Unsupervised Video Summarization via BiLSTM-Attention and Reconstruction-Based Scoring

Authors

  • Manjusha Jagannath Yeola Computer Engineering Department of MIT Academy of Engineering, Savitribai Phule Pune University, India https://orcid.org/0009-0005-7804-2977
  • Sunita S. Barve Computer Engineering Department of MIT Academy of Engineering, Savitribai Phule Pune University, India https://orcid.org/0000-0002-5679-6744

DOI:

https://doi.org/10.47852/bonviewJCCE62028849

Keywords:

multi-head attention, reconstruction learning, temporal feature representation, unsupervised learning, video summarization

Abstract

The goal of video summarization is to automatically summarize a video to produce an informative summary by extracting the representative portion of the video. Recent unsupervised methods have decreased the reliance on manual annotations, but by effectively representing long-range temporal relationships while keeping computations tractable, this is a major challenge. In this paper, we propose a compact frame-based unsupervised video summarization framework based on bidirectional long short-term memory (BiLSTM) network and multi-head attention mechanism for improving the temporal feature representation and keyframe selection. Extensive experiments were performed on two benchmark datasets, SumMe and TVSum, using six different BiLSTM-based attention configurations. To explore the effect of the attention mechanisms and training samples, the performance was evaluated by F1-score, precision, and recall. Experimental results indicate that the configuration of BiLSTM–4 Heads consistently yielded the best trade-off with the highest F1-scores of 0.5896 on the SumMe dataset and 0.6721 on the TVSum dataset. The ablation study also shows that four-head attention configuration consistently outperforms the conventional attention configurations in terms of contextual representation. In general, the proposed framework proves that the integration of BiLSTM, multi-head attention, and reconstruction-driven scoring is a computational efficient and effective approach for fully unsupervised video summarization.



Received: 20 December 2025 | Revised: 23 July 2026 | Accepted: 10 September 2026



Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.



Data Availability Statement

The data that support the findings of this study are openly available in Zenodo at https://zenodo.org/records/4884870 and Github at https://github.com/yalesong/tvsum.



Author Contribution Statement

Manjusha Jagannath Yeola: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Sunita S. Barve: Writing – original draft, Writing – review & editing, Visualization, Supervision, Project administration.

Downloads

Published

2026-11-20

Issue

Section

Research Articles

How to Cite

Yeola, M. J., & Barve, S. S. (2026). Unsupervised Video Summarization via BiLSTM-Attention and Reconstruction-Based Scoring. Journal of Computational and Cognitive Engineering. https://doi.org/10.47852/bonviewJCCE62028849