifying fine-grained differences, subjectivity of key parameters and insufficient interpretability of evaluation results.

[0008] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0009] In a first aspect, this application provides an automatic evaluation method for audiovisual aesthetics, the method comprising the following steps:

[0010] Construct the audiovisual aesthetics assessment task context based on the audiovisual content to be evaluated and its related information;

[0011] The audiovisual content to be evaluated is processed by aligning video frame sequences, audio segments, subtitle text, and timestamps to obtain multiple segment-level audiovisual units;

[0012] Based on the aforementioned segment-level audiovisual unit, and combined with the context of the audiovisual aesthetic evaluation task, the intelligent agent is analyzed through multiple dimensions, including feature extraction, semantic recognition, index normalization, and weighted fusion, to obtain scores corresponding to various evaluation dimensions.

[0013] By estimating the agent through weights, adaptive weights for each evaluation dimension are determined based on content type, modal saliency, fragment complexity, and evaluation objectives, and the corresponding overall aesthetic score is obtained.

[0014] For each of the aforementioned segment-level audiovisual units, a similarity assessment is performed. Each segment-level audiovisual unit with a similarity not lower than a preset intra-group similarity threshold is grouped into the same similar audiovisual sample group, and otherwise it is grouped into a comparable segment group.

[0015] By using a fine-grained preference comparison agent, pairs of similar audiovisual sample groups or comparable segment groups are compared, and relative ranking calibration is performed based on the comparison results to obtain the relative ranking calibration results.

[0016] By comprehensively evaluating the agent and fusing scores, adaptive weights, and relative ranking calibration results corresponding to multiple evaluation dimensions, a total audiovisual aesthetic score and dimension-level diagnostic results are formed; among them...

[0017] The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files.

[0018] The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags;

[0019] The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0020] Based on the above technical solution, the step of constructing an audiovisual aesthetic evaluation task context according to the audiovisual content to be evaluated and its associated information includes the following steps:

[0021] Receive the audiovisual content to be evaluated;

[0022] Receive the associated information,

[0023] The audiovisual content to be evaluated is subjected to format standardization and quality preprocessing.

[0024] The associated information is structured to extract content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags;

[0025] Based on unique content identifiers and version identifiers, the audiovisual content to be evaluated (after standardized formatting or quality preprocessing), the related information after structured processing, the segment timeline, and feature indexes are integrated into an audiovisual aesthetic evaluation task context that can be invoked by multiple agents. This creates a one-to-one and traceable mapping relationship between each audiovisual content to be evaluated or each edit version and its related information and audiovisual aesthetic evaluation task context; among which,

[0026] The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

[0027] Based on the above technical solution, the step of aligning the video frame sequence, audio segment, subtitle text, and timestamp of the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units includes the following steps:

[0028] Shot boundary detection is performed on the video frame sequence of the audiovisual content to be evaluated to obtain multiple shot segments, and keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features, and sharpness features are extracted from each shot segment;

[0029] The audio segments of the audiovisual content to be evaluated are segmented to obtain multiple audio segments, and loudness, beat, accent, pitch variation, spectral centroid, energy variation, speech intelligibility, background noise, musical mood and timbre texture features are extracted.

[0030] Semantic analysis is performed on the subtitle text of the audiovisual content to be evaluated to obtain character speech, emotion words, event nodes, semantic turning points and thematic keywords;

[0031] Based on the timestamps of the audiovisual content to be evaluated, video shots, audio clips, and subtitle text are aligned to form segment-level audiovisual units.

[0032] Based on the above technical solution, the method further includes the following steps:

[0033] Output an evaluation report; among which,

[0034] The evaluation report includes the overall aesthetic score, scores corresponding to multiple evaluation dimensions, and the relative ranking calibration results.

[0035] Based on the above technical solution, the output evaluation report includes the following steps:

[0036] Output the overall aesthetic score, which is in at least one of the following forms: a percentage system, a ten-point system, a grade system, or a ranking system.

[0037] Output scores corresponding to multiple evaluation dimensions, including visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0038] Output the relative ranking calibration results, which include the overall ranking among multiple audiovisual versions, the difference in dimensionality of each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results.

[0039] Secondly, this application provides an automatic audiovisual aesthetic evaluation device, the device comprising:

[0040] The context task construction module is used to construct the context of the audiovisual aesthetic evaluation task based on the audiovisual content to be evaluated and its associated information.

[0041] The segment unit acquisition module is used to perform video frame sequence, audio segment, subtitle text and timestamp alignment processing on the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units;

[0042] The multi-dimensional evaluation module is used to obtain scores corresponding to multiple evaluation dimensions based on the segment-level audiovisual unit and in combination with the context of the audiovisual aesthetic evaluation task, through feature extraction, semantic recognition, index normalization and weighted fusion of the intelligent agent.

[0043] The overall aesthetic evaluation module is used to estimate the agent by weights and determine the adaptive weights of each evaluation dimension based on content type, modal saliency, fragment complexity and evaluation objectives.

[0044] The segment segmentation module is used to evaluate the similarity of each segment-level audiovisual unit, and classify each segment-level audiovisual unit with a similarity not lower than a preset intra-group similarity threshold into the same similar audiovisual sample group, and otherwise classify them into a comparable segment group.

[0045] The relative ranking calibration module is used to perform pairwise comparisons of similar audiovisual sample groups or comparable segment groups by a fine-grained preference comparison agent, and perform relative ranking calibration based on the comparison results to obtain the relative ranking calibration result.

[0046] The comprehensive evaluation module is used to fuse scores, adaptive weights, and relative ranking calibration results corresponding to multiple evaluation dimensions through a comprehensive evaluation agent, forming an overall audiovisual aesthetic score and dimension-level diagnostic results; among them...

[0047] The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files.

[0048] The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags;

[0049] The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0050] Based on the above technical solution, the context task construction module is also used to receive the audiovisual content to be evaluated;

[0051] The context task construction module is also used to receive the associated information.

[0052] The context task construction module is also used to perform format unification and quality preprocessing on the audiovisual content to be evaluated;

[0053] The context task construction module is also used to perform structured processing on the associated information, extracting content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags;

[0054] The context task construction module is also used to integrate the audiovisual content to be evaluated (after unified formatting or quality preprocessing), the structured related information, the segment timeline, and the feature index into an audiovisual aesthetic evaluation task context that can be invoked by multiple intelligent agents, based on the content's unique identifier and version identifier. This ensures that each audiovisual content to be evaluated or each edit version forms a one-to-one and traceable mapping relationship with its related information and the audiovisual aesthetic evaluation task context.

[0055] The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

[0056] Based on the above technical solution, the segment splitting module is also used to perform shot boundary detection on the video frame sequence of the audiovisual content to be evaluated, obtain multiple shot segments, and extract keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features and sharpness features from each shot segment;

[0057] The segment splitting module is also used to segment the audio segments of the audiovisual content to be evaluated, obtain multiple audio segments, and extract loudness, beat, accent, pitch change, spectral centroid, energy change, speech intelligibility, background noise, musical emotion and timbre texture features.

[0058] The segment segmentation module is also used to perform semantic analysis on the subtitle text of the audiovisual content to be evaluated, and to obtain character speech, emotion words, event nodes, semantic turning points and theme keywords;

[0059] The segment splitting module is also used to align video shots, audio clips, and subtitle text based on the timestamps of the audiovisual content to be evaluated, forming segment-level audiovisual units.

[0060] Based on the above technical solution, the device also includes a report output module;

[0061] The report output module is used to output an evaluation report; wherein,

[0062] The evaluation report includes the overall aesthetic score, scores corresponding to multiple evaluation dimensions, and the relative ranking calibration results.

[0063] Based on the above technical solution, the report output module is also used to output the overall aesthetic score, which is in at least one of the following forms: percentage system, ten-point system, grade system or ranking system.

[0064] The report output module is also used to output scores corresponding to multiple evaluation dimensions, including visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0065] The report output module is also used to output the relative ranking calibration results, which include the overall ranking among multiple audiovisual versions, the difference in dimensionality of each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results.

[0066] Thirdly, this application provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps described in the method mentioned in the first aspect.

[0067] Fourthly, this application provides a computer-readable storage medium storing computer instructions for causing a processor to perform the steps described in the method mentioned in the first aspect.

[0068] Compared with the prior art, the advantages of this application are:

[0069] This application aims to address the technical problems in existing technologies, such as the single dimension of audiovisual aesthetic evaluation, insufficient audio-visual collaborative modeling, difficulty in identifying fine-grained differences, subjectivity of key parameters, and insufficient interpretability of evaluation results.

[0070] This application no longer evaluates only the quality of a single image or a single video, but simultaneously models the visual, audio, rhythm, synchronization and emotional semantic relationships, which can more comprehensively reflect the overall aesthetic quality of audiovisual content.

[0071] This application achieves a unified representation of video frames, audio segments, subtitle text, and timestamps through segment-level audiovisual units, which can locate specific low-scoring segments and output interpretable evidence and diagnostic information.

[0072] This application utilizes the collaborative work of multiple intelligent agents, each of which undertakes specific functions such as task decomposition, feature invocation, rule execution, score interpretation, and result verification. This avoids the opaque evaluation process caused by a single model directly outputting the overall score, thereby improving the interpretability and traceability of audiovisual aesthetic evaluation.

[0073] This application uses an adaptive weight estimation mechanism to dynamically generate weight vectors based on content type, modal saliency, and fragment complexity. It then combines these weight vectors with annotation verification results for training, threshold calibration, or review, thereby improving the consistency between the fusion results of each dimension and the annotation verification results.

[0074] This application improves the practicality of the evaluation results in content selection, version comparison, material quality inspection, content review and platform recommendation by using a fine-grained preference comparison and relative ranking calibration mechanism. This is based on preference probability, rating preference probability and calibration score difference between multiple similar audiovisual content or similar segments.

[0075] This application uses a review agent to perform a second check on scoring conflicts and low-confidence results, which can reduce scoring bias caused by misjudgment by a single agent, thereby improving the stability and reliability of audiovisual aesthetic evaluation.

[0076] This application defines key parameters such as shot type, semantic function, segment importance and sample similarity as feature-based computable indicators and mathematical mappings, and uses human preference labels as reference data for training, threshold calibration or review, so that the evaluation process is transformed from subjective experience judgment into a quantifiable and reproducible objective technical process. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a schematic diagram illustrating the principle and flow of the automatic audiovisual aesthetics evaluation method according to an embodiment of this application;

[0079] Figure 2 This is a schematic diagram of the audiovisual aesthetics evaluation task context construction process of the audiovisual aesthetics automatic evaluation method in this application embodiment;

[0080] Figure 3 This is a schematic diagram of the segment-level audiovisual unit construction process of the audiovisual aesthetic automatic evaluation method according to an embodiment of this application;

[0081] Figure 4 This is a schematic diagram of the multi-dimensional intelligent agent scoring process of the audiovisual aesthetic automatic evaluation method according to an embodiment of this application.

