A Deep Learning Pipeline for Adaptive Bayesian-Optimized Keyframe Extraction in Video Summarization

Automated keyframe extraction plays a crucial role in video summarization, facilitating efficient browsing and indexing of long video sequences. Achieving high performance in this task depends on robust feature representation and reliable evaluation metrics. The goal of keyframe extraction is to capture the essence of a video through its most representative frames. This paper introduces a deep learning-based pipeline for video summarization that combines adaptive frame processing with advanced feature extraction strategies. The proposed framework operates in multiple stages: first, it applies percentile-based thresholds for intelligent frame filtering, removing low-quality frames while retaining semantic integrity. Next, content-aware sampling guided by optical flow is used to prioritize motion-rich and informative frames. At the core, an enhanced ConvNeXt architecture with attention-driven spatial pooling ensures discriminative feature representation, which is followed by Bayesian-optimized clustering and Maximal Marginal Relevance selection for precise keyframe identification. Experimental results show that the proposed method achieves an excellent trade-off between precision and recall, while qualitative analysis highlights the generation of compact, semantically meaningful summaries with minimal redundancy. Overall, the proposed approach demonstrates clear improvements over existing state-of-the-art methods.