Optimizing Aedes Mosquito Detection Using YOLOv5-S ROI Localization and Confidence-Weighted Voting via Reinforcement-Based Pruning

Aedes mosquitoes are the principal dengue fever vectors, however, detecting them is still a major challenge, as it requires high classification accuracy, along with the need for efficient computational models for mosquito surveillance. One crucial optimization problem is to strike a balance between classification performance and reducing computational overhead. This study introduces a lightweight detection framework that adopts YOLOv5-S for ROI localization and employs an ensemble of MobileNetV2, ShuffleNetV2 and EfficientNetV2 optimized through reinforcement learning (RL)-based pruning. Confidence weighted voting (CWV) is applied to improve classification reliability further. Results of the experiment show that YOLOv5S obtained a [email protected] of 99.27%. The training and validation accuracies of the generated model in MobileNetV2, EfficientNetV2, and ShuffleNetV2 remained high, after pruning at 99.39%, 98.41%, 99.88%, 99.39%, and 99.57%, 98.414%, respectively. The RL-based pruning also reduced the parameters of MobileNetV2, ShuffleNetV2, and EfficientNetV2 to 1.93M, 0.81M, and 11.93M, achieving sparsity rates of 35.97%, 41.74%, and 43.12%, respectively. The CWV ensemble also increases classification performance even more to 99.94% training accuracy and 99.26% validation accuracy, indicating the CWV ensemble as a promising solution for scalable and cost-effective mosquito surveillance.