[0082] Figure 5 This is a schematic diagram of the adaptive weight estimation process of the audiovisual aesthetic automatic evaluation method according to an embodiment of this application.

[0083] Figure 6 This is a schematic diagram of the fine-grained preference comparison and relative ranking calibration process of the audiovisual aesthetic automatic evaluation method according to an embodiment of this application;

[0084] Figure 7 This is a schematic diagram of the comprehensive evaluation and multi-agent collaborative review process of the audiovisual aesthetic automatic evaluation method according to an embodiment of this application;

[0085] Figure 8 This is a schematic diagram of the evaluation report output process of the audiovisual aesthetic automatic evaluation method according to an embodiment of this application;

[0086] Figure 9 This is a schematic diagram of the principle framework of the audiovisual aesthetic automatic evaluation device according to an embodiment of this application. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0088] The embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0089] This application provides an automatic evaluation method, apparatus, device, and storage medium for audiovisual aesthetics, in order to solve the technical problems in the prior art such as single evaluation dimension of audiovisual aesthetics, insufficient audio-visual collaborative modeling, difficulty in identifying fine-grained differences, subjectivity of key parameters, and insufficient interpretability of evaluation results.

[0090] To achieve the aforementioned technical effects, the overall concept of this application is as follows:

[0091] An automatic evaluation method for audiovisual aesthetics, comprising the following steps:

[0092] S1. Construct the audiovisual aesthetics assessment task context based on the audiovisual content to be evaluated and its related information;

[0093] S2. Align the video frame sequence, audio segment, subtitle text and timestamp of the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units;

[0094] S3. Based on the segment-level audiovisual unit and combined with the audiovisual aesthetic evaluation task context, the intelligent agent is analyzed through multiple dimensions, including feature extraction, semantic recognition, index normalization, and weighted fusion, to obtain scores corresponding to multiple evaluation dimensions.

[0095] S4. By estimating the agent through weights, determine the adaptive weights of each evaluation dimension based on content type, modal saliency, fragment complexity, and evaluation objectives, and obtain the corresponding overall aesthetic score.

[0096] S5. For each of the segment-level audiovisual units, perform similarity evaluation, and classify each segment-level audiovisual unit with a similarity not lower than the preset intra-group similarity threshold into the same similar audiovisual sample group, otherwise classify it into the comparable segment group.

[0097] S6. By using a fine-grained preference comparison agent, pairwise comparisons are performed on similar audiovisual sample groups or comparable segment groups, and relative ranking calibration is performed based on the comparison results to obtain the relative ranking calibration results.

[0098] S7. By comprehensively evaluating the intelligent agent, the scores, adaptive weights, and relative ranking calibration results corresponding to multiple evaluation dimensions are fused to form a total audiovisual aesthetic score and dimension-level diagnostic results; among which...

[0099] The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files.

[0100] The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags;

[0101] The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0102] The embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0103] Firstly, see [the following] Figures 1-8 As shown in the figure, this application provides an automatic audiovisual aesthetic evaluation method, which includes the following steps:

[0104] S1. Construct the audiovisual aesthetics assessment task context based on the audiovisual content to be evaluated and its related information;

[0105] S2. Align the video frame sequence, audio segment, subtitle text and timestamp of the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units;

[0106] S3. Based on the segment-level audiovisual unit and combined with the audiovisual aesthetic evaluation task context, the intelligent agent is analyzed through multiple dimensions, including feature extraction, semantic recognition, index normalization, and weighted fusion, to obtain scores corresponding to multiple evaluation dimensions.

[0107] S4. By estimating the agent through weights, determine the adaptive weights of each evaluation dimension based on content type, modal saliency, fragment complexity, and evaluation objectives, and obtain the corresponding overall aesthetic score.

[0108] S5. For each of the segment-level audiovisual units, perform similarity evaluation, and classify each segment-level audiovisual unit with a similarity not lower than the preset intra-group similarity threshold into the same similar audiovisual sample group, otherwise classify it into the comparable segment group.

[0109] S6. By using a fine-grained preference comparison agent, pairwise comparisons are performed on similar audiovisual sample groups or comparable segment groups, and relative ranking calibration is performed based on the comparison results to obtain the relative ranking calibration results.

[0110] S7. By comprehensively evaluating the intelligent agent, the scores, adaptive weights, and relative ranking calibration results corresponding to multiple evaluation dimensions are fused to form a total audiovisual aesthetic score and dimension-level diagnostic results; among which...

[0111] The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files.

[0112] The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags;

[0113] The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0114] When necessary, an automatic audiovisual aesthetic evaluation system can be constructed based on the technical solutions of the embodiments of this application. For ease of implementation, the data sources, data types, and their roles in the system used in the embodiments of this application are first explained as follows:

[0115] Data sources include: video frame sequences and shot sequences. The input format is image frames arranged in chronological order and their shot boundaries. The acquisition method is to perform video decoding and shot boundary detection on video files or video links. Its purpose is to provide visual features.

[0116] The audio signal is input in the form of time-ordered audio samples, acquired through audio decoding and segmentation, and its purpose is to provide audio characteristics.

[0117] The input format is text with timestamps, and the acquisition method is subtitle extraction or speech recognition. Its purpose is to provide semantic features of the text.

[0118] Semantic tags, including content type tags, topic tags, sentiment tags, style tags, scene tags, and rhythm tags, are used to constrain weight estimation and similarity calculation. In the case of existing labeled data, human preference tags are used for model training, threshold calibration, or result verification.

[0119] All types of data undergo unified preprocessing, including format unification, frame rate standardization, loudness standardization, silent segment detection, black screen segment detection, and abnormal segment marking, before being formed into segment-level audiovisual units.

[0120] Based on this, the technical chain of this application embodiment is "input—feature extraction—semantic recognition—similarity calculation—score fusion—verification—result output":

[0121] After receiving the audiovisual content and related information, the system performs feature extraction and semantic recognition by the segment-level audiovisual unit construction module;

[0122] When multiple versions or comparable segments exist, the similarity calculation and grouping are performed by the preference comparison module.

[0123] The scoring is integrated using a multi-dimensional scoring module and a comprehensive evaluation module.

[0124] The review agent verifies the results in case of scoring conflicts; finally, the evaluation report output module outputs the evaluation results and machine-readable data.

[0125] Furthermore, in the technical solutions of this application embodiment, each intelligent agent undertakes a specific and implementable technical function, rather than being "determined by artificial intelligence" in a general sense. Specifically: The agent performs task decomposition (breaking down segment-level audiovisual units into sub-tasks that can be processed in this dimension), feature retrieval (retrieving corresponding visual, audio, text, or cross-modal features), rule execution (calculating scores according to the normalization and weighted fusion formula of this dimension), score interpretation (outputting evidence segments, timestamps, and diagnostic descriptions), and result verification (outputting confidence scores for subsequent weight reduction and review); the agent performs content type identification, determines the five-dimensional basic weight vector W_base based on the content type (w_vis^base, w_aud^base, w_rhy^base, w_sync^base, w_emo^base correspond to the basic weight coefficients of the five evaluation dimensions of visual aesthetics, audio aesthetics, rhythmic structure, audiovisual synchronization, and emotional semantic consistency, respectively), calculates modal saliency M_mod and segment complexity C_cplx, generates modal weight adjustment ΔW_modality(M_mod) and complexity weight adjustment ΔW_complexity(C_cplx), and performs the following formula:

[0126] W_i=norm(W_base+ΔW_modality(M_mod)+ΔW_complexity(C_cplx));

[0127] The system outputs an adaptive weight vector W_i={w_vis,w_aud,w_rhy,w_sync,w_emo}, and reduces the weight of low-confidence scores; the fine-grained preference comparison agent performs pairwise comparison set construction, outputs P_pref(U_i>U_j) based on preference model or rule table, calculates P_score(U_i>U_j) based on initial comprehensive score, and outputs difference descriptions; the comprehensive evaluation agent performs multi-dimensional scoring, fusion of weights and ranking results, conflict detection, and verification task generation; the verification agent performs the decision to recall relevant dimension analysis agents and adopt alternative scores.

[0128] Objective definition and source distinction of core parameters.

[0129] Furthermore, this application provides objective methods for obtaining core parameters and mathematical mappings, and distinguishes between existing parameters in the prior art and parameters improved or uniquely created in this solution:

[0130] (i) Existing basic feature parameters: lens boundary, key frame, motion intensity, composition features, color distribution, sharpness, loudness, beat, accent point, spectral centroid, timbre, background noise, etc., can be obtained by existing image processing, audio processing and lens detection methods. The embodiments of this application directly call and normalize to the range of 0 to 1.

[0131] (ii) Improved parameters in the embodiments of this application:

[0132] The shot type ST_i is obtained by mapping shot size changes, shot motion trajectory and composition features through shot classification ST_i=argmax_c p(c|shot size, motion, composition), and the shot type with the highest classification confidence is selected (such as close-up, medium shot, long shot, or moving shot).

[0133] The semantic function SF_i is obtained by mapping the subtitle / speech recognition text, event nodes, and shot positions on the timeline through semantic function classification SF_i=argmax_f p(f|text semantics, event, temporal position), and is mapped to functional categories such as opening, foreshadowing, climax, turning point, ending, information statement, or emotional expression.

[0134] The five-dimensional basic weight vector W_base={w_vis^base,w_aud^base,w_rhy^base,w_sync^base,w_emo^base} is determined by the content type and serves as the starting point for adaptive weight calculation. Among them, w_vis^base, w_aud^base, w_rhy^base, w_sync^base, and w_emo^base represent the basic weight coefficients of visual aesthetic score Q_vis, audio aesthetic score Q_aud, rhythmic structure score Q_rhy, audiovisual synchronization score Q_sync, and emotional semantic consistency score Q_emo in the comprehensive fusion, respectively.

[0135] Modal saliency M_mod is calculated from at least one normalized index among audio energy proportion, speech or music proportion, and visual motion intensity; segment complexity C_cplx is calculated from at least one normalized index among shot switching frequency, visual event density, audio event density, and text density.

[0136] The modality weight adjustment ΔW_modality(M_mod) and the complexity weight adjustment ΔW_complexity(C_cplx) are generated by M_mod and C_cplx through a preset rule mapping table or a normalization function, respectively, and are used to dynamically adjust W_base.

[0137] Its technical effect lies in transforming the original reliance on human experience in judging lens type, semantic function, and fixed weight settings into reproducible classification mapping and weight adjustment results.

