Dynamic video summarization using transformers

In today's digital ecosystem, video has become the dominant medium for communication, education, and entertainment, yet the exponential growth of content poses significant challenges for efficient consumption. Manual summarization of long videos is time-consuming and impractical, creating the need for intelligent, automated solutions. This work proposes a dynamic video summarization framework that integrates ResNet-50 for robust spatial feature extraction with a Transformer encoder to capture temporal dependencies through self-attention. Frames are sampled at fixed intervals, encoded with positional information, and assigned importance scores using a scoring head with sigmoid activation, enabling the system to generate concise and coherent summaries that preserve the narrative flow. Experimental evaluation across diverse video domains, including sports, surveillance, and daily-life recordings, demonstrates that the proposed model significantly outperforms LSTM-, GRU-, and vanilla Transformer-based baselines, achieving an F1-score of 79.8% while reducing video length by up to 70%. Qualitative results further highlight the model's ability to retain semantically meaningful content while eliminating redundancy. This study establishes a scalable, efficient, and accurate approach to automated video summarization, with strong potential for applications in e-learning, content retrieval, surveillance monitoring, and healthcare analytics.