HABSS: Heuristic Attention-Based Structure-Aware Summarization with coverage preservation and redundancy reduction for video preprocessing

This study addresses the limitation of conventional video summarization approaches that rely primarily on top-k importance ranking, often overlooking semantic redundancy and representative coverage in the preprocessing stage. To overcome this gap, we propose HABSS (Heuristic Attention-Based Structure-Aware Summarization), a novel preprocessing-oriented video summarization framework that generates compact yet semantically representative frame subsets before final summary construction. The proposed method integrates Vision Transformer (ViT)-based semantic feature extraction, similarity-driven hierarchical clustering, and adaptive cluster-level frame budgeting to preserve content diversity while minimizing redundancy. In addition, an enhanced variant, HABSS-ATTN, incorporates a self-supervised temporal attention scoring module to improve representative frame selection within clusters. Experiments conducted on the benchmark SumMe and TVSum datasets demonstrate that the proposed framework achieves high semantic coverage, selecting representative subsets that preserve up to 81.7% of original content while reducing frame volume by 70–85%. Comparative analyses further show that HABSS improves diversity-aware summarization performance over conventional ranking-based preprocessing strategies, yielding lower redundancy and more balanced semantic representation. These results establish preprocessing-aware summarization as an effective independent optimization stage and highlight the potential of HABSS as a robust foundation for next-generation intelligent video summarization systems.