[0138] (iii) Parameters unique to the embodiments of this application:

[0139] Sample-level similarity Similarity to fragments This is used to objectively construct similar audiovisual sample groups and comparable segment groups. Sim_topic, Sim_scene, Sim_text, Sim_audio, Sim_duration, s_pos, s_vstr, and s_astr are obtained by inverse normalization of the cosine similarity or duration difference of the corresponding feature vectors. s_ST and s_SF are determined by whether the shot type ST_i and semantic function SF_i of the segment pair to be compared are consistent.

[0140] The five-dimensional adaptive weight vector W_i={w_vis,w_aud,w_rhy,w_sync,w_emo} is calculated by W_i=norm(W_base+ΔW_modality(M_mod)+ΔW_complexity(C_cplx)), where w_vis, w_aud, w_rhy, w_sync, and w_emo are the dynamically corrected adaptive weight components of each evaluation dimension.

[0141] The calibrated score S_i' is used to fuse the absolute score and relative ranking; the fragment importance weight μ_i is used to determine the contribution of fragments to the total score by weighting with computable indicators.

[0142] Furthermore, the objective scoring mechanism for segment importance is as follows:

[0143] The importance weight μ_i of the segment is not dependent on subjective human judgment, but is obtained by weighting with computable indicators:

[0144] The segment duration percentage D_i is the ratio of the duration of segment i to the total duration of all L segments;

[0145] The visual and audio event density E_i is the ratio of the sum of the number of visual events and the number of audio events in segment i to the segment duration, normalized to the range of 0 to 1;

[0146] The proportion of the main subject or target object appearing in segment i is the proportion of the number of frames in which the main subject or target object appears in the segment out of the total number of frames in that segment.

[0147] The semantic criticality G_i is determined by the weighted normalized result of the number of topic keyword hits, the confidence of semantic inflection point detection, and the matching degree of plot nodes;

[0148] The impact degree B_i of the abnormal segment is determined by the weighted normalization result of the abnormal segment's deduction magnitude on the dimension score and the duration of the abnormality.

[0149] In one implementation, indicators such as lens scarcity and consistency with preceding and following segments can also be incorporated into the calculation of E_i and G_i as components or extended indicators of the above-mentioned components; when user interaction data is available, user attention can be incorporated into E_i as an optional component or participate in the calculation as an independent component.

[0150] final , norm represents the normalization process that makes the sum of the importance weights of each segment equal to 1, and ρ1 to ρ5 are non-negative coefficients.

[0151] Furthermore, the evaluation annotation and model validation mechanisms are as follows:

[0152] The feature index classifiers, normalization parameters, or scoring thresholds of the intelligent agent for each dimension are trained or calibrated using manually selected preference label data. During the evaluation phase, the feature calculation results are used as the scoring basis, and multiple checks are performed on the scoring results.

[0153] First, a consistency check is performed between P_score(U_i>U_j) and P_pref(U_i>U_j) output by the fine-grained preference comparison agent. If the difference exceeds a preset threshold, a review is triggered.

[0154] Secondly, cross-modal consistency verification is performed on visual emotion vectors, audio emotion vectors, and text semantic vectors;

[0155] Third, confidence is verified by calculating confidence scores based on comparisons with similar samples. Manually selected preference labels are used only as reference data for training, threshold calibration, or result verification; they do not directly replace scores calculated from features, thus ensuring that the scoring system is an objective technical process based on data, models, and formulas.

[0156] In this application embodiment, the technical problems of the prior art, such as the single dimension of audiovisual aesthetic evaluation, insufficient audio-visual collaborative modeling, difficulty in identifying fine-grained differences, subjectivity of key parameters, and insufficient interpretability of evaluation results, are addressed.

[0157] Furthermore, the step of constructing an audiovisual aesthetic evaluation task context based on the audiovisual content to be evaluated and its associated information includes the following steps:

[0158] Receive the audiovisual content to be evaluated;

[0159] Receive the associated information,

[0160] The audiovisual content to be evaluated is subjected to format standardization and quality preprocessing.

[0161] The associated information is structured to extract content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags;

[0162] Based on unique content identifiers and version identifiers, the audiovisual content to be evaluated (after standardized formatting or quality preprocessing), the related information after structured processing, the segment timeline, and feature indexes are integrated into an audiovisual aesthetic evaluation task context that can be invoked by multiple agents. This creates a one-to-one and traceable mapping relationship between each audiovisual content to be evaluated or each edit version and its related information and audiovisual aesthetic evaluation task context; among which,

[0163] The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

[0164] Furthermore, the step of aligning the video frame sequence, audio segment, subtitle text, and timestamp of the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units includes the following steps:

[0165] Shot boundary detection is performed on the video frame sequence of the audiovisual content to be evaluated to obtain multiple shot segments, and keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features, and sharpness features are extracted from each shot segment;

[0166] The audio segments of the audiovisual content to be evaluated are segmented to obtain multiple audio segments, and loudness, beat, accent, pitch variation, spectral centroid, energy variation, speech intelligibility, background noise, musical mood and timbre texture features are extracted.

[0167] Semantic analysis is performed on the subtitle text of the audiovisual content to be evaluated to obtain character speech, emotion words, event nodes, semantic turning points and thematic keywords;

[0168] Based on the timestamps of the audiovisual content to be evaluated, video shots, audio clips, and subtitle text are aligned to form segment-level audiovisual units.

[0169] Furthermore, the method also includes the following steps:

[0170] Output an evaluation report; among which,

[0171] The evaluation report includes the overall aesthetic score, scores corresponding to multiple evaluation dimensions, and the relative ranking calibration results.

[0172] Furthermore, the output evaluation report includes the following steps:

[0173] Output the overall aesthetic score, which is in at least one of the following forms: a percentage system, a ten-point system, a grade system, or a ranking system.

[0174] Output scores corresponding to multiple evaluation dimensions, including visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0175] Output the relative ranking calibration results, which include the overall ranking among multiple audiovisual versions, the dimensional difference of each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results.

[0176] Based on the technical solution of the embodiments of this application, there are multiple specific implementation cases in actual implementation, as follows:

[0177] Specific Implementation Case 1:

[0178] See Figure 1 This embodiment provides an automatic evaluation method for audiovisual aesthetics. This method is applicable to short videos, advertising videos, music videos, film clips, live stream segments, interview videos, instructional videos, documentary clips, social media audio and video content, and audiovisual content composed of image sequences and audio files. This method can be executed by an automatic audiovisual aesthetics evaluation device, which can be implemented in hardware, software, or a combination of both.

[0179] S101, construct an audiovisual aesthetic evaluation task context based on the audiovisual content to be evaluated and its associated information. The task context is used to provide unified input for multiple subsequent intelligent agents, including content type, evaluation target, topic tags, emotion tags, style tags, scene tags, rhythm tags, user target tags, and necessary technical metadata.

[0180] In the above process, the system generates a unique task context identifier for each audiovisual content or edit version to be evaluated, and saves the preprocessed content address, structured association information, segment timeline and feature index in the task context, so that the audiovisual content to be evaluated, the association information and the audiovisual aesthetic evaluation task context form a one-to-one correspondence according to the unique content identifier and the version identifier.

[0181] S102, the audiovisual content to be evaluated is aligned with video frame sequences, audio segments, subtitle text, and timestamps to obtain multiple segment-level audiovisual units. Each segment-level audiovisual unit is used to represent visual information, audio information, text information, and contextual metadata within a time segment.

[0182] S103, based on fragment-level audiovisual units, analyzes the agent across multiple dimensions to obtain visual aesthetic scores, audio aesthetic scores, rhythmic structure scores, audiovisual synchronization scores, and emotional semantic consistency scores. For each dimension, the agent outputs evidence fragments, timestamps, diagnostic descriptions, and confidence levels along with the scores.

[0183] S104, the agent determines the adaptive weights of each evaluation dimension based on content type, modal saliency, fragment complexity and evaluation objective through weight estimation.

[0184] S105 performs pairwise comparisons through a fine-grained preference comparison agent and performs relative ranking calibration based on the comparison results.

[0185] When multiple versions of similar audiovisual content exist, the system calculates the similarity of themes, scenes, text, audio, and duration separately, and then follows the... Obtain sample-level similarity; when the sample-level similarity is not lower than the preset similarity threshold within a group, the corresponding audiovisual content is classified into the same similar audiovisual sample group.

[0186] When there is only a single audiovisual content to be evaluated, the system determines the shot type, semantic function, rhythm position, visual structure and audio structure of the segment, and weights and fuses the similarities corresponding to the above features to obtain the segment-level similarity; when the segment-level similarity is not lower than the preset segment similarity threshold, the corresponding segment is classified into the same comparable segment group.

[0187] S106 integrates multi-dimensional scores, adaptive weights, and ranking calibration results of the intelligent agent through comprehensive evaluation, and calls the review agent for collaborative review when there are score conflicts.

[0188] S107, Output evaluation report.

[0189] Specific Implementation Case 2:

[0190] See Figure 2 In one implementation, the system receives audiovisual content to be evaluated. The audiovisual content to be evaluated can be video files, video links, live stream clips, short drama segments, advertising clips, music videos, film clips, interview videos, instructional videos, documentary clips, social media audio and video content, or audiovisual content formed by combining image sequences and audio files.

[0191] The system further receives related information. This related information may include one or more of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags.

[0192] The system performs format standardization and quality preprocessing on the audiovisual content to be evaluated, including video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking. The system also performs structured processing on related information, extracting content type tags, theme tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags. Based on unique content identifiers and version identifiers, the system integrates the preprocessed audiovisual content, structured related information, segment timelines, and feature indexes to form the audiovisual aesthetic evaluation task context, creating a one-to-one and traceable mapping relationship among the three.

[0193] Specific Implementation Case 3:

[0194] See Figure 3In one implementation, the system performs shot boundary detection on the video to obtain multiple shot segments, and extracts keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features, and sharpness features from each shot segment.

[0195] The system segments audio into multiple segments and extracts features such as loudness, tempo, accents, pitch variations, spectral centroid, energy variations, speech intelligibility, background noise, musical emotion, and timbre. The system also performs semantic analysis on subtitles or speech-recognized text to obtain speech characters, emotion words, event nodes, semantic transition points, and key themes.

