Hybrid ViT and ConvGRU Framework Achieves 88.88% Accuracy in Detecting Dengue-Infected Mosquitoes
A recent research paper available on arXiv (2608.11582) introduces a three-phase approach to differentiate dengue virus-infected mosquitoes from control specimens by examining their movement patterns in video footage. Initially, YOLO 11M is employed to identify mosquitoes and eliminate background noise. Next, visual features are extracted using a Vision Transformer (ViT). The final step involves classifying the videos with a convolutional Gated Recurrent Unit (ConvGRU). When compared to RNN, LSTM, GRU, and their convolutional counterparts, ConvGRU outperformed them, achieving an accuracy of 88.88%, precision of 84.45%, recall of 82.82%, and an F1 score of 82.81%. This study tackles the difficulties posed by the small size and intricate backgrounds of video frames, which traditional AI techniques often struggle with, indicating potential advancements in disease vector monitoring and management.
Key facts
- The study is published on arXiv with ID 2608.11582.
- The framework uses YOLO 11M for mosquito detection and background removal.
- Vision Transformer (ViT) is used for visual feature extraction.
- ConvGRU classifier is used for video classification.
- ConvGRU achieved 88.88% accuracy, 84.45% precision, 82.82% recall, and 82.81% F1 score.
- Comparative models included RNN, LSTM, GRU, and their convolutional versions.
- The challenge is identifying dengue-infected mosquitoes from control mosquitoes in video.
- The framework addresses limitations of conventional AI methods in feature extraction from video frames.
Entities
Institutions
- arXiv