Multimodal Feature Fusion Video Description Model Integrating Attention Mechanisms and Contrastive Learning

To avoid the issue of significant redundancy in the spatiotemporal features extracted from multimodal video description methods and the substantial semantic gaps between different modalities within video data. Building upon the TimeSformer model, this paper proposes a two-stage video description approach (Multimodal Feature Fusion Video Description Model Integrating Attention Mechanism and Contrastive Learning, MFFCL). The TimeSformer encoder extracts spatiotemporal attention features from the input video and performs feature selection. Contrastive learning is employed to establish semantic associations between the spatiotemporal attention features and textual descriptions. Finally, GPT2 is employed to generate descriptive text. Experimental validations on the MAVD, MSR-VTT, and VATEX datasets were conducted against several typical benchmark methods, including Swin-BERT and GIT. The results indicate that the proposed method achieves outstanding performance on metrics such as Bleu-4, METEOR, ROUGE-L, and CIDEr. The spatiotemporal attention features extracted by the model can fully express the video content and that the language model can generate complete video description text.

@article{Zhihao2024,
  title     = {Multimodal Feature Fusion Video Description Model Integrating Attention Mechanisms and Contrastive Learning},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {4},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Wang Zhihao and Che Zhanbin},
  doi       = {10.14569/IJACSA.2024.0150440},
  url       = {https://doi.org/10.14569/IJACSA.2024.0150440}
}