[0196] Based on timestamps, the system aligns video shots, audio clips, and subtitle text to form segment-level audiovisual units U_i={F_i,A_i,T_i,M_i}. Here, F_i represents the video frame sequence and visual features, A_i represents the audio signal and audio features, T_i represents the subtitle text or speech-recognized text, and M_i represents the timestamp, shot boundaries, content type, and contextual metadata. When grouping single videos, the shot type is determined by ST_i=argmax_cp(c|shot type, motion, composition), and the semantic function is determined by SF_i=argmax_fp(f|text semantics, event, temporal position). For the segment pair (U_i, U_j) to be compared, s_ST is 1 when ST_i and ST_j are the same, and 0 otherwise. s_SF is 1 when SF_i and SF_j are the same, and 0 otherwise. s_pos, s_vstr, and s_astr are obtained by the cosine similarity of the normalized feature vectors of rhythm position, visual structure, and audio structure, respectively. The segment-level similarity Sim_seg is then obtained by weighting s_ST, s_SF, s_pos, s_vstr, and s_astr.

[0197] Specific Implementation Case 4:

[0198] See Figure 4 The visual aesthetics analysis agent reads the video frame sequence and visual features in the segment-level audiovisual unit, calculates the fusion results of normalized indicators such as picture composition, subject prominence, color coordination, light and shadow layers, camera movement, picture stability, visual style consistency C_style, and visual anomaly inverse index C_attr, and obtains the visual aesthetics score Q_vis.

[0199] The audio aesthetics analysis agent reads audio segments and features, calculates the fusion results of normalized indicators such as loudness stability A_loud, timbre quality, noise suppression A_noise, musicality, speech intelligibility, dynamic range A_dyn, and spatial awareness, and obtains the audio aesthetics score Q_aud. The rhythm structure analysis agent jointly reads video shot boundaries, motion intensity changes, music beats, and energy changes, evaluates the matching degree between shot transitions, action rhythms, music rhythms, and content progression, and obtains the rhythm structure score Q_rhy.

[0200] The audiovisual synchronization analysis agent analyzes the correspondence between visual event time points and audio event time points to obtain the audiovisual synchronization score Q_sync. The emotion semantic consistency analysis agent extracts the visual emotion vector e_v,i, the audio emotion vector e_a,i, and the text semantic emotion vector e_t,i, and calculates the cosine consistency of the three-modal emotion vectors to obtain the emotion semantic consistency score Q_emo.

[0201] In one implementation, the visual aesthetics analysis agent extracts and normalizes the following indicators for each segment-level audiovisual unit: composition rule matching degree, subject prominence, color harmony, light and shadow hierarchy, camera movement smoothness, image stability, visual style consistency, and visual anomaly inverse indicators. It then calculates a visual aesthetics score based on a formula, the specific formula for which is:

[0202] .

[0203] In one implementation, the audio aesthetics analysis agent extracts and normalizes loudness stability, timbre smoothness, noise suppression, music layering, speech intelligibility, dynamic range reasonableness, and sound-image or dual-channel spatial consistency for each segment-level audiovisual unit, and calculates an audio aesthetics score according to the following formula:

[0204] .

[0205] In one implementation, the rhythm structure analysis agent calculates the matching degree between the shot switching interval and the target rhythm template, the coordination degree between the motion intensity change and the shot switching, the matching degree between the visual event and the musical beat or downbeat point, and the consistency between the semantic turning point of the subtitle, the event node and the content progression speed for each segment-level audiovisual unit, and calculates the rhythm structure score according to the formula as follows:

[0206] .

[0207] In one implementation, the audiovisual synchronization score and the emotion semantic consistency score can be calculated using the following formulas:

[0208] The formula for calculating the audiovisual synchronization score is as follows:

[0209] .

[0210] The formula for calculating the emotion semantic consistency score is as follows:

[0211] .

[0212] Specific Implementation Case 5:

[0213] See Figure 5 The weight estimation agent reads the audiovisual aesthetics assessment task context, segment-level audiovisual units, and the preliminary scoring results of the agent analyzing each dimension, and determines the basic weights based on the content type. The details are as follows:

[0214] W_base={w_vis^base,w_aud^base,w_rhy^base,w_sync^base,w_emo^base}; where,

[0215] w_vis^base, w_aud^base, w_rhy^base, w_sync^base, and w_emo^base represent the basic weight coefficients of the five evaluation dimensions: visual aesthetics, audio aesthetics, rhythmic structure, audiovisual synchronization, and emotional semantic consistency, respectively.

[0216] For music videos, w_aud^base, w_rhy^base, and w_sync^base in W_base are greater than the arithmetic mean of the remaining components of W_base;

[0217] For interview videos, w_aud^base, w_emo^base, and w_vis^base in W_base are greater than the arithmetic mean of the other components of W_base;

[0218] For ad videos, w_vis^base, w_emo^base, and w_rhy^base in W_base are greater than the arithmetic mean of the other components of W_base;

[0219] For re-edited videos, mixed videos, or multiple versions of footage, w_vis^base, w_sync^base, w_rhy^base, and w_emo^base in W_base are greater than the arithmetic mean of the other components of W_base.

[0220] The weight estimation agent also calculates the modal saliency M_mod and fragment complexity C_cplx, and adjusts the weights accordingly; among them,

[0221] M_mod is calculated from at least one normalized index among audio energy proportion, speech or music proportion, and visual motion intensity.

[0222] C_cplx is calculated from at least one normalized metric among shot switching frequency, visual event density, audio event density, and text density.

[0223] The weight estimation agent generates modality weight adjustment ΔW_modality(M_mod) based on M_mod using a preset rule mapping table or normalization function, and generates complexity weight adjustment ΔW_complexity(C_cplx) based on C_cplx using a preset rule mapping table or normalization function.

[0224] Finally, follow the formula below:

[0225] W_i=norm(W_base+ΔW_modality(M_mod)+ΔW_complexity(C_cplx));

[0226] The adaptive weight vector is obtained, i.e., W_i={w_vis,w_aud,w_rhy,w_sync,w_emo}, where,

[0227] w_vis, w_aud, w_rhy, w_sync, and w_emo are the adaptive weight components of the segment in the dimensions of visual aesthetics, audio aesthetics, rhythmic structure, audiovisual synchronization, and emotional semantic consistency, respectively, and the sum of each component is 1. When the score difference between different agents exceeds a preset threshold, the weight estimation agent combines the confidence level and the evidence segment to reduce the weight of low-confidence scores.

[0228] Specific Implementation Case Six:

[0229] See Figure 6 The system constructs pairwise comparison sets based on similar audiovisual sample groups or comparable segment groups. Similar audiovisual sample groups can correspond to multiple edited versions of the same advertising video, multiple post-production versions of the same original material, different packaging versions of the same program segment, multiple candidate segments under the same theme, or multiple audiovisual segments with similar content in the same video collection. The relationship within the group is determined by weighted similarity based on theme, scene, text, audio, and duration.

[0230] Comparable segment groups can correspond to segments within the same video that have similar shot types, semantic functions, rhythmic positions, visual structures, or audio structures. The relationships within these groups are determined by the comprehensive similarity of the aforementioned segment features.

[0231] The fine-grained preference comparison agent compares each audiovisual unit pair (U_i, U_j), inputting ΔQ_vis, ΔQ_aud, ΔQ_rhy, ΔQ_sync, ΔQ_emo, ΔS_ij, and the evidence difference vector into the preference model or rule table to obtain the preference probability P_pref (U_i>U_j), the comparison basis, and the difference description; where,

[0232] ΔQ_vis, ΔQ_aud, ΔQ_rhy, ΔQ_sync, and ΔQ_emo represent the differences between U_i and U_j in the corresponding dimension scores, respectively.

[0233] ΔS_ij represents the difference between S_i and S_j, i.e., S_i−S_j. The evidence difference vector is composed of the differences in feature values, timestamps, or question types of the evidence fragments between the two.

[0234] The system also combines multi-dimensional ratings and the adaptive weight vector W_i of segment i to calculate the initial comprehensive score S_i, and calculates the pairwise rating preference probability P_score (U_i>U_j) based on the initial comprehensive score.

[0235] When P_pref(U_i>U_j) is not lower than the preset preference threshold, the system determines the ranking relationship of U_i relative to U_j;

[0236] When P_pref(U_i>U_j) is lower than the preset preference threshold, the system does not determine that U_i has a priority relationship with U_j.

[0237] When the difference between P_pref(U_i>U_j) and P_score(U_i>U_j) exceeds a preset threshold, the system triggers a review agent to re-examine the visual, audio, and cross-modal evidence for the corresponding segment. The system forms a global ranking based on the pairwise comparison results and performs ranking calibration on the initial composite score to obtain the calibrated score S_i'.

[0238] S_i=w_vis(i)·Q_vis(i)+w_aud(i)·Q_aud(i)+w_rhy(i)·Q_rhy(i)+w_sync(i)·Q_sync(i)+w_emo(i)·Q_emo(i);

[0239] P_score(U_i>U_j)=exp(S_i) / (exp(S_i)+exp(S_j));

[0240] .

[0241] Specific Implementation Case Seven:

[0242] See Figure 7 The intelligent agent is comprehensively evaluated based on visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, emotional semantic consistency score, adaptive weight vector, and sorting calibration results.

[0243] The system calculates the overall aesthetic score Score_AV of the audiovisual content to be evaluated based on the calibrated score, and further calculates the weighted overall score Score_AV* based on the importance weight of the segments.

[0244] The importance weight of a segment is determined by the segment duration ratio, the density of visual and audio events, the proportion of the subject or target object, the semantic keyness, and the impact of abnormal segments on the overall evaluation results.

[0245] When conflicts arise between scores across different dimensions, the system invokes a review agent for collaborative review. Conflict scenarios include at least one of the following: visual aesthetics score not lower than a first preset threshold and audiovisual synchronization score lower than a second preset threshold; audio aesthetics score not lower than a third preset threshold and emotional semantic consistency score lower than a fourth preset threshold; overall aesthetics score not lower than a fifth preset threshold and ranking preference probability lower than a sixth preset threshold; and segment-level scores varying between adjacent segments beyond a seventh preset threshold.

[0246] The comprehensive evaluation agent generates a review task based on the conflict type. The review task carries conflict fragments, original dimension scores, evidence fragments, feature values, confidence levels, and ranking preference results. The review agent then calls upon the relevant dimension analysis agent again on the same fragment-level audiovisual unit to obtain review scores, review evidence, and alternative scores.

[0247] When the difference between the alternative score and the original score exceeds the review threshold and the confidence level of the alternative score is higher than that of the original score, the comprehensive evaluation agent uses the alternative score to update the corresponding dimension score, weight, or ranking calibration result.

[0248] Otherwise, retain the original score and record the reason for the conflict. The final evaluation results can be written to the evaluation database to support subsequent retrieval of similar samples, preference comparison, and quality review of similar content.

[0249] ;

[0250] ;

[0251] .

