A Deep Image Representation and Temporal Correlation Modeling Approach for Film Scene Style Evolution Analysis

초록

The structured representation and dynamic modeling of non-rigid, high-level aesthetic visual styles remain core challenges in computer vision. The ambiguity and temporal dynamics of such styles make it difficult for traditional methods to achieve accurate characterization. As aAtypical medium for this challenge, film visual styles evolve dynamically along with the narrative progression, integrating multiple aesthetic attributes such as color, lighting, and composition. This places high demands on the representational capabilities and temporal modeling precision of analysis methods. To address these challenges, we propose an endto-end general framework for "deep image representationA- temporal correlation modelingAstyle evolution analysis." The core innovation of this framework lies in constructing a triune deep image representation that integrates local textures, global semantics, and style prototypes, tailored to the multi-dimensional nature of aesthetic styles. A narrative-guided hierarchical attention masking mechanism is designed to enhance the relevance of dynamic evolution modeling. The key contributions of this research include: the construction and public release of the FilmStyleEvoBench benchmark dataset, accompanied by standard evaluation tasks and metrics; and cross-domain validation through painting, architectural videos, and user-generated content, which demonstrates the generalization potential of the method. Experiments based on FilmStyleEvoBench and cross-domain datasets show that the proposed method significantly outperforms existing comparison methods in style recognition, evolution change point detection, and temporal correlation quantification tasks, with stable and effective cross-domain transfer performance. This method not only solves key issues in the analysis of film visual style evolution but also provides a universal methodology for visual aesthetic computation and structured understanding of long videos, while empowering film industry creative support and digital humanities quantitative research.

키워드