Efficient Hierarchical Temporal Audio-Video Cross-Attention Fusion Network for Audio-Enhanced Text-To-Video Retrieval

With video and audio being integral to modern multimedia content, accurately retrieving relevant segments based on textual queries is crucial for enhancing user experience and information accessibility. However, contextual misalignment across video segments presents significant challenges, particularly when different segments exhibit varying degrees of relevance to specific portions of a text query. To address this issue, a novel Hierarchical Temporal Audio-Video Cross-Attention Fusion Network has been developed. This model utilizes a Video Swim Feature Pyramid video encoder to enhance the extraction of multi-scale spatial features and capture intricate details within videos. Additionally, a Temporal RoBERTa Graph Network serves as the text encoder, enabling a deep understanding of relationships within the text and allowing for minute interpretations of queries that encompass multiple themes. To effectively align video and audio representations with textual queries, the model employs a Hierarchical multiscale spatial-temporal attention mechanism. Furthermore, an Audio Spectrogram Short-Term Memory Transformer is utilized to capture the temporal dynamics of complex audio streams. To refine audio-text alignment, the model incorporates a Threshold-Based audio-text Dynamic Time cross-attention block, which selectively filters irrelevant audio components and dynamically adjusts for temporal misalignments. The experimental results demonstrate that the proposed model significantly enhances retrieval accuracy by effectively aligning video and audio representations with textual queries, resolving multi-scene transitions, and isolating relevant audio cues among complex soundscapes.