[0252] Specific Implementation Case Eight:

[0253] See Figure 8The evaluation report output module outputs an overall audiovisual aesthetic score. This overall score can be in at least one of the following formats: percentage, 10-point scale, graded scale, or ranking scale. The evaluation report also outputs dimensional scores, including visual aesthetic score, audio aesthetic score, rhythmic structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0254] The evaluation report further outputs segment-level diagnostic results. These results include dimensionality differences, low-score thresholds, evidence segment indexes, impact magnitude values, and suggested modification direction codes or text; the evidence segment index includes at least start and end timestamps, keyframe indexes, and audio segment indexes. For multiple audiovisual versions, the evaluation report outputs overall ranking, pairwise preference comparison results, dimensionality differences for each version, and ranking calibration contributions. For a single video, the evaluation report outputs aesthetic quality ranking of comparable segment groups, low-score segment thresholds, and problem segment location results.

[0255] Machine-readable evaluation data can be in the form of JSON files, tabular files, database records, or data returned by an API. Content management platforms, video filtering systems, video quality inspection systems, content review systems, or recommendation systems can directly access the evaluation results of this invention using this data.

[0256] Specific Implementation Case Nine:

[0257] See Figure 9 The audiovisual aesthetics automatic evaluation device includes a task context construction module, a segment-level audiovisual unit construction module, a multi-dimensional scoring module, an adaptive weight estimation module, a preference comparison and ranking calibration module, a comprehensive evaluation and review module, and an evaluation report output module.

[0258] The task context construction module is used to construct the audiovisual aesthetic evaluation task context based on the audiovisual content to be evaluated and its associated information. The segment-level audiovisual unit construction module is used to perform video, audio, text, and timestamp alignment processing on the audiovisual content to be evaluated. The multi-dimensional scoring module is used to call a multi-dimensional analysis agent to calculate the scores, evidence, and confidence levels for each dimension according to the formulas described in Implementation Case 4.

[0259] The adaptive weight estimation module is used to determine the adaptive weights W_i for each evaluation dimension according to specific implementation case five.

[0260] The preference comparison and ranking calibration module is used in Implementation Case 3 and Implementation Case 6 to calculate fine-grained preference results and relative ranking results using Sim_sample, Sim_seg, P_pref(U_i>U_j), P_score(U_i>U_j) and Si'.

[0261] The comprehensive evaluation and review module is used to determine and integrate various intermediate results and perform score conflict review as described in Case Study 7, based on the Score_AV, Score_AV*, μ_i and review threshold.

[0262] The evaluation report output module is used to output readable reports and machine-readable data. Each module performs the Sim_sample, Sim_seg, Q, W_i, Si', and μ_i calculations and verifies the threshold judgments as described in Implementation Cases 3 to 7.

[0263] Specific Implementation Case 10:

[0264] In the following embodiments, w_vis^base, w_aud^base, w_rhy^base, w_sync^base, and w_emo^base represent the basic weight coefficients of the five evaluation dimensions: visual aesthetics, audio aesthetics, rhythmic structure, audiovisual synchronization, and emotional semantic consistency, respectively.

[0265] In one embodiment, the audiovisual content to be evaluated is a 60-second music video clip. The system first performs video and audio decoding on the music video clip, extracting the video frame sequence, music beats, shot transitions, and subtitle text. Subsequently, the system divides the video into multiple segment-level audiovisual units and calls upon visual aesthetics analysis agents, audio aesthetics analysis agents, rhythmic structure analysis agents, audiovisual synchronization analysis agents, and emotion semantic consistency analysis agents to perform scoring for each segment.

[0266] The weight estimation agent identifies the content as a music video and determines that w_aud^base, w_rhy^base, and w_sync^base in W_base are greater than the arithmetic mean of the remaining components of W_base.

[0267] Calculate M_mod and C_cplx and generate ΔW_modality(M_mod) and ΔW_complexity(C_cplx).

[0268] The adaptive weight vector is obtained by using W_i=norm(W_base+ΔW_modality(M_mod)+ΔW_complexity(C_cplx)).

[0269] The evaluation report outputs W_i, each Q_*, S_i, μ_i, low-score segment threshold, and corresponding evidence segments. The evidence segments include start and end timestamps, keyframe indexes, and audio segment indexes.

[0270] In another embodiment, the objects to be evaluated are three edited versions of the same advertising video. The system constructs segment-level audiovisual units for each of the three versions, and then constructs similar audiovisual sample groups based on theme, scene, text, duration, and audio structure.

[0271] The fine-grained preference comparison agent performs pairwise comparisons of corresponding segments in different versions, determines the ranking relationship between versions based on P_pref(U_i>U_j) and a preset preference threshold, and determines the overall audiovisual aesthetic ranking by combining P_score(U_i>U_j), S_i' and the dimensional differences of each version.

[0272] The system then performs sorting calibration and outputs the overall audiovisual aesthetic sorting of the three versions, Sim_sample, P_pref(U_i>U_j), P_score(U_i>U_j), Si', and the dimensional differences between each version.

[0273] In another embodiment, the object to be evaluated is a video interview. The system identifies the content type as an interview video, and the weight estimation agent determines that w_aud^base, w_emo^base, and w_vis^base in W_base are greater than the arithmetic mean of the remaining components of W_base.

[0274] Calculate M_mod and C_cplx and obtain the adaptive weight vector according to W_i=norm(W_base+ΔW_modality(M_mod)+ΔW_complexity(C_cplx)).

[0275] If the audio aesthetic analysis agent detects that the background noise level exceeds the preset noise threshold and the subtitle recognition confidence is lower than the preset confidence threshold, the comprehensive evaluation agent will reduce the weight of the low confidence score and output Wi, Qaud, Qemo, low confidence weight reduction result, diagnostic description and corresponding timestamp in the evaluation report. The diagnostic description includes the background noise exceeding the preset noise threshold, the speech clarity being lower than the preset clarity threshold and the numerical value of its impact.

[0276] Specific Implementation Case Eleven:

[0277] This specific implementation case further illustrates the complete technical chain from data input to result output. After receiving video clips, the system extracts shot type features from the shot sequence using a visual model and maps shot types according to ST_i=argmax_c p(c|shot type, motion, composition). A language model or multimodal model extracts semantic function features from subtitles or speech recognition text and maps semantic functions according to SF_i=argmax_f p(f|text semantics, event, temporal position). Then, visual, audio, and text features are mapped to a unified representation space, and the similarity between clips or between clips and the target rhythm template is calculated. Subsequently, the dimensional analysis agents obtain dimensional scores according to formulas, and the comprehensive evaluation agent weights and fuses them according to the importance of the clips μ_i, ultimately generating an overall aesthetic score, dimensional scores, and machine-readable data. Each step in this chain—"input—feature extraction—semantic recognition—similarity calculation—score fusion—result output"—has clearly defined inputs, processing, and outputs, facilitating reproduction and review.

[0278] Specific Implementation Case Twelve:

[0279] This specific implementation case illustrates one approach to evaluation labeling and model validation. During the training phase, manually labeled data is used to train or calibrate the feature index classifiers and scoring thresholds for each dimension's analytical agent. For example, pairwise preference labels constrain the preference probability output of the fine-grained preference comparison agent, and dimensional scoring labels calibrate the normalization parameters for each dimension. During the evaluation phase, the feature calculation results are used as the scoring basis, and consistency checks are performed: when the difference between P_score(U_i>U_j) and P_pref(U_i>U_j) output by the fine-grained preference comparison agent exceeds a preset threshold, a review agent is triggered for review; simultaneously, cross-modal consistency checks are performed on visual emotion vectors, audio emotion vectors, and text semantic vectors, and confidence checks are performed on comparison results with similar samples. The manually labeled data is always used only as reference data for training, threshold calibration, or review, and does not directly replace the scores obtained from feature calculations.

[0280] Furthermore, the embodiments of this application also provide the following specific implementation methods to further illustrate the method described in the first aspect:

[0281] The audiovisual aesthetics assessment task context is constructed based on the audiovisual content to be evaluated and its associated information; the data sources of the audiovisual content to be evaluated and its associated information include video frame sequences, shot sequences, audio signals, subtitle text or speech recognition text, as well as semantic tags such as content type tags and topic tags, and also include human preference tags when there is existing labeled data;

[0282] The audiovisual content to be evaluated is processed by aligning video frame sequences, audio segments, subtitle text, and timestamps to obtain multiple segment-level audiovisual units;

[0283] Based on the aforementioned segment-level audiovisual unit, the intelligent agent is analyzed through multiple dimensions, including feature extraction, semantic recognition, index normalization, and weighted fusion, to obtain visual aesthetics scores, audio aesthetics scores, rhythmic structure scores, audiovisual synchronization scores, and emotional semantic consistency scores.

[0284] The agent determines adaptive weights for each evaluation dimension based on content type, modal saliency, fragment complexity, and evaluation objective by weighting estimation.

[0285] A fine-grained preference comparison agent performs pairwise comparisons of similar audiovisual sample groups or comparable segment groups, and performs relative ranking calibration based on the comparison results;

[0286] The visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, and emotional semantic consistency score are obtained by normalizing, weighting, and calculating the confidence level of the visual features, audio features, rhythmic features, cross-modal synchronization features, and emotional semantic features in the corresponding segment-level audiovisual units. These scores are computable based on data, models, and formulas. When there are artificial preference labels, these labels are used for model training, threshold calibration, or result verification, and are not used as the final score to replace the above feature-based score calculation.

[0287] By comprehensively evaluating the intelligent agent, the multi-dimensional scores, adaptive weights, and ranking calibration results are fused to form an overall audiovisual aesthetic score and dimensional diagnostic results.

[0288] Output an evaluation report, which includes an overall aesthetic score, scores for each dimension, segment-level problem identification, relative ranking results, improvement suggestions, and machine-readable evaluation data.

[0289] Furthermore, the step of constructing the audiovisual aesthetic evaluation task context based on the audiovisual content to be evaluated and its associated information includes:

[0290] The audiovisual content to be evaluated is received, which includes at least one of the following: video files containing video tracks and audio tracks, video links, live broadcast clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files.

[0291] The associated information is received, which includes at least one of the following: content type, evaluation target, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, editing version identifier, content source, reference sample, and human preference label;

[0292] The audiovisual content to be evaluated is subjected to format unification and quality preprocessing. The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

[0293] The associated information is structured to extract content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags;

[0294] Based on unique content identifiers and version identifiers, the audiovisual content to be evaluated after unified formatting or quality preprocessing, the related information after structured processing, the segment timeline, and feature index are integrated into an audiovisual aesthetic evaluation task context that can be called by multiple intelligent agents. This enables each audiovisual content to be evaluated or each edit version to form a one-to-one and traceable mapping relationship with its related information and audiovisual aesthetic evaluation task context.

