What question did this study set out to answer?

The study aims to develop a deep learning framework that effectively detects violence in video data with explainability features.

February 28, 2026Scientific ReportsOpen Access

An explainable deep learning framework for video violence detection using unsupervised keyframe selection and attention-based CNN

View Full Paper

Ask AI

Bookmark

Authors

This framework demonstrates improved violence detection in videos, indicating potential for real-time safety applications.

The study aims to develop a deep learning framework that effectively detects violence in video data with explainability features.
Developed an Explainable Attention-Enhanced CNN framework.
Utilized unsupervised keyframe selection based on similarity clustering.
Incorporated attention modules to improve feature discrimination.
Employed Grad-CAM++ for visual interpretability of model decisions.
Conducted experiments on five benchmark datasets.
Achieved an average accuracy of 94.6% and an F1-score of 93.9%.
Outperformed state-of-the-art models like C3D and ResNet-LSTM.
Demonstrated near-real-time processing efficiency at approximately 62 FPS.
Showed significant improvements in performance due to keyframe selection and attention modules (p < 0.05).
Established large effect sizes (η² = 0.76) in statistical analyses.

Azim et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark

February 28, 2026Scientific ReportsOpen Access

View Full Paper

Ask AI

Bookmark

Authors

This framework demonstrates improved violence detection in videos, indicating potential for real-time safety applications.

The study aims to develop a deep learning framework that effectively detects violence in video data with explainability features.
Developed an Explainable Attention-Enhanced CNN framework.
Utilized unsupervised keyframe selection based on similarity clustering.
Incorporated attention modules to improve feature discrimination.
Employed Grad-CAM++ for visual interpretability of model decisions.
Conducted experiments on five benchmark datasets.
Achieved an average accuracy of 94.6% and an F1-score of 93.9%.
Outperformed state-of-the-art models like C3D and ResNet-LSTM.
Demonstrated near-real-time processing efficiency at approximately 62 FPS.
Showed significant improvements in performance due to keyframe selection and attention modules (p < 0.05).
Established large effect sizes (η² = 0.76) in statistical analyses.

Azim et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark