Video Captioning using Deep Learning with Greedy Search (VCDLGS)

The exponential growth of digital video content has intensified the demand for automated video captioning systems that can bridge visual understanding and natural language processing. Video captioning, which generates descriptive sentences for video sequences, plays a crucial role in enhancing accessibility, content retrieval, and human–computer interaction. This paper presents a novel Video Captioning using Deep Learning with Greedy Search (VCDLGS) model that addresses the challenges of temporal dynamics and contextual dependencies inherent in video content. The proposed framework integrates EfficientNet for robust visual feature extraction, GloVe semantic embeddings for enhanced linguistic representation, and a sequence-to-sequence Long Short-Term Memory architecture with multimodal attention mechanisms. The model employs a greedy search decoding strategy to generate coherent and contextually relevant captions efficiently. Comprehensive evaluation on the Microsoft Research Video Description Corpus dataset demonstrates the effectiveness of our approach, achieving competitive performance with a BLEU score of 64.82 (23.12 point improvement over S2VT baseline), METEOR score of 46.10 (16.90 point improvement), and CIDEr score of 144.00 (92.30 point improvement). These results represent substantial advances over several state-of-the-art baselines, with greedy search providing 42% faster inference than beam search while maintaining comparable quality. The VCDLGS model contributes to advancing automated video understanding technology while providing an efficient solution suitable for real-time applications with 18.5 fps processing capability. This work establishes a foundation for improved content accessibility and multimedia comprehension across diverse domains.