[0295] Furthermore, the audiovisual content to be evaluated undergoes video frame sequence, audio segment, subtitle text, and timestamp alignment processing to obtain multiple segment-level audiovisual units, including:

[0296] Shot boundary detection is performed on the video to obtain multiple shot segments, and keyframes, representative frame sequences, motion intensity features, subject region features, composition features, color distribution features, and sharpness features are extracted from each shot segment;

[0297] The audio is segmented to obtain multiple audio segments, and features such as loudness, beat, accent, pitch variation, spectral centroid, energy variation, speech intelligibility, background noise, musical mood, and timbre texture are extracted.

[0298] Semantic analysis is performed on subtitle text or speech recognition text to obtain character speech, emotion words, event nodes, semantic turning points and thematic keywords;

[0299] The video shots, audio clips, and subtitle text are aligned based on timestamps to form a segment-level audiovisual unit U_i={F_i,A_i,T_i,M_i}; where F_i represents the video frame sequence and visual features corresponding to the segment, A_i represents the audio signal and audio features corresponding to the segment, T_i represents the subtitle text or speech recognition text corresponding to the segment, and M_i represents the timestamp, shot boundary, content type, and contextual metadata of the segment.

[0300] When multiple versions of similar audiovisual content exist, the topic similarity Sim_topic is calculated based on the semantic vector cosine similarity of the topic tags or topic keywords; the scene similarity Sim_scene is calculated based on the feature vector similarity of scene tags, main area features, composition features, and color distribution features; the text similarity Sim_text is calculated based on the semantic vector cosine similarity of subtitle text or speech recognition text; the audio similarity Sim_audio is calculated based on the feature vector similarity of beat, energy change, spectral centroid, timbre texture, and musical emotion features; and the duration similarity Sim_duration is calculated based on the reverse normalization result of the video duration difference. Each similarity is normalized to the range of 0 to 1.

[0301] The above similarity scores are weighted and fused to obtain the sample-level similarity. , where λ1 to λ5 are non-negative weights and their sum is 1;

[0302] When Sim_sample is not lower than the preset similarity threshold within the group, the corresponding audiovisual content will be classified into the same similar audiovisual sample group.

[0303] When there is only a single audiovisual content to be evaluated, the shot type ST_i is determined through shot classification mapping based on shot boundary detection results, shot size changes, and shot motion characteristics. The shot classification mapping takes the shot category that maximizes the classification confidence, i.e., ST_i = argmax_c p(c|shot size, motion, composition). The semantic function SF_i is determined through semantic function classification mapping based on subtitle text, speech recognition text, or event nodes. The semantic function classification mapping takes the function category that maximizes the classification confidence, i.e., SF_i = argmax_f p(f|text semantics, event, temporal position). The rhythmic position is determined based on the normalized position, beat point, and semantic inflection point of the segment in the overall timeline. The visual structure is determined based on the main area features, composition features, color distribution features, and motion intensity features. The audio structure is determined based on loudness, beat, energy change, spectral centroid, and timbre texture features. For the segment-level audiovisual unit pair (U_i, U_j) to be compared, the shot type consistency s_ST is determined when ST_i and ST_j are the same. Set 1 if SF_i is the same as SF_j, otherwise set 0. The semantic functional consistency s_SF is 1 if SF_i is the same as SF_j, otherwise set 0. The rhythm position proximity s_pos, visual structure similarity s_vstr, and audio structure similarity s_astr are obtained by the cosine similarity of the normalized feature vectors of the corresponding rhythm position, visual structure, and audio structure, respectively.

[0304] The fragment-level similarity is obtained by weighted fusion of s_ST, s_SF, s_pos, s_vstr, and s_astr. , where θ1 to θ5 are non-negative weights and their sum is 1, and each consistency or similarity is normalized to the range of 0 to 1;

[0305] When Sim_seg is not lower than the preset segment similarity threshold, the corresponding segments are classified into the same comparable segment group.

[0306] Furthermore, the multi-dimensional analysis agents include visual aesthetics analysis agents, audio aesthetics analysis agents, rhythmic structure analysis agents, audiovisual synchronization analysis agents, and emotion semantic consistency analysis agents;

[0307] The aforementioned visual aesthetic analysis agent is used to read the following indicators from the visual features of fragment-level audiovisual units: composition rule matching degree C_comp, subject prominence C_subj, color harmony C_color, light and shadow hierarchy C_light, camera motion smoothness C_motion, image stability C_stab, visual style consistency C_style, and visual anomaly inverse index C_attr, and normalizes each indicator to the range of 0 to 1.

[0308] according to Calculate the visual aesthetic score of the i-th segment-level audiovisual unit.

[0309] C_attr is determined by sharpness, occlusion rate and abnormal image ratio, and α1 to α8 are non-negative weights that sum to 1.

[0310] The audio aesthetic analysis agent is used to read loudness stability A_loud, timbre smoothness A_timbre, noise suppression A_noise, music level index A_music, speech intelligibility A_speech, dynamic range reasonableness A_dyn, and sound image or binaural spatial consistency A_space from the audio features of the segment-level audiovisual units, and normalize each index to the range of 0 to 1.

[0311] according to Calculate the audio aesthetic score of the i-th segment-level audiovisual unit, where A_noise is obtained by inverse normalization of the background noise level, and β1 to β7 are non-negative weights that sum to 1.

[0312] The rhythm structure analysis agent is used to jointly read video shot boundaries, motion intensity changes, music beats, accent points, and energy changes. It calculates the matching degree R_cut between shot transition intervals and the target rhythm template, the coordination degree R_motion between motion intensity changes and shot transitions, the matching degree R_beat between visual events and music beats or accent points, and the consistency degree R_progress between subtitle semantic turning points, event nodes, and content progression speed. Calculate the rhythmic structure score of the i-th segment-level audiovisual unit, where γ1 to γ4 are non-negative weights and their sum is 1.

[0313] The audiovisual synchronization analysis agent is used to analyze the correspondence between visual event time points and audio event time points, according to... Calculate the audiovisual synchronization score Q_sync for the i-th segment-level audiovisual unit, where t_v,k represents the occurrence time of the k-th visual event in segment i, t_a,j represents the occurrence time of the j-th audio event in segment i, τ represents the time tolerance parameter, and N_i represents the number of visual events in segment i.

[0314] The emotion semantic consistency analysis agent is used to extract visual emotion vectors, audio emotion vectors, and text semantic vectors, respectively.

[0315] according to Calculate the emotional semantic consistency score Q_emo for the i-th segment-level audiovisual unit, where e_v,i, e_a,i, and e_t,i represent the visual emotion vector, audio emotion vector, and textual semantic emotion vector of segment i, respectively, cos(·) represents the cosine similarity function, and k1, k2, and k3 are weight coefficients that satisfy k1 + k2 + k3 = 1;

[0316] The multi-dimensional analysis agents are also used to output evidence fragments, timestamps, diagnostic descriptions, and confidence levels;

[0317] The evidence fragments are audiovisual units at the segment level whose contribution to the corresponding dimension's scoring weight exceeds a preset threshold or leads to a deduction. They include at least start and end timestamps, keyframe indexes, audio segment indexes, and corresponding feature values. The diagnostic descriptions are generated based on indicators below a preset threshold, abnormal detection results, or cross-modal inconsistency results. The confidence level is calculated based on input quality, feature completeness, model output stability, and consistency with similar samples.

[0318] Furthermore, each dimension analysis agent includes an input parsing unit, a feature reading unit, a scoring calculation unit, and an evidence output unit. The input parsing unit is used to read segment-level audiovisual units and decompose them into sub-tasks that can be processed in this dimension. The feature reading unit is used to call the corresponding video features, audio features, text features, or cross-modal features. The scoring calculation unit is used to calculate and output the score and confidence level of the corresponding dimension according to the normalization and weighted fusion rules of this dimension. The evidence output unit is used to output the scoring basis, problem description, and corresponding timestamp.

[0319] Furthermore, the audiovisual synchronization score is calculated for each segment-level audiovisual unit. The audiovisual synchronization score Q_sync(i) of the i-th segment-level audiovisual unit is calculated as follows:

[0320]

[0321] Where t_v,k represents the time point of occurrence of the k-th visual event in segment i, t_a,j represents the time point of occurrence of the j-th audio event in segment i, τ represents the time tolerance parameter, and N_i represents the number of visual events in segment i.

[0322] Furthermore, the emotion semantic consistency score is calculated for each segment-level audiovisual unit. The emotion semantic consistency score Q_emo(i) of the i-th segment-level audiovisual unit is calculated as follows:

[0323]

[0324] Where e_v,i represents the visual emotion vector of segment i, e_a,i represents the audio emotion vector of segment i, e_t,i represents the textual semantic emotion vector of segment i, cos(·) represents the cosine similarity function, and k1, k2, k3 are weight coefficients, satisfying k1+ k2+ k3=1.

[0325] The evidence fragments output by the agent for each dimension are used to explain the source of the score, including start and end timestamps, keyframe indexes, audio segment indexes, feature values ​​that trigger the score or deduction, and the corresponding dimension names;

[0326] Diagnostic descriptions are generated based on indicators below a threshold, abnormal detection results, or cross-modal inconsistencies.

[0327] The confidence score is calculated based on input quality, feature completeness, model output stability, and consistency with similar samples.

[0328] Furthermore, the step of determining the adaptive weights for each evaluation dimension through a weight estimation agent includes:

[0329] Read the initial scoring results of the audiovisual aesthetics assessment task context, segment-level audiovisual units, and analytical agents for each dimension;

[0330] The basic weight W_base is determined based on the content type, where W_base is the five-dimensional basic weight vector for the corresponding content type: W_base={w_vis^base,w_aud^base,w_rhy^base,w_sync^base,w_emo^base}.

[0331] w_vis^base, w_aud^base, w_rhy^base, w_sync^base, and w_emo^base represent the basic weight coefficients of the five evaluation dimensions—visual aesthetics, audio aesthetics, rhythmic structure, audiovisual synchronization, and emotional semantic consistency—in the initial comprehensive score fusion, and each component is a non-negative number.

[0332] For music videos, w_aud^base, w_rhy^base, and w_sync^base in W_base are greater than the arithmetic mean of the other components of W_base;

[0333] For interview videos, w_aud^base, w_emo^base, and w_vis^base in W_base are greater than the arithmetic mean of the other components of W_base;

[0334] For ad videos, w_vis^base, w_emo^base, and w_rhy^base in W_base are greater than the arithmetic mean of the other components of W_base;

[0335] For re-edited videos, mixed videos, or multiple versions of material, w_vis^base, w_sync^base, w_rhy^base, and w_emo^base in W_base are greater than the arithmetic mean of the other components of W_base;

[0336] Calculate the modal saliency M_mod and the fragment complexity C_cplx; where M_mod is calculated by at least one normalized index among audio energy proportion, speech or music proportion, and visual motion intensity, and C_cplx is calculated by at least one normalized index among shot switching frequency, visual event density, audio event density, and text density.

[0337] Based on M_mod, the modal weight adjustment ΔW_modality(M_mod) is generated through a preset rule mapping table or a normalization function, and based on C_cplx, the complexity weight adjustment ΔW_complexity(C_cplx) is generated through a preset rule mapping table or a normalization function.

[0338] The adaptive weight vector W_i={w_vis,w_aud,w_rhy,w_sync,w_emo} for each evaluation dimension is obtained by following the formula W_i=norm(W_base+ΔW_modality(M_mod)+ΔW_complexity(C_cplx)).

[0339] Where w_vis, w_aud, w_rhy, w_sync, and w_emo are the adaptive weight components of segment i in the dimensions of visual aesthetics, audio aesthetics, rhythmic structure, audiovisual synchronization, and emotional semantic consistency, respectively. They can also be denoted as w_vis(i), w_aud(i), w_rhy(i), w_sync(i), and w_emo(i), and the sum of each component is 1.

[0340] When the score difference between different agents exceeds a preset threshold, the low confidence score is downweighted by combining confidence level and evidence fragments.

[0341] Furthermore, the step of performing pairwise comparisons of similar audiovisual sample groups or comparable segment groups through a fine-grained preference comparison agent, and performing relative ranking calibration based on the comparison results, includes:

[0342] Based on the similar audiovisual sample groups or comparable segment groups, a pairwise comparison set is constructed, wherein the pairwise comparison set includes multiple audiovisual unit pairs (U_i, U_j);

[0343] For each audiovisual unit i, the initial comprehensive score is calculated by combining the multi-dimensional scores and the adaptive weight vector W_i of segment i. ,

[0344] Where w_vis(i), w_aud(i), w_rhy(i), w_sync(i), and w_emo(i) are the corresponding components of W_i;

[0345] By inputting ΔQ_vis, ΔQ_aud, ΔQ_rhy, ΔQ_sync, ΔQ_emo, ΔS_ij and the evidence difference vector into the preference model or rule table through a fine-grained preference comparison agent, the preference probability P_pref (U_i>U_j), comparison basis, and difference description are obtained.

[0346] Where ΔQ_vis, ΔQ_aud, ΔQ_rhy, ΔQ_sync, and ΔQ_emo represent the difference between U_i and U_j in the corresponding dimension scores, respectively, and ΔS_ij represents the difference between S_i and S_j, i.e., S_i−S_j. The evidence difference vector is composed of the difference in feature values, timestamps, or question types of the evidence fragments between the two.

[0347] The pairwise rating preference probability P_score(U_i>U_j)=exp(S_i) / (exp(S_i)+exp(S_j)) is calculated based on the initial comprehensive score, and the ranking relationship of U_i relative to U_j is determined when P_pref(U_i>U_j) is not lower than the preset preference threshold.

[0348] When the difference between P_pref(U_i>U_j) and P_score(U_i>U_j) exceeds a preset threshold, the review agent is triggered to re-examine the visual evidence, audio evidence and cross-modal evidence of the corresponding segment.

[0349] A global ranking of audiovisual units is formed based on the pairwise comparison results, and the initial comprehensive scores are then calibrated to rank them.

[0350] Furthermore, the calibrated score obtained from the sorting calibration is:

[0351]

[0352] Where S_i' represents the calibrated score of the i-th audiovisual unit, Ŝ_i represents the normalized initial composite score, R̂_i represents the sorted score converted and normalized from the pairwise comparison results, and η represents the fusion coefficient, and 0≤η≤1.

[0353] Furthermore, the process of fusing multi-dimensional scores, adaptive weights, and ranking calibration results through comprehensive evaluation of the intelligent agent includes:

[0354] Read the visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, emotional semantic consistency score, adaptive weight vector, and sorting calibration results;

[0355] The overall aesthetic score of the audiovisual content to be evaluated is calculated based on the calibrated score. ;

[0356] A weighted overall score is calculated based on the importance of each segment. , where μ_i represents the importance weight of the i-th audiovisual unit, and satisfies Σ[i=1→L]μ_i=1;

[0357] The segment importance weight μ_i is calculated based on the segment duration ratio D_i, the density of visual and audio events E_i, the proportion of the main or target object O_i, the semantic keyness G_i, and the impact of abnormal segments on the overall evaluation result B_i. Here, D_i is the ratio of the duration of segment i to the total duration of all L segments; E_i is the ratio of the sum of the number of visual and audio events in segment i to the segment duration, normalized to the range of 0 to 1; O_i is the proportion of the number of frames in segment i where the main or target object appears to the total number of frames in that segment; G_i is determined by the weighted normalized result of the number of hit keywords, the confidence of semantic turning point detection, and the matching degree of plot nodes; and B_i is determined by the weighted normalized result of the deduction magnitude of abnormal segments on the dimension score and the duration of the abnormality. Each component is a computable indicator based on features. ρ1 to ρ5 are non-negative coefficients, and norm represents the normalization process that makes the sum of the importance weights of each segment equal to 1.

[0358] When conflicts arise between scores of different dimensions, a review agent is invoked for collaborative review. The conflicts include at least one of the following: visual aesthetics score not lower than a first preset threshold and audiovisual synchronization score lower than a second preset threshold; audio aesthetics score not lower than a third preset threshold and emotional semantic consistency score lower than a fourth preset threshold; overall aesthetics score not lower than a fifth preset threshold and ranking preference probability lower than a sixth preset threshold; and the variation of segment-level scores between adjacent segments exceeding a seventh preset threshold. The comprehensive evaluation agent generates a review task based on the conflict type. The review task includes conflicting segments, original dimension scores, evidence segments, corresponding feature values, confidence levels, and ranking preference results.

[0359] Based on the type of conflict, the review agent re-invokes the conflict-related dimension analysis agent on the same segment-level audiovisual unit to obtain the review score, review evidence, and alternative score.

[0360] When the difference between the alternative score and the original score exceeds the review threshold and the confidence level of the alternative score is higher than that of the original score, the comprehensive evaluation agent uses the alternative score to update the corresponding dimension score, weight, or ranking calibration result; otherwise, the original score is retained and the reason for the conflict is recorded to form the final evaluation result.

[0361] Furthermore, the output evaluation report includes:

[0362] Output an overall audiovisual aesthetic score, wherein the overall audiovisual aesthetic score is in at least one of the following forms: a percentage system, a ten-point system, a grade system, or a ranking system;

[0363] Output dimension-level scores, which include visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0364] Output fragment-level diagnostic results, which include low-score fragment timestamps, corresponding problem types, dimensional differences, low-score thresholds, evidence fragment indexes, impact degree values, and suggested modification direction codes or texts. The evidence fragment index includes at least start and end timestamps, keyframe indexes, and audio fragment indexes.

[0365] Output relative ranking results, which include the overall ranking among multiple audiovisual versions, the difference in dimensionality of each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results.

[0366] The output includes interpretable evaluation text and machine-readable evaluation data, wherein the machine-readable evaluation data includes at least one of JSON files, tabular files, database records, or interface-returned data.

[0367] Furthermore, the method also includes evaluation annotation and model validation steps:

[0368] Acquire human preference label data, which includes pairwise preference labels, dimension rating labels, or threshold labels for audiovisual units;

[0369] The artificial preference label data is used to train or calibrate the feature index classifiers, normalization parameters, or scoring thresholds of the intelligent agent for each dimension of analysis.

[0370] During the evaluation phase, the feature calculation results are used as the scoring basis, and the scoring results are verified. The verification includes: consistency verification based on P_score(U_i>U_j) and P_pref(U_i>U_j) output by the fine-grained preference comparison agent; cross-modal consistency verification between visual emotion vectors, audio emotion vectors, and text semantic vectors; and confidence verification based on comparison results with similar samples.

[0371] When the verification fails, the review agent is triggered to review the data. The manual preference label data does not directly replace the score calculated from the features.

[0372] Secondly, see Figure 9 As shown, this application provides an automatic audiovisual aesthetic evaluation device, which includes:

[0373] The context task construction module is used to construct the context of the audiovisual aesthetic evaluation task based on the audiovisual content to be evaluated and its associated information.

[0374] The segment unit acquisition module is used to perform video frame sequence, audio segment, subtitle text and timestamp alignment processing on the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units;

[0375] The multi-dimensional evaluation module is used to obtain scores corresponding to multiple evaluation dimensions based on the segment-level audiovisual unit and in combination with the context of the audiovisual aesthetic evaluation task, through feature extraction, semantic recognition, index normalization and weighted fusion of the intelligent agent.

[0376] The overall aesthetic evaluation module is used to estimate the agent by weights and determine the adaptive weights of each evaluation dimension based on content type, modal saliency, fragment complexity and evaluation objectives.

[0377] The segment segmentation module is used to evaluate the similarity of each segment-level audiovisual unit, and classify each segment-level audiovisual unit with a similarity not lower than a preset intra-group similarity threshold into the same similar audiovisual sample group, and otherwise classify them into a comparable segment group.

[0378] The relative ranking calibration module is used to perform pairwise comparisons of similar audiovisual sample groups or comparable segment groups by a fine-grained preference comparison agent, and perform relative ranking calibration based on the comparison results to obtain the relative ranking calibration result.

[0379] The comprehensive evaluation module is used to fuse scores, adaptive weights, and relative ranking calibration results from multiple evaluation dimensions through a comprehensive evaluation agent, forming a total audiovisual aesthetic score and dimension-level diagnostic results; among them...

[0380] The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files.

[0381] The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags;

[0382] The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0383] In this application embodiment, the technical problems of the prior art, such as the single dimension of audiovisual aesthetic evaluation, insufficient audio-visual collaborative modeling, difficulty in identifying fine-grained differences, subjectivity of key parameters, and insufficient interpretability of evaluation results, are addressed.

[0384] Furthermore, the context task construction module is also used to receive the audiovisual content to be evaluated;

[0385] The context task construction module is also used to receive the associated information.

[0386] The context task construction module is also used to perform format unification and quality preprocessing on the audiovisual content to be evaluated;

[0387] The context task construction module is also used to perform structured processing on the associated information, extracting content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags;

[0388] The context task construction module is also used to integrate the audiovisual content to be evaluated (after unified formatting or quality preprocessing), the structured association information, the segment timeline, and the feature index into an audiovisual aesthetic evaluation task context that can be invoked by multiple agents, based on the content's unique identifier and version identifier. This ensures that each audiovisual content to be evaluated or each edit version forms a one-to-one and traceable mapping relationship with its association information and the audiovisual aesthetic evaluation task context.

[0389] The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

[0390] Furthermore, the segment splitting module is also used to perform shot boundary detection on the video frame sequence of the audiovisual content to be evaluated, obtain multiple shot segments, and extract keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features, and sharpness features from each shot segment;

[0391] The segment splitting module is also used to segment the audio segments of the audiovisual content to be evaluated, obtain multiple audio segments, and extract loudness, beat, accent, pitch change, spectral centroid, energy change, speech intelligibility, background noise, musical emotion and timbre texture features.

[0392] The segment segmentation module is also used to perform semantic analysis on the subtitle text of the audiovisual content to be evaluated, and to obtain character speech, emotion words, event nodes, semantic turning points and theme keywords;

[0393] The segment splitting module is also used to align video shots, audio clips, and subtitle text based on the timestamps of the audiovisual content to be evaluated, forming segment-level audiovisual units.

[0394] Furthermore, the device also includes a report output module;

[0395] The report output module is used to output an evaluation report; wherein,

[0396] The evaluation report includes the overall aesthetic score, scores corresponding to multiple evaluation dimensions, and the relative ranking calibration results.

[0397] Furthermore, the report output module is also used to output the overall aesthetic score, which is in at least one of the following forms: percentage system, ten-point system, grade system or ranking system.

[0398] The report output module is also used to output scores corresponding to multiple evaluation dimensions, including visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

[0399] The report output module is also used to output the relative ranking calibration results, which include the overall ranking among multiple audiovisual versions, the difference in dimensionality of each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results.

[0400] In summary, the audiovisual aesthetic automatic evaluation device provided in this application embodiment has the same technical principle as the audiovisual aesthetic automatic evaluation method provided in the first aspect in terms of technical problems, technical solutions and technical effects, so it will not be described in detail here.

[0401] Thirdly, embodiments of this application provide an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps described in the method mentioned in the first aspect.

[0402] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions for causing a processor to perform the steps described in the method mentioned in the first aspect.

[0403] In the description of this application, it should be noted that the terms "upper," "lower," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0404] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0405] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. An automatic evaluation method for audiovisual aesthetics, characterized in that, The method includes the following steps: Construct the audiovisual aesthetics assessment task context based on the audiovisual content to be evaluated and its related information; The audiovisual content to be evaluated is processed by aligning video frame sequences, audio segments, subtitle text, and timestamps to obtain multiple segment-level audiovisual units; Based on the aforementioned segment-level audiovisual unit, and combined with the context of the audiovisual aesthetic evaluation task, the intelligent agent is analyzed through multiple dimensions, including feature extraction, semantic recognition, index normalization, and weighted fusion, to obtain scores corresponding to various evaluation dimensions. By estimating the agent through weights, adaptive weights for each evaluation dimension are determined based on content type, modal saliency, fragment complexity, and evaluation objectives, and the corresponding overall aesthetic score is obtained. For each of the aforementioned segment-level audiovisual units, a similarity assessment is performed. Each segment-level audiovisual unit with a similarity not lower than a preset intra-group similarity threshold is grouped into the same similar audiovisual sample group, and otherwise it is grouped into a comparable segment group. By using a fine-grained preference comparison agent, pairs of similar audiovisual sample groups or comparable segment groups are compared, and relative ranking calibration is performed based on the comparison results to obtain the relative ranking calibration results. By comprehensively evaluating the agent and fusing scores, adaptive weights, and relative ranking calibration results corresponding to multiple evaluation dimensions, a total audiovisual aesthetic score and dimension-level diagnostic results are formed; among them... The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files. The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags; The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

2. The automatic evaluation method for audiovisual aesthetics as described in claim 1, characterized in that, The process of constructing an audiovisual aesthetic evaluation task context based on the audiovisual content to be evaluated and its associated information includes the following steps: Receive the audiovisual content to be evaluated; Receive the associated information, The audiovisual content to be evaluated is subjected to format standardization and quality preprocessing. The associated information is structured to extract content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags; Based on unique content identifiers and version identifiers, the audiovisual content to be evaluated (after standardized formatting or quality preprocessing), the related information after structured processing, the segment timeline, and feature indexes are integrated into an audiovisual aesthetic evaluation task context that can be invoked by multiple agents. This creates a one-to-one and traceable mapping relationship between each audiovisual content to be evaluated or each edit version and its related information and audiovisual aesthetic evaluation task context; among which, The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

3. The automatic evaluation method for audiovisual aesthetics as described in claim 1, characterized in that, The process of aligning the video frame sequence, audio segment, subtitle text, and timestamp of the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units includes the following steps: Shot boundary detection is performed on the video frame sequence of the audiovisual content to be evaluated to obtain multiple shot segments, and keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features, and sharpness features are extracted from each shot segment; The audio segments of the audiovisual content to be evaluated are segmented to obtain multiple audio segments, and loudness, beat, accent, pitch variation, spectral centroid, energy variation, speech intelligibility, background noise, musical mood and timbre texture features are extracted. Semantic analysis is performed on the subtitle text of the audiovisual content to be evaluated to obtain character speech, emotion words, event nodes, semantic turning points and thematic keywords; Based on the timestamps of the audiovisual content to be evaluated, video shots, audio clips, and subtitle text are aligned to form segment-level audiovisual units.

4. The automatic audiovisual aesthetics evaluation method as described in claim 1, characterized in that, The method further includes the following steps: Output an evaluation report; The output evaluation report includes the following steps: Output the overall aesthetic score, which is in at least one of the following forms: a percentage system, a ten-point system, a grade system, or a ranking system. Output scores corresponding to multiple evaluation dimensions, including visual aesthetics score, audio aesthetics score, rhythmic structure score, audiovisual synchronization score, and emotional semantic consistency score. The relative ranking calibration results are output, including the overall ranking among multiple audiovisual versions, the dimensional differences between each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results; wherein... The evaluation report includes the overall aesthetic score, scores corresponding to multiple evaluation dimensions, and the relative ranking calibration results.

5. An automatic audiovisual aesthetic evaluation device, characterized in that, The device includes: The context task construction module is used to construct the context of the audiovisual aesthetic evaluation task based on the audiovisual content to be evaluated and its associated information. The segment unit acquisition module is used to perform video frame sequence, audio segment, subtitle text and timestamp alignment processing on the audiovisual content to be evaluated to obtain multiple segment-level audiovisual units; The multi-dimensional evaluation module is used to obtain scores corresponding to multiple evaluation dimensions based on the segment-level audiovisual unit and in combination with the context of the audiovisual aesthetic evaluation task, through feature extraction, semantic recognition, index normalization and weighted fusion of the intelligent agent. The overall aesthetic evaluation module is used to estimate the agent by weights and determine the adaptive weights of each evaluation dimension based on content type, modal saliency, fragment complexity and evaluation objectives. The segment segmentation module is used to evaluate the similarity of each segment-level audiovisual unit, and classify each segment-level audiovisual unit with a similarity not lower than a preset intra-group similarity threshold into the same similar audiovisual sample group, and otherwise classify them into a comparable segment group. The relative ranking calibration module is used to perform pairwise comparisons of similar audiovisual sample groups or comparable segment groups by a fine-grained preference comparison agent, and perform relative ranking calibration based on the comparison results to obtain the relative ranking calibration result. The comprehensive evaluation module is used to fuse scores, adaptive weights, and relative ranking calibration results from multiple evaluation dimensions through a comprehensive evaluation agent, forming a total audiovisual aesthetic score and dimension-level diagnostic results; among them... The audiovisual content to be evaluated includes at least one of the following: video files containing video tracks and audio tracks, video links, live stream clips, short drama clips, advertising clips, music videos, film and television clips, interview videos, teaching videos, documentary clips, social media audio and video, or audiovisual content formed by combining image sequences and audio files. The associated information includes at least one of the following: content type, evaluation objective, target audience, video duration, resolution, frame rate, audio sampling rate, subtitle text, edit version identifier, content source, reference sample, and human preference tags; The scores corresponding to the various evaluation dimensions include visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score.

6. The audiovisual aesthetic automatic evaluation device as described in claim 5, characterized in that: The context task construction module is also used to receive the audiovisual content to be evaluated; The context task construction module is also used to receive the associated information. The context task construction module is also used to perform format unification and quality preprocessing on the audiovisual content to be evaluated; The context task construction module is also used to perform structured processing on the associated information, extracting content type tags, topic tags, emotion tags, style tags, scene tags, rhythm tags, and user target tags; The context task construction module is also used to integrate the audiovisual content to be evaluated (after unified formatting or quality preprocessing), the structured association information, the segment timeline, and the feature index into an audiovisual aesthetic evaluation task context that can be invoked by multiple agents, based on the content's unique identifier and version identifier. This ensures that each audiovisual content to be evaluated or each edit version forms a one-to-one and traceable mapping relationship with its association information and the audiovisual aesthetic evaluation task context. The quality preprocessing includes at least one of video decoding, audio decoding, frame rate normalization, loudness normalization, silent segment detection, black screen segment detection, subtitle extraction, and abnormal segment marking.

7. The audiovisual aesthetic automatic evaluation device as described in claim 5, characterized in that: The segment splitting module is also used to perform shot boundary detection on the video frame sequence of the audiovisual content to be evaluated, obtain multiple shot segments, and extract keyframes, representative frame sequences, motion intensity features, main area features, composition features, color distribution features, and sharpness features from each shot segment; The segment splitting module is also used to segment the audio segments of the audiovisual content to be evaluated, obtain multiple audio segments, and extract loudness, beat, accent, pitch change, spectral centroid, energy change, speech intelligibility, background noise, musical emotion and timbre texture features. The segment segmentation module is also used to perform semantic analysis on the subtitle text of the audiovisual content to be evaluated, and to obtain character speech, emotion words, event nodes, semantic turning points and theme keywords; The segment splitting module is also used to align video shots, audio clips, and subtitle text based on the timestamps of the audiovisual content to be evaluated, forming segment-level audiovisual units.

8. The automatic audiovisual aesthetic evaluation device as described in claim 5, characterized in that, The device also includes a report output module; The report output module is used to output an evaluation report; The report output module is also used to output the overall aesthetic score, which is in at least one of the following forms: percentage system, ten-point system, grade system or ranking system. The report output module is also used to output scores corresponding to multiple evaluation dimensions, including visual aesthetics score, audio aesthetics score, rhythm and structure score, audiovisual synchronization score, and emotional semantic consistency score. The report output module is also used to output the relative ranking calibration results, which include the overall ranking among multiple audiovisual versions, the difference in dimensionality between each version, the ranking calibration contribution, the ranking of comparable segments within a single video, and the pairwise preference comparison results; wherein... The evaluation report includes the overall aesthetic score, scores corresponding to multiple evaluation dimensions, and the relative ranking calibration results.

9. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to perform the method of any one of claims 1 to 4.

Images